Reuse and Sharing of Electronic Health Record Data
with a focus on Primary Care and Disease Coding
Annet Sollie
Reuse and Sharing of Electronic Health Record DataPhD thesis, with a summary in Dutch.
No part of this thesis may be reproduced without prior permission of the author.
ISBN: 978-94-6182-757-9
Author: J.W. (Annet) Sollie, copyright © 2016
Cover: Janneke Laarakkers, PlanPuur
Lay-out & Print: Off Page, Amsterdam
VRIJE UNIVERSITEIT
Reuse and Sharing of Electronic Health Record Data
with a focus on Primary Care and Disease Coding
ACADEMISCH PROEFSCHRIFT
ter verkrijging van de graad Doctor aan
de Vrije Universiteit Amsterdam,
op gezag van de rector magnificus
prof.dr. V. Subramaniam,
in het openbaar te verdedigen
ten overstaan van de promotiecommissie
van de Faculteit der Geneeskunde
op vrijdag 27 januari 2017 om 11.45 uur
in de aula van de universiteit,
De Boelelaan 1105
door
Johanna Wilhelmina Sollie
geboren te Zwolle
promotoren:
prof.dr. M.E. Numans
prof.dr. R.H. Sijmons
copromotor:
dr. C.W.Helsper
thesis committee: prof. dr. H.E. van der Horst VU Medisch Centrum, Amsterdam
prof. dr. R.A.M.J. Damoiseaux Universitair Medisch Centrum, Utrecht
prof. dr. M.C. Cornel VU Medisch Centrum, Amsterdam
prof. dr. S. Brinkkemper Universiteit Utrecht
dr.ir. R. Cornet Academisch Medisch Centrum Amsterdam
The printing of this thesis was financially supported by SBOH, employer of GP-trainees
in the Netherlands
5
thesis committee prof. dr. H.E. van der Horst VU Medisch Centrum, Amsterdam prof. dr. R.A.M.J. Damoiseaux Universitair Medisch Centrum, Utrecht
prof. dr. M.C. Cornel VU Medisch Centrum, Amsterdam prof. dr. S. Brinkkemper Universiteit Utrecht dr.ir. R. Cornet Academisch Medisch Centrum Amsterdam
The printing of this thesis was financially supported by SBOH, employer of GP-trainees in the Netherlands
Voor mijn kinderen:
Kristel
Nathalie
Mike
Marilyn
9
ContEntS
Chapter 1 General Introduction, aims and outline of the thesis 11
Quality of Data – Literature ReviewChapter 2 Quality of Data in the Primary Care Electronic Health Record
(EHR) System 25(Published: Huisarts & Wetenschap August 2013)
Quality and Reusability of Primary Care EHR Data – Hands-on identification of bottlenecksChapter 3 Reusability of coded data in the primary care Electronic Medical
Record: a dynamic cohort study concerning cancer diagnoses 35(Published: Int. J. Med. Inf. August 2016)
Chapter 4 Do GPs know their cancer patients? Assessing the quality
of cancer registration in Dutch Primary Care: a cross sectional
validation study 57(Published: BMJ Open September 2016)
Chapter 5 Primary Care management of women with breast cancer related
concerns - a dynamic cohort study using a Network Database 79(Published: Eur. J. Cancer Care June 2016)
Strategies & Solutions for improving data quality and enabling data re-use and sharingChapter 6 A new coding system for metabolic disorders demonstrates
gaps in the international disease classifications ICD-10 and
SNOMED-CT which can be barriers to genotype-phenotype
data sharing 103(Published: Human Mutation July 2013)
Chapter 7 SORTA: a System for Ontology-based Recoding and Technical
Annotation of biomedical phenotype data 119(Published: Database, the journal of biological databases and curation
September 2015)
Chapter 8 Proposed roadmap to stepwise integration of genetics in family
medicine and clinical research 143(Published: Clinical & Translational Medicine, February 2013)
Summary
Chapter 9 Summarizing Discussion 157
Chapter 10 Nederlandse Samenvatting, Dankwoord, Curriculum Vitae 175
CHaPtERGeneral Introduction, aims
and outline of the thesis1
13
1
Gen
era
l Intr
od
uc
tIon
, aIm
s an
d o
utlIn
e of th
e thesIs
GEnERaL IntRoDuCtIon
To introduce the field of study of this thesis, the following case scenario is presented:
Case ScenarioJanuary 2016: a 48-year-old male consults his General Practitioner (GP) for a persistent
mucus-producing cough. He is worried because he has been coughing for more than
8 weeks now and used to be a solid smoker until he stopped 5 years ago. The cough is
interfering with his sporting activities and his sleep. The patient has no fever, no hoarseness
and does not cough up blood but did lose some weight over the last few months. The
GP performs a physical examination that turns out to be normal. According to the Dutch
College of General Practitioners’ guideline “Acute Cough” the GP decides to send the
patient to the hospital for a chest X-ray.
In her* Electronic Health Record (EHR) system the GP registers the consultation according to
the Subjective Objective Analysis Plan (SOAP)**[1] primary care registration structure as follows:
8
Chapter 1. General Introduction, aims and outline of the thesis
To introduce the field of study of this thesis, the following case scenario is presented:
Case Scenario January 2016: a 48-year-old male consults his General Practitioner (GP) for a
persistent mucus-producing cough. He is worried because he has been coughing for more
than 8 weeks now and used to be a solid smoker until he stopped 5 years ago. The cough is
interfering with his sporting activities and his sleep. The patient has no fever, no hoarseness
and does not cough up blood but did lose some weight over the last few months. The GP
performs a physical examination that turns out to be normal. According to the Dutch College
of General Practitioners’ guideline “Acute Cough” the GP decides to send the patient to the
hospital for a chest X-ray.
In her* Electronic Health Record (EHR) system the GP registers the consultation
according to the Subjective Objective Analysis Plan (SOAP)**[1] primary care registration
structure as follows:
The GP also creates a new Episode in the registry. [2]. Episodes are used to cluster
consultations that belong together and, as such, provide an overview of diagnoses. Based on
the text in the Analysis-line, the system suggests, among other codes, an ICPC-1
(International Classification of Primary Care) [3,4] code R84# (malignancy bronchus/lung)
for this consultation. She accepts this suggestion since no code or data-field is available to
register suspected cancer. Thinking the diagnosis of lung cancer is highly likely in this
patient, the GP uses the code for that disorder to make sure her opinion is recorded
prominently in the patient’s EHR.
Journal Date Description ICPC 15-01-2016 S Cough > 8 weeks, productive, no reported fever or haemoptysis.
Weight loss of 3 kg, insomnia, and exercise induced dyspnoea. Previous smoker (25 pack-years, stopped > 5 years)
O Not appear ill nor in respiratory distress. Respiratory rate 16/min. Chest examination: normal vesicular breath sounds, no added sounds. Hoarseness -. No palpable lymph nodes, throat: normal.
A Cough DD lung cancer? R84# P CXR and appt. 1 week
The GP also creates a new Episode in the registry. [2]. Episodes are used to cluster
consultations that belong together and, as such, provide an overview of diagnoses. Based
on the text in the analysis-line, the system suggests, among other codes, an ICPC-1
(International Classification of Primary Care) [3,4] code R84# (malignancy bronchus/lung)
for this consultation. She accepts this suggestion since no code or data-field is available
to register suspected cancer. Thinking the diagnosis of lung cancer is highly likely in this
patient, the GP uses the code for that disorder to make sure her opinion is recorded
prominently in the patient’s EHR.
9
The chest X-ray turns out to be normal and the GP shares this news with her patient in
a consultation by telephone. That day work is hectic and she forgets to remove the lung
cancer code from the EHR episode and to replace it with the appropriate code for this
consultation: the symptom code for “Cough (ICPC-1 code R05)”. The patient, reassured and
relieved, recovers completely in the following weeks.
However, on a sunny Saturday morning early March, the patient visits an out-of-hours
clinic because he sprained his ankle during the first football match of the season. He is
frightened out of his senses when the GP on duty asks him attentively if he has already started
his treatment for lung cancer.
A month later, his anonymized EHR record is sent to a large research database at the
nearby University where investigators are working on various projects such as early detection
of cancer. In this database he is now registered as a lung cancer patient and researchers assess
his consultations, lab results and radiology investigations in the 2 years prior to the diagnosis.
Little does the patient know that coming December he will also be counted as a “cancer case”
in the assessment of the GPs workload by an insurance company with the aim to determine
financial reimbursement and to calculate a quality indicator to asses quality of cancer care
provided by the GP.
Obviously, the GP did not make a mistake in her diagnostic workup with this patient,
however, she did make a mistake during her ICPC coding routine in the first consultation and
again, during the second consultation. The first mistake is that she accepted a “diagnosis”
code for a hypothesis, the second mistake is that she did not correct the code of the hypothesis
in the Episode into a symptom code [2]. Unfortunately, both easy-to-make mistakes are likely
to be common, because of time pressure in daily Primary Care practice (10 minutes per
consultation including expanding registration constraints) and EHR system design and
interface which have not yet fully evolved. These error-prone coding routines do not disturb
the medical process in every-day practice. However, problems do arise when medical files
containing these errors are used by someone else in the setting of clinical care or when they
are reused for other purposes such as quality management or research.
Episodes Start date Description ICPC 15-01-2016 Cough DD Lung cancer? R84
14
1
Gen
era
l Intr
od
uc
tIon
, aIm
s an
d o
utlIn
e of th
e thesIs
The chest X-ray turns out to be normal and the GP shares this news with her patient in
a consultation by telephone. That day work is hectic and she forgets to remove the lung
cancer code from the EHR episode and to replace it with the appropriate code for this
consultation: the symptom code for “Cough (ICPC-1 code R05)”. The patient, reassured
and relieved, recovers completely in the following weeks.
However, on a sunny Saturday morning early March, the patient visits an out-of-hours
clinic because he sprained his ankle during the first football match of the season. He is
frightened out of his senses when the GP on duty asks him attentively if he has already started
his treatment for lung cancer.
A month later, his anonymized EHR record is sent to a large research database at
the nearby University where investigators are working on various projects such as early
detection of cancer. In this database he is now registered as a lung cancer patient and
researchers assess his consultations, lab results and radiology investigations in the 2 years
prior to the diagnosis. Little does the patient know that coming December he will also
be counted as a “cancer case” in the assessment of the GPs workload by an insurance
company with the aim to determine financial reimbursement and to calculate a quality
indicator to asses quality of cancer care provided by the GP.
Obviously, the GP did not make a mistake in her diagnostic workup with this patient,
however, she did make a mistake during her ICPC coding routine in the first consultation
and again, during the second consultation. The first mistake is that she accepted a
“diagnosis” code for a hypothesis, the second mistake is that she did not correct the code
of the hypothesis in the Episode into a symptom code [2]. Unfortunately, both easy-to-
make mistakes are likely to be common, because of time pressure in daily Primary Care
practice (10 minutes per consultation including expanding registration constraints) and
EHR system design and interface which have not yet fully evolved. These error-prone
coding routines do not disturb the medical process in every-day practice. However,
problems do arise when medical files containing these errors are used by someone else
in the setting of clinical care or when they are reused for other purposes such as quality
management or research.
Reuse of EHR dataComputer scientists and medical informaticists used to shudder at the thought: reuse of
EHR data for other purposes [5]. In the early nineties, van der Lei warned against the reuse
of routine care EHR data for other purposes such as research by launching his fist law of
informatics [6]: “data shall only be used for the purpose for which they were collected”.
In his opinion, medical data is recorded for a specific purpose and that purpose has an
influence on what data is recorded and how, thereby limiting its quality and usability for
other purposes[7]. This is illustrated in the case scenario described above: a registration
that makes perfect sense to GPs in everyday practice but can be interpreted completely
wrong by re-users of data.
15
1
Gen
era
l Intr
od
uc
tIon
, aIm
s an
d o
utlIn
e of th
e thesIs
Today (2016), however, reuse of EHR data for other purposes than its traditional
role of recording and supporting the healthcare process, is becoming commonplace.
Anonymized Primary Care EHR records are already being re-used as a data source
for purposes of quality assessment and for producing indicators of care for insurance
companies[8,9], for pro-active indicated prevention[10] in specified risk groups[10], and
for research[11,12].
EHR data is potentially valuable to improve care: for instance to identify and
subsequently monitor within practice populations patients with an increased risk of
disease such as cardiovascular events, cancer or cognitive derailment and psychiatric
disorders [13,14]. These patients can be included in pro-active indicated prevention
projects, which are thought to be promising tools in managing the ever-increasing
workload of family physicians. Although studies regarding the effectiveness of indicated
prevention are not as promising as some would have expected[15,16], there are examples
of these projects running successfully such as those for the frail elderly with multi-
morbidity or complex care-needs [17,18].
Policy makers have discovered the possibilities of EHR data collected during routine
care to calculate quality indicators, which are increasingly often being used as a
supposedly accurate basis to assess quality of care [8,9]. For Dutch primary care alone,
over one hundred quality indicators have already been established today and more are
being developed. Because manual assessment of these indicators is a time-consuming
burden for healthcare professionals, policy makers aim for automatic calculation based
on extracted routine care data[9].
There are also substantial benefits of reusing EHR data for research purposes. In
the Netherlands, as in the UK, the US and many other western countries, GPs have
been registering data on their patients electronically for more than 20 years and are
sharing their anonymized patient data with practice-based research network (PBRN)
databases [12,19,20]. Retrospective research possibilities using this voluminous “data
goldmine” that comprises many years of follow-up seem endless. As no patient or
data recruitment is necessary: scientific discoveries can be accelerated while cutting
costs at the same time. Furthermore, studies that would otherwise be difficult
to perform, for instance in the field of rare diseases, or concerning health events
that would otherwise be difficult to capture, could be accomplished with routinely
collected EHR data [21].
Besides reuse of data for purposes of pro-active indicated prevention, for assessment
of quality indicators and for research, EHR data is also increasingly being shared for
the purpose of every-day care. In the Netherlands, data sharing through the National
Switch Point (LSP) project between GPs and out-of-hours clinics or other medical
professionals is being expanded further, despite delays and patient participation
numbers lagging behind.
16
1
Gen
era
l Intr
od
uc
tIon
, aIm
s an
d o
utlIn
e of th
e thesIs
The problemAlthough the potential for re-use of EHR data is huge and examples of re-use in health care,
policy making and research already exist, the concerns that prompted Van der Lei to state
his first law of Informatics many years ago are still valid. Indeed, studies into reusability of
hospital clinical data for research demonstrate that data is often incomplete, incorrect and
inaccurate, including episode list errors and inaccurate diagnostic codes. [22,23]. In contrast,
little is known about EHR data quality and its suitability for reuse and sharing within the
domain of primary care. The scarce literature in this field that reports on only a few aspects
of data quality; mainly ‘completeness’ and ‘correctness’ (model table 1), is giving rise to
concerns. There are signals that data quality is suboptimal [24–26] but information on the
extent of the problem, as well as information on relevant determinants of data quality,
causes for and consequences of suboptimal quality is currently missing. This prompts the
question: are we in fact reusing and sharing a data-goldmine or quite the opposite?
table 1. Dimensions of data quality^
Dimension Description
Completeness A characteristic of Information Quality measuring the degree to which all required data is known
Correctness Conforming to an approved or conventional standard, conforming to or agreeing with fact, logic or known truth
Concordance Or: consistency. The condition of adhering together, the ability to be asserted together without contradiction
Plausibility Or: accuracy to surrogate source. A measure of the degree to which data agrees with an original, acknowledged authorative source of data (in this context: general medical knowledge)
Currency The quality or state of information of being up-to-date and not outdated
^ Definitions from the IAIDQ (International Association for Information and Data Quality (http://iaidq.org/main/glossary.shtml)
aImS anD outLInE of tHIS tHESIS
In summary, EHR data is increasingly being shared and reused and we expect this trend
to persist and expand in the coming years. Currently, we do not know enough about the
quality and resulting reusability of this data for other purposes, and not enough about the
suitability for data sharing as a means to improve every-day care. Therefore, there is a need
to quantify and explore this problem within Primary Care and this is the first aim of this thesis.
We decided to explore a number of dimensions of data quality in Primary Care using various
approaches, including literature study. We focused on diagnosis registry as a central item in
17
1
Gen
era
l Intr
od
uc
tIon
, aIm
s an
d o
utlIn
e of th
e thesIs
the EHR with special attention for disease coding since re-users rely heavily on coded data.
We also wanted to identify bottle-necks in reuse hands-on, thus by actually doing research
with EHR data, not only coded but also free-text. The choice for cancer as a disease under
study in the first part of this thesis is not accidental: there is a reference standard available
through the Netherlands Cancer Registry[27]. Furthermore patient care for cancer-survivors
is partly subject to transition to Primary Care, which means reliable data should be available.
In summary, in the first part of this thesis (figure 1) we aim to:
1. assess data quality in the Primary Care EHR by
y Chapter 2: Studying literature to provide an overview of current knowledge on
data quality in Primary Care;
y Chapter 3: Studying the quality of coded cancer diagnosis registration
(numerical comparison on population level with external reference for three
common cancer types)
y Chapter 4: Assessing diagnosis registry (completeness and correctness of
cancer registration through record linkage to an external reference for four
common cancers)
y Chapter 5: Doing research using coded as well as free-text data concerning
GP management of women with breast cancer related concerns;
In the second part of this thesis we try to contribute to solutions in order to improve data quality
and enable reuse and sharing of EHR data. We decided to broaden our horizons by working
with rare diseases as opposed to common diseases (cancer) in the first part of this thesis, but
also by searching participation with medical specialists (hospital EHRs) and bio-informaticians.
Because we found coding errors to be a major cause for suboptimal data quality, we decided
to focus more on disease coding by actually developing a more complete coding system in the
field of rare diseases and by participating in the development of a coding tool. In the study
described in chapter 5 we discovered a lack of application of available genetics-knowledge in
Primary Care, partly caused by current design and limitations in the EHR. Hence we decided
to develop a practical roadmap on this subject, including items to improve EHR data quality
in Primary Care. In summary, in the second part of this thesis (figure 1) we aim to:
2. Find strategies and solutions to improve quality of EHR data and to contribute to
the enabling of reuse and sharing of EHR data.
y Chapter 6: Explore coding pathways by actually developing a disease coding
system for (rare) metabolic diseases in cooperation with paediatricians;
y Chapter 7: Participate in the development a tool for retrospective coding of text
and mapping of existing coding systems in cooperation with bio-informaticians;
y Chapter 8: Propose practical improvements to the EHR and to coding
systems by developing a “roadmap” to stepwise integration of genetics in family
medicine and clinical research.
18
1
Gen
era
l Intr
od
uc
tIon
, aIm
s an
d o
utlIn
e of th
e thesIs
15
Thes
is:
Reus
e an
d sh
arin
g of
Ele
ctro
nic H
ealth
Rec
ord
Data
– w
ith a
focu
s on
Prim
ary
Care
and
dise
ase
codi
ng
Qual
ity &
Reu
sabi
lity
of D
ata
Hand
s on
Iden
tifica
tion
of b
ottle
neck
s and
are
as fo
r im
prov
emen
t
Stra
tegi
es &
Solu
tions
to e
nabl
e re
use
& sh
arin
g of
EHR
dat
a
Chap
ter 2
Lite
ratu
re re
view
Da
ta q
ualit
y in
the
Elec
tron
ic He
alth
Rec
ord
of th
e GP
Chap
ter 4
Dia
gnos
tic D
ata
Qua
lity
asse
ssm
ent u
sing
rec
ord
linka
ge:
Do GP
s kno
w the
ir ca
ncer
patie
nts?
a rec
ord l
inkag
e stud
y asse
ssing
the
quali
ty of
canc
er re
gistry
in
Prim
ary C
are
Chap
ter 5
Doi
ng re
sear
ch u
sing
co
ded
& fr
ee-t
ext E
HR
data
Pr
imar
y Car
e man
agem
ent o
f wo
men w
ith br
east
canc
er re
lated
co
ncer
ns; a
dyna
mic c
ohor
t stud
y us
in g a
netw
ork d
ataba
se
Chap
ter 3
Qua
lity
& R
eusa
bilit
y of
Cod
ed D
iagn
osis
Regi
stry
Re
usab
ility o
f cod
ed da
ta in
the
Prim
ary C
are E
HR; a
dyna
mic
coho
rt stu
dy co
ncer
ning c
ance
r dia
gnos
es
Chap
ter 8
Im
prov
ing
the
EHR
& C
odin
g Sy
stem
s Pr
opos
ed ro
adma
p to
stepw
ise in
tegra
tion o
f ge
netic
s in f
amily
med
icine
an
d clin
ical r
esea
rch
Chap
ter 6
Dev
elop
men
t of
a D
isea
se C
odin
g sy
stem
A n
ew co
ding s
ystem
for
metab
olic d
isord
ers
demo
nstra
tes ga
ps in
int
erna
t. dise
ase c
lassif
.
Qual
ity o
f Dat
a Lit
erat
ure
Revi
ew
Chap
ter 9
Sum
mar
izing
Di
scus
sion
Tr
ansla
tion
of o
vera
ll re
sults
in
to ch
eckl
ist fo
r reu
se a
nd
data
shar
ing.
Sug
gest
ions
for
impr
ovem
ent o
f dat
a qu
ality
Chap
ter 7
Mat
chin
g an
d en
rich
ing
dise
ase
codi
ng
syst
ems
Deve
lopme
nt an
d us
e of a
tool
for co
ding
syste
ms: p
ilot m
etabo
lic
disea
ses
Figu
re 1
: Str
uctu
re o
f the
The
sis
fig
ure
1. S
truc
ture
of
the
Thes
is
19
1
Gen
era
l Intr
od
uc
tIon
, aIm
s an
d o
utlIn
e of th
e thesIs
We conclude this thesis with a summarizing discussion including a translation of results
into a checklist for EHR data reuse and sharing in chapter 9.
Notes* Whenever in this thesis the GP is referred to as female (“her” or “she”), the reader can
replace this by “his”, or “he” if so preferred
**A “SOAP”-journal entry consists of four data fields. The first is “Subjective” and is used
to register in plain text what the patient describes, such as complaints and the reason
for the encounter. The second is “Objective” and includes the GPs findings; results from
clinical examination and measurements, mostly in text format. The third “Analysis”
is used to register the diagnosis or most important symptom and is coded using the
International Classification of Primary Care version 1 (ICPC-1) 2009 coding system. The
final is “Plan”, comprising the GPs medication prescriptions, referrals to medical specialists
and follow-up appointments. # Code for malignant neoplasm bronchus/lung, from the International Classification of
Primary Care version 1 (ICPC-1) 2009 coding system, published by the WHO (http://www.
rivm.nl/who-fic/cdromthesaurus/Pagerenglish.pdf )
20
1
Gen
era
l Intr
od
uc
tIon
, aIm
s an
d o
utlIn
e of th
e thesIs
REfEREnCES1. Van der Zanden G. Quality Assessment of Medical Health Records using Information
Extraction, Master’s thesis. 2010, Universiteit Twente, Enschede.2. The Dutch College of General Practitoners. Guideline adequate EHR registry. Revised
version 2013. Available at: https://www.nhg.org/themas/publicaties/richtlijn-adequate-dossiervorming-met-het-epd.
3. Bentsen BG. International classification of primary care. Scand J Prim Health Care 1986;4:43–50.
4. WHO. International Classification of Primary Care; ICPC-2. Available at: http://www.who.int.proxy.library.uu.nl/classifications/icd/adaptations/icpc2/en/(December 2013 last accessed)Title.
5. Burnum JF. The Misinformation Era: The Fall of the Medical Record. Ann Intern Med 1989;110:482. doi:10.7326/0003-4819-110-6-482
6. van der Lei J. Use and abuse of computer-stored medical records. Methods Inf Med 1991;30:79–80.http://www.ncbi.nlm.nih.gov/pubmed/1857252 (accessed 23 Mar2015).
7. van der Lei J. Information and communication technology in health care: do we need feedback? Int J Med Inform 2002;66:75–83. doi:10.1016/S1386-5056(02)00039-4
8. Blumenthal D. Meaningful use: an assessment. An interview with David Blumenthal, M.D., National Coordinator for Health Information Technology, Office of the National Coordinator. Interview by Mark Hagland. Healthc Inform 2011;28:40,44.
9. Dentler K, Numans ME, ten Teije A, et al. Formalization and computation of quality measures based on electronic medical records. J Am Med Inform Assoc 2014;21:285–91. doi:10.1136/amiajnl-2013-001921 [doi]
10. Bleijenberg N, Drubbel I, Ten Dam VH, et al. Proactive and integrated primary care for frail older people: design and methodological challenges of the Utrecht primary care PROactive frailty intervention trial (U-PROFIT). BMC Geriatr 2012;12:16. doi:10.1186/1471-2318-12-16 [doi]
11. Terry AL, Chevendra V, Thind A, et al. Using your electronic medical record for research: a primer for avoiding pitfalls. Fam Pract 2010;27:121–6. doi:10.1093/fampra/cmp068 [doi]
12. Dolor RJ, Fagnan LJ. PBRN conference summary and update. Ann Fam Med 2014;12:176–7. doi:10.1370/afm.1631 [doi]
13. Bodenheimer T. The future of primary care: transforming practice. N Engl J Med 2008;359:2086,2089. doi:10.1056/NEJMp0805631 [doi]
14. Neuwirth EE, Schmittdiel JA, Tallman K, et al. Understanding panel management: a comparative study of an emerging approach to population care. Perm J 2007;11:12–20.
15. Kruis AL, Boland MRS, Assendelft WJJ, et al. Effectiveness of integrated disease management for primary care chronic obstructive pulmonary disease patients: results of cluster randomised trial. BMJ 2014;349:g5392. doi:10.1136/bmj.g5392
16. Caley M, Chohan P, Hooper J, et al. The impact of NHS Health Checks on the prevalence of disease in general practices: a controlled study. Br J Gen Pract 2014;64:e516–21. doi:10.3399/bjgp14X681013
17. Chen EH, Bodenheimer T. Improving Population Health Through Team-Based Panel Management. Arch Intern Med (American Med Assoc 2011;171:(2 pages). doi:10.1001/archinternmed.2011.395
18. Loo TS, Davis RB, Lipsitz L a., et al. Electronic Medical Record Reminders and Panel Management to Improve Primary Care of Elderly Patients. Arch Intern Med 2011;171:1552–8. doi:10.1001/archinternmed.2011.394
19. Carey IM, Cook DG, De Wilde S, et al. Implications of the problem orientated medical record (POMR) for research using electronic GP databases: a comparison of the Doctors Independent Network Database (DIN) and the General Practice Research Database (GPRD). BMC Fam Pract 2003;4:14. doi:10.1186/1471-2296-4-14 [doi]
21
1
Gen
era
l Intr
od
uc
tIon
, aIm
s an
d o
utlIn
e of th
e thesIs
20. Peterson KA, Lipman PD, Lange CJ, et al. Supporting better science in primary care: a description of practice-based research networks (PBRNs) in 2011. J Am Board Fam Med 2012;25:565–71. doi:10.3122/jabfm.2012.05.120100 [doi]
21. Muller S. Electronic medical records: the way forward for primary care research? Fam Pract 2014;31:127–9. doi:10.1093/fampra/cmu009
22. Coorevits P, Sundgren M, Klein GO, et al. Electronic health records: new opportunities for clinical research. J Intern Med 2013;274:547–60. doi:10.1111/joim.12119
23. Danciu I, Cowan JD, Basford M, et al. Secondary use of clinical data: the Vanderbilt approach. J Biomed Inform 2014;52:28–35. doi:10.1016/j.jbi.2014.02.003
24. Boggon R, van Staa TP, Chapman M, et al. Cancer recording and mortality in the General Practice Research Database and linked cancer registries. Pharmacoepidemiol Drug Saf 2013;22:168–75. doi:10.1002/pds.3374
25. Nielen MMJ, Ursum J, Schellevis FG, et al. The validity of the diagnosis of inflammatory arthritis in a large population-based primary care database. BMC Fam Pract 2013;14:79. doi:10.1186/1471-2296-14-79
26. Hammad TA, Margulis A V, Ding Y, et al. Determining the predictive value of Read codes to identify congenital cardiac malformations in the UK Clinical Practice Research Datalink. Pharmacoepidemiol Drug Saf 2013;22:1233–8. doi:10.1002/pds.3511
27. Netherlands CCCT. The Netherlands Cancer Registry. http://cijfersoverkanker.nl (accessed 20 Mar2012).
Quality of Data – Literature Review
CHaPtERQuality of Data in the Primary Care Electronic Health Record
(EHR) System
Annet Sollie
2
Published (as a shortened version) in Huisarts & Wetenschap, august 2013: Annet Sollie. Hoe is de kwaliteit van data in het HIS? Huisarts & Wetenschap. 2013 Augustus. nr 8:403-403
(Published: Huisarts & Wetenschap August 2013)
27
2
Qu
alit
y o
f Da
ta in
the Pr
ima
ry C
ar
e eleCtr
on
iC h
ealth
reC
or
D (eh
r) Sy
Stem
IntRoDuCtIon
In the light of increasing possibilities for re-use and sharing of data we investigated various
quality aspects of routine healthcare data from Primary Care. To explore this, we carried
out a literature study, as described in this chapter.
What is the question from General Practice? Most General Practitioners (GPs) in the Netherlands have been registering the routine
care data of their enlisted patients in an Electronic Health Record (EHR) system for
many years now. However, the use of advanced digital techniques to improve the
reuse of this data, for indicated prevention projects - projects aimed at preventing
the onset of a disease in an individual with an increased risk for that disease –, for
harvesting indicators to monitor quality of care, or for research purposes, is still not
very thriving. Our research question is: what is known about the quality of data in the
Primary Care EHR?
What is the current policy?Routine care data are not being used extensively yet (2012) for purposes other than daily
care. This is explained, among other reasons like suboptimal adherence to registration
guidelines and supposed privacy violation, mainly by insufficient insight into the quality
and thus reusability of this data.
WHat IS tHE RELEvanCE of tHIS ISSuE?
It is important for GPs themselves as well as for other potential re-users to know what
the quality of the data in their EHR is, since we expect routine care data to be reused
more and more for purposes of:
y Indicated prevention: better care at lower cost for example by actively detecting
patients in the primary care EHR with an increased risk of certain diseases, preferably
followed by structured proactive care programs aiming at the identification and
restoration of care gaps.
y Quality management: Increasing use of quality indicators harvested from EHR systems,
thus fully depending on data entered during routine care by GPs for providing the
correct information;
y Research: reuse of existing, readily available routine care data would enable
researchers to avoid building separate databases, provided that advanced coding
and anonymization tools are installed;
y Data Sharing: automated, not anonymized, digital sharing of data for supporting
remote healthcare through for instance a National EHR:
28
2
Qu
alit
y o
f Da
ta in
the Pr
ima
ry C
ar
e eleCtr
on
iC h
ealth
reC
or
D (eh
r) Sy
Stem
WHat Do WE knoW fRom ExIStInG LItERatuRE?
Within the field of information technology, “data quality” is assessed based on a number
of dimensions, on which unfortunately there is no full agreement. The dimensions
mentioned in the literature are the state of completeness, correctness, currency, plausibility
and concordance that make data whether or not appropriate for a specific use. In this
context, completeness and correctness of data are the most commonly used dimensions[1].
We searched the existing (Dutch and English) literature from 2008 – 2012 for publications
addressing the quality of EHR data in the context of primary care (Figure 1) and found
eight articles, most of them focusing on the dimensions of completeness and to a lesser
degree to correctness. We studied the references of these eight studies but did not find
any additional publications.
The available literature shows that quality of data varies among GP practices
and among data categories (demographics, vital signs, laboratory, risk, prescribing
information, allergies/intolerances and diagnoses).
Kristianson et al[2] extracted data from 776 diabetes patients from a Swedish
EHR and found that demographic data (date of birth and gender) was recorded
22
Figure 1: Search-strings, results & criteria for inclusion and exclusion
The available literature shows that quality of data varies among GP practices and among data
categories (demographics, vital signs, laboratory, risk, prescribing information,
allergies/intolerances and diagnoses).
Kristianson et al[2] extracted data from 776 diabetes patients from a Swedish EHR and found
that demographic data (date of birth and gender) was recorded complete and correct for most
patients (94- 100%). He also found that data on the prescription of medication can be trusted
and is useful, because of the ATC Coding system that is being used, although information on
details like correct dosages is frequently missing. Information on vital signs (blood pressure,
pulse and respiratory rate), weight and Body Mass Index (BMI) was registered incomplete
and inconsistent, mostly because of free-text registration. With the exception of data on
diagnoses and drug therapy (medication prescription), few data could be used without
extensive data management.
Kwaliteit AND huisartsinformatiesyste
em OR HIS OR huisarts informatie
systeem OR huisarts
Dutch Literature (NTVG en H&W) Full text 2008 t/m 2012
routine zorg data OR routine zorg gegevens
English Literature (Pubmed) Title & Abstract 2008 t/m 2012
(data AND quality) AND ((electronic AND
record) OR ehr OR emr) AND (general
practice OR gp)
(routine AND care) AND (data AND
quality) AND (general practice OR gp)
Result Search / Selected
NTVG 22/0 H&W 7/0
Result Search / Selected
NTVG 42/0 H&W 77/6
Result Search / Selected
Pubmed 25/8
Result Search / Selected
Pubmed 22/4
After removing duplicates & reading: 3
After removing duplicates & reading: 5
Inclusion Criteria Exclusion Criteria: Research article Opinion About quality of data About quality of care About routine care data Data collected for research / questionnaires Setting: primary care Setting: hospital / other
figure 1. Search-strings, results & criteria for inclusion and exclusion
29
2
Qu
alit
y o
f Da
ta in
the Pr
ima
ry C
ar
e eleCtr
on
iC h
ealth
reC
or
D (eh
r) Sy
Stem
complete and correct for most patients (94- 100%). He also found that data on the
prescription of medication can be trusted and is useful, because of the ATC Coding
system that is being used, although information on details like correct dosages is
frequently missing. Information on vital signs (blood pressure, pulse and respiratory
rate), weight and Body Mass Index (BMI) was registered incomplete and inconsistent,
mostly because of free-text registration. With the exception of data on diagnoses
and drug therapy (medication prescription), few data could be used without extensive
data management.
Fokkens et al[3] extracted the data from 196 diabetes patients from Dutch EHRs
and compared this with data registered in the software application “Diabcare”, a
structured registration program used in many practices for diabetes care in addition to
the EMR. They found that the registration of blood pressure scores quite high on the
dimension of “completeness”. This also applies to results of laboratory investigations.
However, data on risk factors (smoking and weight) and results of eye and foot
examinations were registered less frequently in the EHR system compared to the
Diabcare system.
Australian researchers[4] showed 33 patients their own medical file and asked them to
check for completeness and correctness. The results of this study show that demographic
data and data on allergies was recorded complete and correct for respectively 94% and
61% of the patients. However, 35% of patients found that relevant information was
missing in their EHR and 51% found erroneous information in their EHR.
Two Dutch[5,6] and one Swiss[7] study that focused on diagnosis coding in the EHR
concluded that the quality of diagnosis coding was “encouraging” but could improve
and varied among practices. Visscher et al[6] studied ICPC coding of diagnoses in 311
general practices during 2011/2012 and showed that for 86% of the consultations the
General Practitioner (GP) assigned a meaningful* ICPC code. Akkerman et al[5] found
that the annual incidences and prevalences per 1.000 person-years for several ICPC-coded
diseases in the EHR registries of 153 GPs from the regions of Utrecht and Almere (total
198.000 patients) were comparable with incidences and prevalences from the Second
Dutch National Survey of General Practice (2-DNSGP), which contains EHR data validated
with patient health interviews. Swiss[7] researchers organized a “hot-coding” week
among 24 GPS in 2010 and showed that the mean index of ICPC-codes in relation to the
number of consultations rose significantly (from 1.31 to 1.52 mean ICPC codes per patient
visit). This implies that the currently observed numbers of diagnoses per consultation is
underestimated, suggesting room for improvement.
Another Dutch study[8] by Jabaaij (112.000 patient EHRs from 32 practices) shows that
57-99% of Episodes of care in the EHR, which are used to group consultations, are ICPC-
coded. Consultations are linked to an Episode in 62-100% of cases and prescriptions are
30
2
Qu
alit
y o
f Da
ta in
the Pr
ima
ry C
ar
e eleCtr
on
iC h
ealth
reC
or
D (eh
r) Sy
Stem
linked to an Episode in 33-99% of cases. Again, this study shows remarkable variability
in the use of codes and completeness between GP practices.
ConCLuSIonS
The currently available literature on data quality in the primary care EHR is relatively scarce.
However, literature shows that registry of demographic data and results of laboratory
tests is nearly complete and correct and that registry of medication prescriptions is
correct but not always complete and up-to-date (current). Registration of diagnoses using
ICPC codes is fairly good but not complete, while the registration of vital parameters,
allergies/intolerances, weight, Body Mass Index (BMI) and risk factors is unsatisfactory.
To summarize, studies have focused mainly on assessing the completeness of data in the
EHR, and conclude that for demographic data and coded data (medication prescriptions,
laboratory results and diagnoses) this completeness is ‘fairly’ good. Little is known on
other dimensions of data quality in the primary care EHR as well as on extraction (im-)
possibilities. There is certainly room for improvement.
What is the most important research question? Is the quality of primary care routine care data of an acceptable standard in order to
make reuse of data for purposes of health and research possible?
* Meaningful ICPC codes: codes in the range 01-29 (complaints), 70-99 (diagnoses) + A44
(vaccination), R44 (influenza vaccine) and X37 (cervical smear). Codes A79 (no disease) and
A99 (generalized other / unspecified disease (s)) are not considered to be “meaningful”.
31
2
Qu
alit
y o
f Da
ta in
the Pr
ima
ry C
ar
e eleCtr
on
iC h
ealth
reC
or
D (eh
r) Sy
Stem
REfEREnCES1. Weiskopf NG, Weng C. Methods and dimensions of electronic health record data quality
assessment: enabling reuse for clinical research. J Am Med Inform Assoc 2013;20:144–51. doi:10.1136/amiajnl-2011-000681
2. Kristianson KJ, Ljunggren H, Gustafsson LL. Data extraction from a semi-structured electronic medical record system for outpatients: a model to facilitate the access and use of data for quality control and research. Health Informatics J 2009;15:305–19. doi:10.1177/1460458209345889 [doi]
3. Fokkens AS, Wiegersma PA, Reijneveld SA. A structured registration program can be validly used for quality assessment in general practice. BMC Health Serv Res 2009;9:241. doi:10.1186/1472-6963-9-241
4. Tse J, You W. How accurate is the electronic health record? - a pilot study evaluating information accuracy in a primary care setting. Stud Health Technol Inform 2011;168:158–64.
5. Akkerman A, Verheij T, Veen R, et al. Interactieve medische informatie van het Huisartsen Netwerk Utrecht en de Almere Zorggroep (‘Interacive medical information from the General Practitioners Network Utrecht and the Almere Group of Care’). Huisarts Wet 2008;51:90–5. doi:10.1007/BF03086658
6. Visscher, S. ten Veen, P. Verheij PR. Kwaliteit van ICPC-Codering (‘Quality of ICPC Coding’. Huisarts en Wet j 2012;10:459–459.
7. Chmiel C, Bhend H, Senn O, et al. The FIRE project: a milestone for research in primary care in Switzerland. Swiss Med Wkly 2011;140:w13142. doi:10.4414/smw.2011.13142 [doi]
8. Jabaaij L, Njoo K, Visscher S, Van den Hoogen H, Tiersma W, Levelink H et al. Verbeter uw verslaglegging, gebruik de EPD-scan-h. Huisarts Wet 2009;52:240–6.
Quality and Reusability of Primary Care EHR Data – Hands-on identification
of bottlenecks
CHaPtERReusability of coded data
in the primary care Electronic Medical Record: a dynamic cohort
study concerning cancer diagnoses
Annet Sollie, General Practitioner-in-Training/PhD Fellow, Rolf H. Sijmons, Clinical Geneticist, Professor of Medical Translational Genetics, Charles Helsper, MD, PhD, Epidemiologist, Mattijs E Numans, General Practitioner, Professor of Primary Care
3
Published august 2016 in the International Journal of medical Informatics as: Annet Sollie, Rolf H. Sijmons, Charles Helsper, Mattijs E. Numans. Reusability of coded data in the primary care Electronic Medical Record: a dynamic cohort study concerning cancer diagnoses.
37
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
abStRaCt
objectives To assess quality and reusability of coded cancer diagnoses in routine primary care data.
To identify factors that influence data quality and areas for improvement.
methodsA dynamic cohort study in a Dutch network database containing 250,000 anonymized
electronic medical records (EMRs) from 52 general practices was performed. Coded data
from 2000 to 2011 for the three most common cancer types (breast, colon and prostate
cancer) was compared to the Netherlands Cancer Registry.
measurementsData quality is expressed in Standard Incidence Ratios (SIRs): the ratio between the
number of coded cases observed in the primary care network database and the expected
number of cases based on the Netherlands Cancer Registry. Ratios were multiplied by
100% for readability.
ResultsThe overall SIR was 91.5% (95%CI 88.5 – 94.5) and showed improvement over the years.
SIRs differ between cancer types: from 71.5% for colon cancer in males to 103.9% for
breast cancer. There are differences in data quality (SIRs 76.2% – 99.7%) depending on
the EMR system used, with SIRs up to 232.9% for breast cancer. Frequently observed
errors in routine healthcare data can be classified as: lack of integrity checks, inaccurate
use and/or lack of codes, and lack of EMR system functionality.
ConclusionsRe-users of coded routine primary care Electronic Medical Record data should be aware
that 30% of cancer cases can be missed. Up to 130% of cancer cases found in the EMR
data can be false-positive. The type of EMR system and the type of cancer influence the
quality of coded diagnosis registry. While data quality can be improved (e.g. through
improving system design and by training EMR system users), re-use should only be taken
care of by appropriately trained experts;
38
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
IntRoDuCtIon
Reuse of electronic medical record (EMR) data is a hot topic, not only in hospitals[1,2] but
also in primary care[3]. An example is the international trend to calculate quality indicators
automatically based on data collected during routine care. For Dutch primary care alone,
over one hundred quality indicators are established and more are being developed[4].
Because manual assessment of these indicators is a time-consuming burden for healthcare
professionals, policy makers aim for automatic calculation based on extracted, mainly
coded, routine care data [5,6].
Furthermore, risk assessment for prevention projects, followed by structured panel
management procedures as well as chronic disease management to improve proactive
care[7–10], are becoming more and more popular. These are thought to be promising
tools in managing the increasing workload of family physicians, but again they rely
strongly on the analysis of routine care diagnostic data to identify patients who could
be included in preventive care and chronic disease management programmes, such as
the frail elderly[11] or cancer patients.
Also reuse of data for primary care research purposes such as early detection of
cancer is almost becoming commonplace, as is demonstrated by the rapidly evolving
practice-based research networks (PBRNs) in Europe, Canada and the USA. These
networks provide a basic facility for primary care research, and often use anonymized
data uploaded by participating practices to a central database[12–14].
It is important that primary care organization regard their (routine care) data as
a significant and valuable organizational asset. It is equally important that they realize that
in the wrong hands (personnel without appropriate expertise and training in handling
of routine care data), data re-use can actually cause harm. In order to truly value routine
primary healthcare data and to re-use this data reliably, the data should represent the
true situation as closely as possible. Despite the examples of actual reuse mentioned
above, there are serious concerns about the quality and subsequent reusability of EMR
data in primary care[1,2,15–17].
In medical informatics, data quality is assessed using various “dimensions”[15,18].
Although there is no uniform accepted model or method to assess data quality in primary
care, Gray Weiskopf [15] summarized the five common dimensions of data quality based
on extensive literature research: Completeness, Correctness, Concordance, Plausibility
and Currency (figure 1).
39
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
figure 1. Dimensions of data quality
Dimension Description
Completeness No missing data; a truth about a patient is present in the EHR
Correctness An element that is present in the EHR is true
Concordance There is agreement between elements in the EHR or between the EHR and other data sources
Plausibility An element in the EHR makes sense in light of other knowledge about what that element is measuring or representing
Currency An element in the EHR is a relevant representation of the patient state at a given point in time
Little has been published on the quality of data from primary care records. A few studies
(see Discussion) have assessed their completeness and, to a lesser extent, their correctness,
but information on other dimensions is lacking.
When assessing data quality, focusing on the coded registration of diagnoses has
most priority, because this is a central item being used in analyses. We focus on diagnoses
of cancer because it is a high impact diagnosis that we expect to be registered and coded
correctly in the EMR for purposes of care. Furthermore, the national Netherlands Cancer
Registry (NCR)[19] provides an accessible and supposedly reliable reference standard. To
assess quality and usability of coded cancer diagnoses for re-use using available reference
data we decided that we could assess and study three dimensions of data quality using
our data infrastructure: completeness, correctness and concordance with the reference
standard. To find focus points for improvement, we identified factors that influence
data quality.
mEtHoDS
Design We performed a dynamic cohort study in a Dutch network database containing 250,000
anonymized electronic medical records (EMRs) from 52 general practices. We used
a 4 step study approach, as described in Figure 2, to determine Standardized Incidence
Rate Ratios (SIRs) between January 1st 2000 and December 31st 2011.
First, we determined our reference standard: the expected incidence rates based on
the Netherlands Cancer Registry (NCR) [19] and Statistics Netherlands[20].
Second, observed incidence rates of cancer in coded routine primary care data were
determined using the Julius General Practitioners’ Network (JGPN), for patients with
a diagnosis of breast-, prostate-, and/or colon cancer.
40
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
1d Calculate expected number of cases for each
cancer in study population
Statistics Netherlands Dutch National Cancer Registry Study population inJGPN database
1a Download total population, male/female, 2000 – 2011
1b Download number of breast, prostate and colon cancer cases
in 2000 - 2011
1c Calculate number of person years in JGPN
database for 2000 - 2011
Observed cases
2a Select patients with cancer episode ICPC code
2c Select patients without cancer episode but with cancer consultation ICPC
code
2d Select patients without cancer ICPC code but with
cancer medication
2b Work-up data: manual correction of diagnosis
and/or date of diagnosis
Observed cases,
corrected
Observed cases, corr & consult
Observed cases, corr
& med
Expected cases
3 Calculate
SIR & 95%CI
SIRs and 95%CIs
Figure 2: Flowchart Methods
figure 2. Flowchart Methods
Third, we calculated the SIRs for three four-year time slots between January 1st 2000
– December 31st 2011, for each EMR system, for the three cancer types and for overall
cancer diagnoses. Finally, we identified, listed and classified the common errors we
found in the extracted EMR data.
Patients & EMR Data: the JGPN databaseThe JGPN database[11,21] comprises anonymized routine care data, updated
quarterly from 52 GP practices (120 GPs, 250,000 patients) that share their data
with the Julius Centre for Health Sciences and Primary Care, University Medical
Centre Utrecht, the Netherlands. This population is considered representative for
the Dutch population (table 1)[22]. GP’s were not aware of this study at the time
of registry, neither did they receive specific training on coding or registry in the
EMR nor was the data improved in any way. Hence the data can be considered as
regular routine care data.
41
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
table 1. Representativeness of patients in GPs in the JGPN in 2013
Patients
JGPn netherlands
Total N:231,556 Total N:16,779,575
male sex n (% male of total)
110,973(48%)
8,307,339(50%)
% age < 20 of total (% male of group)
21%(51%)
23%(51%)
% age 20-65 of total (% male of group)
65%(48%)
60%(50%)
% age > 65 of total (% male of group)
14%(43%)
17%(45%)
The JGPN data include consultations and clustered disease episodes with ICPC-coded[23]
diagnoses, ATC (Anatomical Therapeutic Chemical)-coded prescribed medication, coded
laboratory test results, and for some patients coded referrals and letters from medical
specialists.
In the Netherlands GP medical encounters are registered according to the “SOAP-
system”[24]. A SOAP-journal consists of four data fields. The first is “Subjective” and
comprises patient complaints and the reason for the encounter, registered in free
text fields. The second is “Objective” and includes results from clinical examination
and measurements, also mostly in free text format. The third “Analysis” is used to
register the diagnosis or most important symptom and is coded using the International
Classification of Primary Care version 1 (ICPC-1) 2009 coding system[23]. The final
datafield is “Plan”, comprising the GPs medication prescriptions, referrals to medical
specialists and follow-up appointments. The list “Episodes”, also coded using ICPC-1
codes, clusters consultations concerning the same diagnosis for an individual patient.
This list provides an overview of active and non-active diagnoses and complaints with
corresponding start and/or end dates.
ICPC diagnosis codes are available for the more common types of cancer, including
the three cancers under study. There are no codes available to specify recurrence of
cancer, for suspected cancer or for treatments of cancer. The GP manually enters the
ICPC code for the concerning ‘cancer’ diagnosis during consultation or after receiving
confirmation of the diagnosis from secondary care correspondence. The GP decides
when a new Episode is created for the cancer diagnosis and which consultations
are added to this Episode. According to the Dutch College of General Practitioners’
guideline for correct registration, every cancer diagnosis should be registered as an
episode in the EMR and consultations about relevant complaints or treatments should
be attached to it [25].
42
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
Reference dataAll oncology-related information from hospitals, as triggered by the pathology report
for newly found cancers, is sent to the NCR[19]. Specially trained staff members enter
relevant data about these diagnoses in the registry. Also, cancer diagnoses reported
in hospital patient discharge files, for which no pathological investigation is being
performed, are included in the NCR as clinical diagnoses for most hospitals. The NCR
claims to be almost complete (>95% of all cancers) for the population of the Netherlands
and without false-positive records since 1989. However, no solely in primary care
diagnosed cancer cases are entered into the NCR registry. The registration delay at the
NCR is reported to be 3 - 9 months after the pathologist confirmed the cancer and is
claimed to be decreasing. There is some evidence that the quality of the NCR data is
complete and accurate[26,27].
Reference data on the size of the Dutch population were obtained from Statistics
Netherlands[20], which is responsible for collecting and processing data for the official
national statistics to be used by policymakers and scientific researchers.
Data collection and analysis
Step 1: Calculation of expected number of cases from reference data
The expected incidences in the study population were calculated using steps 1a to 1d,
as demonstrated in Figure 2. First, to determine the size of the reference population,
the total number of males and females between 20 – 90 years on January 1st of each
year from 2000 to 2011 were downloaded from Statistics Netherlands[20] in 5-year age
categories. The mean of two consecutive years was used to determine the size of the
population on July 1st of each year. Next, to determine the incidence rates for cancer, the
absolute national number of the three types of cancer patients under study and aged
20 – 90 years was obtained from the NCR[19] for each of the years 2000 – 2011, also
in 5-year age categories. We included invasive and non-invasive (in situ) breast cancer
(females only), colon cancer (separately for males and females) and prostate cancer
(males). Next, to determine the size of the study population, the number of person-years
in the JGPN database was calculated from the number of registered patients per age
category for each EMR system on July 1st of each year. Finally, the expected incidences
in our study population were calculated per category per EMR system per year using
the following formula:
(absolute number of cancer cases / population) * number of person-years in database.
Step 2: Determining the observed number of cases: selecting patients
from study population, data extraction/work-up
To obtain the observed incidences, we used a three step search procedure to select all
cases of breast- colon-, and prostate cancer cases from 2000 to 2011 from the JGPN
43
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
database (Figure 2). These were counted per EMR system, per 5-year age category and
per year. Practices were using one of the following EMR systems: Medicom, Promedico
or MicroHIS. There are more than 10 EMR systems available for general practices in the
Netherlands, but these three are the most frequently used in the country as well as in
the JGPN-affiliated practices. Because of low numbers of users and patients we decided
not to use data of additional practices in the database using other EMR systems. The
data models of all Dutch Primary Care EMR systems are based on the frequently updated
Reference Model issued by the Dutch College of General Practitioners[28] and they differ
mainly in user interface.
First (step 1), we extracted relevant data for all patients with one or more ICPC
episode codes for breast- (X76 and X67.1), colon- (D75), and/or prostate cancer
(Y77 and Y77.1). This included year of birth, gender and cancer episodes with their
ICPC code and description. We continued by removing duplicate records and then
manually checking the episode description entered by the GP along with the registered
ICPC code and date. This manual check was performed by the first author, who is
researcher, ICT expert and GP and took about 30 hours. If the information in the
episode description was sufficiently clear, corrections were made, mainly in the date
of diagnosis (for instance: Episode date 2011, description “Breast cancer diagnosed
and treated in 2002”, date of diagnosis was changed into 2002). If the ICPC code
clearly did not match the description (for instance code Breast Cancer X76, description
“mother has had breastcancer”), the record was excluded. We counted the number
of errors (corrections and exclusions) per EMR system and per time period. Patients
reported as having recurrent cancer (by the GP, in the text description) were counted
only at the occurrence of the primary cancer, in line with NCR prevalence reports.
Patients with “cancer in medical history”, with or without further specification, were
excluded because the original date of diagnosis was considered to occur before our
observation period.
After this procedure, we concentrated on finding missing episode-diagnoses (step 2)
by selecting all the patients without a coded cancer-related episode but who were
registered with one or more registered consultations with an ICPC code for breast- (X76,
X76.1), colon- (D75), and/or prostate (Y77, Y77.1) cancer. Extracted data included year
of birth, sex, as well as all other encounter SOAP registries, including other ICPC codes.
Finally (step 3) we continued to check all patients without an episode or consultation
coded for a relevant cancer type but who were prescribed specific cancer drugs (3)
from 2000 – 2011. This procedure should theoretically result in finding all the patients
diagnosed with cancers in our observation period. Note that we performed the three
step search procedure for the total JGPN population but only present the data for adults
to compare with the reference population (age at diagnosis between 20-90) in the
results section.
44
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
figure 3. List of drug prescriptions used in EMR search of cancer cases
Cancer type name of drug atC code*
Breast cancer Gosereline/zoladex L02AE03
Tamoxifen/nolvadex/istubal/valodex L02BA01
Anastrozol/arimidex L02BG03
Letrozol/femara/letroman L02BG04
Exemestaan/aromasin L02BG06
Trastuzumab/Herceptin/herclon L01XC03
Bevacizumab/Avastin L01XC07
Colon cancer Cetuximab/erbitux L01XC06
Panitumumab/vectibix L01XC08
Bevacizumab/Avastin L01XC07
Prostate cancer Gosereline/zoladex L02AE030
*ATC = Anatomical Therapeutic Chemical coding system
Step 3: Calculation of standard incidence ratios
SIRs[29] were calculated as the ratio between the observed and expected number
of cancer cases, for our three cancer types combined and for the different types
separately, differentiated per four-year period and per EMR system. All analyses
were stratified by sex. Because the differences between the observed and expected
number of cases may be due to random fluctuations in disease occurrence, 95%
confidence intervals (CIs) were computed assuming Byar’s approximation[29]. If the
95%CI includes 100%, the difference between the observed and expected cases is
likely to have occurred by chance. All data analysis and calculations were carried out
using Microsoft Excel 2010.
RESuLtS
The combined SIR for breast, colon, and prostate cancer between 2000 and 2011 was
91.5%, (95%CI 88.5 – 94.5). This means there is a significant difference between the
observed number of cases in the EMR and the expected number according to the NCR
(table 2).
45
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
table 2. Quality of diagnosis registry in the EMR
Expected cases observed cases* SIR** 95%CI of SIR***
n n % %
overall Cancer 3926 3594 91.5% 88.5 - 94.5
Follow-up period
2000 - 2003 incl 1010 670 66.3% 61.3 - 71.3
2004 - 2007 incl 1294 1239 95.7% 90.3 - 100.9
2008 - 2011 incl 1623 1685 103.8% 98.8 - 108.6
EMR system
Promedico 2132 2126 99.7% 95.4 - 103.9
Medicom 574 537 93.7% 85.4 - 101.1
MicroHIS 1221 931 76.2% 71.3 - 81.0
Cancer type
Breast Cancer Female 1685 1750 103.9% 98.9 - 108.5
Colon Cancer Female 599 476 79.5% 72.2 - 86.3
Colon Cancer Male 662 473 71.5% 65.0 - 77.8
Prostate Cancer Male 981 895 91.2% 85.1 - 97.0
* Observed cases after work-up** SIR is the Standard Incidence Ratio and is the ratio between observed and expected cases, expressed as a percentage, numbers in bold print are statistically significant*** If the 95%CI includes 100.0, the difference between the observed and expected number of cases is not statistically significant
The SIRs varied over time: from 2000 to 2003 the combined SIR was 66.3%, (95%CI
61.3 – 71.3), from 2004 to 2007 it was 95.7% (95%CI 90.3 – 100.9), and from 2008
to 2011 it was 103.8% (95%CI 98.8 – 108.6). For colon cancer in males the SIR was
71.5% (65.0 – 77.8), while in women it was 79.5% (95%CI 72.2 – 86.3). The SIR for
prostate cancer was 91.2% (85.2 – 97.0) and for breast cancer 103.9% (98.9 – 108.5).
Furthermore, a statistically significant deviation from the expected number of cases was
found in the combined SIR for the MicroHIS EMR system: 76.2% (71.3 – 81.0). Almost
every SIR for the Promedico system was over 100% in recent years, which means that
a higher than expected number of cases (indicating false positive diagnoses) was found
in the EMR (S1 table 3).
46
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
S1 table 3. Quality of diagnosis registry stratified to cancer type and EMR system (Continued)
Expected cases
n
observed cases*
nSIR**
%
95%CI of SIR#
%
observedCases after
Cons***n (% extra)
SIR**%
95%CI of SIR#
%
breast Cancer female
Promedico
2000 – 2003 262.1 232 88.5% 76.8 – 99.2 449 (94%) 171.3% 154.1 – 185.1
2004 – 2007 292.7 350 119.6% 106.4 – 131.0 674 (93%) 230.2% 211.0 – 244.9
2008 – 2011 337.0 411 121.9% 109.5 – 132.7 785 (91%) 232.9% 214.9 – 246.8
Medicom
2000 – 2003 37.7 58 153.7% 108.3 – 180.2 59 (2%) 156.3% 110.4 – 182.9
2004 – 2007 80.8 97 120.0% 94.1 – 139.7 99 (2%) 122.5% 96.2 – 142.2
2008 – 2011 141.5 124 87.7% 71.6 – 101.6 291(135%) 111.9% 178.6 – 224.5
MicroHIS
2000 – 2003 135.8 108 79.5% 64.1 – 93.3 110 (2%) 81.0% 65.4 – 94.9
2004 – 2007 174.1 150 86.2% 71.9 – 98.9 156 (4%) 89.6% 75.0 – 102.5
2008 – 2011 223.5 220 98.4% 84.9 – 110.4 232 (5%) 103.8% 89.9 – 116.0
Colon Cancer female
Promedico
2000 – 2003 92.9 41 44.1% 31.3 – 57.5 55 (34%) 59.2% 43.7 – 73.9
2004 – 2007 109.4 98 89.6% 71.1 – 105.4 98 (0%) 89.6% 71.1 – 105.4
2008 – 2011 129.4 159 122.9% 102.3 – 139.3 182 (15%) 140.7% 118.3 – 157.8
Medicom
2000 – 2003 10.4 6 57.6% 19.4 – 91.8 6 (0%) 57.6% 19.4 – 91.8
2004 – 2007 25.1 17 67.7% 36.8 – 94.0 18 (6%) 71.7% 39.4 – 98.2
2008 – 2011 45.2 33 72.2% 48.0 – 94.4 33 (0%) 72.2% 48.0 – 94.4
MicroHIS
2000 – 2003 42.3 25 59.1% 36.8 – 79.9 25 (0%) 59.1% 36.8 – 79.9
2004 – 2007 61.3 34 55.4% 37.4 – 72.9 34 (0%) 55.4% 37.4 – 72.9
2008 – 2011 82.5 63 76.4% 57.1 – 93.3 66 (5%) 80.0% 60.2 – 97.2
Colon Cancer male
Promedico
2000 – 2003 97.4 39 40.0% 28.2 – 52.6 46 (18%) 47.2% 34.1 – 60.6
2004 – 2007 120.8 104 86.1% 68.9 – 101.0 114 (10%) 94.3% 76.2 – 109.8
2008 – 2011 146.9 138 93.9% 77.6 – 108.1 157 (14%) 106.9% 89.2 – 121.7
Medicom
2000 – 2003 11.1 9 80.9% 31.8 – 114.3 9 (0%) 80.9% 31.8 – 114.3
2004 – 2007 28.4 14 49.4% 25.9 – 72.8 15 (7%) 52.9% 28.3 – 76.7
2008 – 2011 54.2 42 77.5% 53.6 – 97.7 42 (0%) 77.5% 53.6 – 97.7
47
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
S1 table 3. Quality of diagnosis registry stratified to cancer type and EMR system (Continued)
Expected cases
n
observed cases*
nSIR**
%
95%CI of SIR#
%
observedCases after
Cons***n (% extra)
SIR**%
95%CI of SIR#
%
MicroHIS
2000 – 2003 42.6 21 49.3% 29.6 – 69.1 21 (0%) 49.3% 29.6 – 69.1
2004 – 2007 66.5 32 48.1% 32.3 – 64.2 32 (0%) 48.1% 32.3 – 64.2
2008 – 2011 93.7 74 79.0% 60.5 – 95.2 78 (5%) 83.3% 64.2 – 99.7
Prostate Cancer
Promedico
2000 – 2003 140.2 90 64.2% 50.9 – 76.8 114 (27%) 81.3% 65.9 – 95.0
2004 – 2007 186.3 239 128.3% 110.8 – 142.6 253 (6%) 135.8% 117.8 – 150.4
2008 – 2011 216.3 225 104% 89.8 – 116.4 243(8%) 112.3% 97.4 – 125.1
Medicom
2000 – 2003 15.2 14 92% 44.0 – 123.3 14 (0%) 92% 44.0 – 123.3
2004 – 2007 43.9 47 106.9% 74.2 – 130.8 47 (0%) 106.9% 74.2 – 130.8
2008 – 2011 79.9 76 95.1% 72.6 – 113.5 78 (3%) 97.6% 74.7 – 116.1
MicroHIS
2000 – 2003 121.9 27 22.2% 14.8 – 31.2 27 (0%) 22.2% 14.8 – 31.2
2004 – 2007 104.8 57 54.4% 40.5 – 67.9 58 (1%) 55.3% 41.3 – 69.0
2008 – 2011 72.0 120 166.6% 132.6 – 188.9 132 (10%) 183.2% 147.1 – 206.1
* Observed cases after work-up** SIR is the Standard Incidence Ratio and is the ratio between observed and expected cases, expressed as a percentage, numbers in bold print are significant*** Observed cases after work-up and adding patients without episode ICPC for cancer but with journal consultations coded as cancer # If the 95%CI includes 100.0, the difference between the observed and expected number of cases is not statistically significant
Note that all data presented is after work-up, which was performed by manual checking
(see methods). This led to records being excluded because of erroneous coding in 6% of
cancer cases found in Promedico, 10% in Medicom and 5% in MicroHIS.
Corrections in date of diagnosis were made in 11% of cancer cases in Promedico,
14% in Medicom and 6% in MicroHIS. There were no trends visible. In total, work-up
was necessary in 17% of the Promedico cancer cases found, 24% of the Medicom cases
and 11% of the MicroHIS cases.
Frequent errors encountered in the EMR during work-up could be classified as follows:
lack of integrity checks, inaccurate use or lack of ICPC codes, lack of EMR functionality,
and inaccurate registration of the date of diagnosis (S2 table 3).
48
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
S2 table 4. List of frequent errors in the EMR registries classified by type
no.type of inaccuracy
Dimension of quality Examples
1 Lack of integrity checks in EMR
Concordance •X76 (malignant neoplasm breast female) given to a male patient
•X76 given to a very young patient•Two or more episodes for the same disease
2 Inaccurate use of ICPC codes
Correctness •X76 for a patient with a tick bite or D75 for malignant neoplasm bladder (wrong coding)
•X76 for suspicion of breast cancer (e.g. lump)•See also examples under 3
3 Lack of ICPC codes
Correctness •D75 for anal cancer (no ICPC code available)•For genetic risk (cancer or known mutated gene in family,
gene carrier, or gene tested negative)•Two disease episodes for bilateral breast cancer
(no ICPC code available)
3 Lack of EMR functionality
Completeness •Use of disease registry (ICPC codes) for crucial symptoms & signs (rectal bleeding, increased PSA)
•Use of disease registry (ICPC codes) for further investigation (mammography, colposcopy) and periodic investigation (e.g. annual preventive screening)
•Use of disease registry (ICPC codes) for treatment•Use of disease registry (ICPC codes) for preventive surgery
(mastectomy) and breast reconstruction•No possibility to register relapsing cancer•No possibility to enter cancer staging (TNM)•No functionality to register suspicion of disease or
exclusion of disease
8 Date of diagnosis incorrect or missing
Currency •Registration of “breast cancer 1993”with data 1-1-2010•Registration of “colon cancer in medical history” on
1-3-2009
Adding patients without an episode ICPC code for cancer, but with one or more journal
consultations with a cancer code, increased the total number of cases found by 0 – 7%
for MicroHIS and Medicom (table 5). For Promedico this was up to 27%. For breast cancer,
however, over 90% more cases were found in Promedico in the follow-up periods. In
Medicom the increase in number of breast cancer cases was 135% between 2008 – 2011,
counting patients with journal consultation cancer codes.
The number of cases found by searching the EMR for cancer medication was low and
added only to breast- and prostate cancer numbers. Because the date of prescription
was not available for all the records, only the total number of cases found from 2000 to
2011 were counted (Table 5).
49
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
table 5. Number of cases found by searching EMR for cancer medication prescriptions
Expected cases
observed cases*
n
medication casesm (%)
observed cases including
medication cases
overall Cancer 3926.4 3594 164 (4.5%) 3758
Breast Cancer Female 1685.3 1750 105 (6.0%) 1855
Promedico 891.9 993 33 1026
Medicom 260.0 279 20 299
MicroHIS 533.4 478 52 530
Colon Cancer Female 59.1 476 1 (0.2%) 477
Promedico 331.7 298 0 298
Medicom 81.2 56 0 56
MicroHIS 186.1 122 1 123
Colon Cancer Male 661.5 473 0 (0%) 437
Promedico 365.1 281 0 281
Medicom 93.7 65 0 65
MicroHIS 202.7 127 0 127
Prostate Cancer Male 980.6 895 586 (6.4%) 953
Promedico 542.8 554 16 570
Medicom 139.1 137 26 163
MicroHis 298.7 204 16 220
DISCuSSIon
Principal findingsThe overall SIR was 91.5% (95%CI 88.5 – 94.5). Comparability of incidence rates improved
significantly over the years, from a SIR of 66.3% in 2000 – 2003 (95%CI 61.3 – 71.3) to
103.8% in 2008 – 2011 (95%CI 98.8 – 108.6). SIRs differ between cancer types: from 71.5%
(95%CI 65.0 – 77.8) for colon cancer in males to 103.9% (95%CI 98.9 – 108.5) for breast
cancer. There are differences in data quality (SIRs 76.2% – 99.7%) depending on the EMR
system used, with SIRs up to 232.9% for breast cancer in one EMR system in recent years.
The most frequent errors in the EMR can be classified as: lack of integrity checks,
inaccurate use and/or lack of codes, and lack of EMR functionality.
Strengths and limitationsTo the best of our knowledge, this is the first study to assess data quality in primary
care EMRs, taking into account the type of diagnosis, adaptation time and type of EMR
50
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
system. The major strengths of our study are the size of the cohort, the extensive EMR
data content we had available that was sheer routine care hence not improved in any way
and the use of SIRs to assess data quality. Furthermore, the JGPN database comprises
well-documented information on its enlisted patients. The characteristics of these patients
do not differ from the overall Dutch population (table 1)[22]. Furthermore we used only
coded data (ICPC codes and medication) to select cancer diagnosis from the EMR and
chose not to exploit the free-text part of the EMR. Although this may be viewed as a
limitation, current re-use of EMR data is often based on coded data only, which means
our study provides an insight that could be useful.
A limitation is that we chose to work with coded diagnoses extracted from the
EMR which were not validated using information from the EMR other than the Episode
description. This means false-positive diagnosis could have been included and false-
negative diagnosis could have been missed. Furthermore, we assumed our reference
standard to be correct (see methods). If this is not the case, this may have biased our
results. Last but not least working with routine care data has its limitations that should
be taken into account, such as variance in registry between GPs despite the availability of
a guideline for adequate registry issued by the Dutch College of General Practitioners[25]
for a number of years now.
Comparison with the literatureTse et al[30] asked 33 Australian patients to check their own medical files for
completeness and accuracy; 35% of them found relevant information missing from
their EMR and 51% found incorrect data. Two Dutch[31,32] studies and one Swiss[33]
study investigated the coding of diagnoses in primary care EMRs; they all concluded
that the quality of coding in general was fairly good but that it varied widely between
general practices. Visscher et al.[32] found a meaningful (i.e. no non-specific) ICPC
code for 86% of consultations at 311 general practices in the Netherlands. Akkerman
et al.[31] concluded that the incidence and prevalence data for 28 diseases (including
colon- and kidney cancers), registered by 135 GPs in the Netherlands, were not
significantly higher or lower than the data from Visscher and Ten Veen[21], which
were also from a GP registry.
A study by De Clercq[34] et al compared routine care data from over 10,000 Belgian
patients, extracted from GPs’ EMRs and their answers to an electronic questionnaire. They
studied ten healthcare conditions using clinical and biological parameters (cholesterol,
blood pressure, and body mass index), diagnoses (hypertension, diabetes, and past
cardiovascular events), and drug prescriptions (anti-diabetic drugs, aspirin, statins, and
anti-hypertensive drugs). They found a relatively fair agreement (Kappa≥0.40) between
the two data collection methods for seven conditions, but no agreement for the biological
parameters. The prevalence of diagnoses and drug prescriptions was relatively lower in
the EMR data than in that collected by the questionnaire.
51
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
Pascoe et al[16] recruited five GP centres in the UK for a retrospective analysis of
primary care medical records on registration of cancer diagnoses compared to their
regional cancer registry. One in five of all primary care patients with cancer was not
identified when a search for all patients with cancer was conducted using codes for
malignancy, while one in five patient records with a code for malignancy that was
confirmed in the cancer registry also lacked the necessary documentation to verify the
cancer type, date of diagnosis, or any other aspect of the cancer. Overall, codes for cancer
in these EMRs had a poor level of completeness (29.4%) and correctness (65.6%) when
the UK Cancer Registry was used as the reference. These results, including the possible
false-positive registry of cancer diagnoses, are in line with our study.
In Canada, the use of EMRs in primary care is relatively low. Terry et al[17] studied
the EMR as a potential source of data for research and concluded that care must be
taken when using the EMR data for research purposes. They concluded that more time
is needed for EMR training and to standardize how data are registered.
In summary, the limited literature on data quality in primary care EMRs shows that the
registration of diagnoses is reasonably, but often incomplete when compared to external
sources. Prevalence’s are generally lower in the EMRs and relevant data is sometimes
missing. This is in line with our own findings, although we found improvement over time.
For recent years prevalence’s of cancer diagnoses can be higher than expected based on
external reference data, indicating registry of false-positive diagnoses.
Meaning and impactThe quality of coded diagnosis registry in primary care has improved in recent years (2008
and up) but re-users of data have to understand the pitfalls. Quality and usability of
coded cancer diagnoses in routine care data remains substandard. Re-use should only be
performed by people appropriately trained and with expertise in data management and
analysis. It is important that primary care workers that enter data routinely are involved
in this process for adequate interpretation of data.
Work-up of data (manual checking small parts of the EMR and correction in 11 – 24%
of cases) improves the quality of routine care data. Be aware that 30% of cases can be
missed when searching for cancer diagnoses even in recent years and be aware that
there are differences in the quality of registration of diagnoses between cancer types
and types of EMR system used. For the diagnoses of breast- and prostate cancers, a large
number of false-positive cancer cases can be found. Depending on the purpose of reuse,
this means that it is probably necessary to validate the cancer diagnoses for instance
by linking records with an external data source such as the National Cancer Registry or
a hospital database to eliminate false-positive cases.
Searching for extra cases by checking for journal consultation codes or prescribed cancer
medication seems to be ineffective, since relatively few extra cases were found. When extra
cases are being found, they tend to lead to a higher than expected incidence (false-positive).
52
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
The differences between types of cancer and types of EMR are remarkable: for colon
cancer, especially in Medicom and MicroHIS, the SIRs are relatively low (72.2% – 79.0%,
both significant), which suggests that GPs fail to register colon cancer cases in these
systems. For prostate cancer, SIRs in recent years are good, but higher than expected
incidence rates (possible false-positives) are found in MicroHIS (166.6%). Likewise there
seems to be over-registration of all cancer types in Promedico in recent years. The most
probable explanation is that GPs register a clinical suspicion of cancer with an ICPC
code for cancer. A less likely explanation is that the NCR is less complete than previously
demonstrated[19] Also, the differences between EMRs suggests that EMR system design
directly influences data quality.
Furthermore, MicroHIS has the lowest SIR overall (76.2% 95%CI 71.3 – 81.0), since
it does not force its users to choose an ICPC code for every consultation and episode
(unlike Promedico and Medicom). It also needs the lowest number of corrections (6%).
This suggests that although obligatory coding for each encounter results in a more
complete registry, it also leads to more registration errors.
The list of frequent errors we encountered in the EMR systems (S2 table 4) reveals
ways to improve data quality. EMR software suppliers should extend their standard
procedures for integrity checks and provide EMR data entry fields for crucial symptoms,
disease recurrence, for common tests and treatments, and for familial or genetic risk
factors that might be the main reason for consulting the doctor but do not necessarily
mean that the definite diagnosis has been confirmed. Certain ICPC codes could be
developed to enable better registration of these items. In addition, GPs and their staff
should be trained to make adequate EMR registrations. Also, EMR software designers
should cooperate with interaction designers to discover ways to improve data quality at
entry by enhanced system design. Furthermore, new technology in the area of natural
language recognition could be incorporated into an EMR to help users by suggesting
structured data entry options[35].
Unanswered questions and future researchAlthough with this study information on data quality has been produced on a population
level, the next step would be to assess data quality on an individual level. Therefore
the possibilities for an individual patient-based linkage of records between the EMR
and National Cancer Registries, first in an anonymized research setting but eventually
as a dynamic link to everyday practice, are worth investigating. Future research should
furthermore focus on assessing the quality and reusability of data for other parts of the
EMR and on assessing more dimensions of data quality using the dimensions determined
by Gray Weiskopf[15]. Also, investigating which aspects of the user interfaces in the
different EMR lead to differences in data quality would be a worthwhile exercise. Last
but not least research is needed to determine why GPs seem to miss and over-register
cancer cases, since this can provide additional clues to improve data quality.
53
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
aCknoWLEDGEmEntS
We thank the GPs in the Utrecht area for sharing their anonymized EMR data with us
for this study, Julia Velikopolskaia for her assistance in extracting data from the Julius
General Practitioners’ Network Database and Jackie Senior for editing the manuscript.
54
3
Reu
sab
ility
of c
od
ed d
ata
in th
e pRim
aRy
ca
Re eh
R
REfEREnCES1. Danciu I, Cowan JD, Basford M, et al. Secondary use of clinical data: the Vanderbilt approach.
J Biomed Inform 2014;52:28–35. doi:10.1016/j.jbi.2014.02.0032. Coorevits P, Sundgren M, Klein GO, et al. Electronic health records: new opportunities for
clinical research. J Intern Med 2013;274:547–60. doi:10.1111/joim.121193. Muller S. Electronic medical records: the way forward for primary care research? Fam Pract
2014;31:127–9. doi:10.1093/fampra/cmu0094. Nivel. Feasibility study indicators primary care and Etalage+ data (Haalbaarheidsstudie
indicatoren huisartsenzorg en Etalage+-gegevens). 2011. 5. Dentler K, Numans ME, ten Teije A, et al. Formalization and computation of quality measures
based on electronic medical records. J Am Med Inform Assoc 2014;21:285–91. doi:10.1136/amiajnl-2013-001921 [doi]
6. Blumenthal D. Meaningful use: an assessment. An interview with David Blumenthal, M.D., National Coordinator for Health Information Technology, Office of the National Coordinator. Interview by Mark Hagland. Healthc Inform 2011;28:40,44.
7. Chen EH, Bodenheimer T. Improving Population Health Through Team-Based Panel Management. Arch Intern Med (American Med Assoc 2011;171:(2 pages). doi:10.1001/archinternmed.2011.395
8. Bodenheimer T. The future of primary care: Transforming Practice. N Engl J Med 2008;359(20):2086–9.
9. Neuwirth EE, Schmittdiel JA, Tallman K, et al. Understanding panel management: a comparative study of an emerging approach to population care. Perm J 2007;11:12–20.
10. Loo TS, Davis RB, Lipsitz L a., et al. Electronic Medical Record Reminders and Panel Management to Improve Primary Care of Elderly Patients. Arch Intern Med 2011;171:1552–8. doi:10.1001/archinternmed.2011.394
11. Bleijenberg N, Drubbel I, Ten Dam VH, et al. Proactive and integrated primary care for frail older people: design and methodological challenges of the Utrecht primary care PROactive frailty intervention trial (U-PROFIT). BMC Geriatr 2012;12:16. doi:10.1186/1471-2318-12-16 [doi]
12. Dolor RJ, Fagnan LJ. PBRN conference summary and update. Ann Fam Med 2014;12:176–7. doi:10.1370/afm.1631 [doi]
13. Carey IM, Cook DG, De Wilde S, et al. Implications of the problem orientated medical record (POMR) for research using electronic GP databases: a comparison of the Doctors Independent Network Database (DIN) and the General Practice Research Database (GPRD). BMC Fam Pract 2003;4:14. doi:10.1186/1471-2296-4-14 [doi]
14. Peterson KA, Lipman PD, Lange CJ, et al. Supporting better science in primary care: a description of practice-based research networks (PBRNs) in 2011. J Am Board Fam Med 2012;25:565–71. doi:10.3122/jabfm.2012.05.120100 [doi]
15. Weiskopf NG, Weng C. Methods and dimensions of electronic health record data quality assessment: enabling reuse for clinical research. J Am Med Inform Assoc 2013;20:144–51. doi:10.1136/amiajnl-2011-000681
16. Pascoe SW, Neal RD, Heywood PL, et al. Identifying patients with a cancer diagnosis using general practice medical records and Cancer Registry data. Fam Pract 2008;25:215–20. doi:10.1093/fampra/cmn023 [doi]
17. Terry AL, Chevendra V, Thind A, et al. Using your electronic medical record for research: a primer for avoiding pitfalls. Fam Pract 2010;27:121–6. doi:10.1093/fampra/cmp068 [doi]
18. McGilvray Danette. Executing Data Quality Projects. Elsevier 2008. 19. Netherlands CCCT. The Netherlands Cancer Registry. http://cijfersoverkanker.nl (accessed
20 Mar2012).
20. Statistics Netherlands. Available at: http://statline.cbs.nl (December 2013, date last accessed).
21. Hoogendoorn M, Szolovits P, Moons LMG, et al. Utilizing uncoded consultation notes from electronic medical records for predictive modeling of colorectal cancer. Artif Intell Med 2016;69:53–61. doi:10.1016/j.artmed.2016.03.003
22. Hak E, Rovers MM, Sachs APE, et al. Is asthma in 2-12 year-old children associated with physician-attended recurrent upper respiratory tract infections? Eur J Epidemiol 2003;18:899–902.http://www.ncbi.nlm.nih.gov/pubmed/14561050 (accessed 4 Aug2014).
23. Bentsen BG. International classification of primary care. Scand J Prim Health Care 1986;4: 43–50.
24. Van der Zanden G. Quality Assessment of Medical Health Records using Information Extraction, Master’s thesis. 2010.
25. Dutch College of General Practitoners. Guideline adequate EHR registry. Revised version 2013. Available at: https://www.nhg.org/themas/publicaties/richtlijn-adequate-dossiervorming-met-het-epd.
26. Schouten LJ, Jager JJ, van den Brandt PA. Quality of cancer registry data: a comparison of data provided by clinicians with those of registration personnel. Br J Cancer 1993;68:974–7.http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=1968711&tool=pmcentrez&rendertype=abstract (accessed 20 Aug2015).
27. Van Leersum NJ, Snijders HS, Henneman D, et al. The Dutch surgical colorectal audit. Eur J Surg Oncol 2013;39:1063–70. doi:10.1016/j.ejso.2013.05.008
28. The Dutch College of General Practitoners. No Title. Ref. Model EHR Syst. Available https//www.nhg.org/themas/publicaties/his-referentiemodel.
29. Breslow NE DN. Statistical Methods in Cancer Research: Volume II—The Design and Analysis of Cohort Studies. Lyon: : International Agency for Research on Cancer 1987.
30. Tse J, You W. How accurate is the electronic health record? - a pilot study evaluating information accuracy in a primary care setting. Stud Health Technol Inform 2011;168: 158–64.
31. Akkerman A, Verheij T, Veen R, et al. Interactieve medische informatie van het Huisartsen Netwerk Utrecht en de Almere Zorggroep (‘Interacive medical information from the General Practitioners Network Utrecht and the Almere Group of Care’). Huisarts Wet 2008;51:90–5. doi:10.1007/BF03086658
32. Visscher, S. ten Veen, P. Verheij PR. Kwaliteit van ICPC-Codering (‘Quality of ICPC Coding’. Huisarts en Wet j 2012;10:459–459.
33. Chmiel C, Bhend H, Senn O, et al. The FIRE project: a milestone for research in primary care in Switzerland. Swiss Med Wkly 2011;140:w13142. doi:10.4414/smw.2011.13142 [doi]
34. De Clercq E, Moreels S, Bossuyt N, et al. Routinely-collected general practice data from the electronic patient record and general practitioner active electronic questioning method: a comparative study. Stud Health Technol Inform 2013;192:510–4.
35. Barrett N, Weber-Jahnke JH, Thai V. Engineering natural language processing solutions for structured information from clinical text: extracting sentinel events from palliative care consult letters. Stud Health Technol Inform 2013;192:594–8.http://www.ncbi.nlm.nih.gov/pubmed/23920625 (accessed 4 Aug2014).
CHaPtERDo GPs know their cancer patients?
Assessing the quality of cancer registration in Dutch Primary Care:
a cross sectional validation study
Annet Sollie, General practitioner in training / PhD Fellow, Jessika Roskam, Medical student, Rolf H. Sijmons, Professor of Medical Genetics, Mattijs E Numans, Professor of General Practice, Charles W. Helsper, MD, PhD, Clinical Epidemiologist.
4
Published September 2016 in bmJ open as: Sollie A, Helsper CW, Ader RJ, Ausems MG, van der Woudern JC, Numans ME. Do GPs know their patients with cancer? Assessing the quality of cancer registration in Dutch primary care: a cross-sectional validation study. BMJ Open. 2016 Sep 15;6(9):e012669. doi: 10.1136/bmjopen-2016-012669.
59
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
abStRaCt
ObjectivesTo assess the quality of cancer registry in Primary Care.
Design and SettingA cross-sectional validation study using linked data from primary care Electronic Health
Records (EHRs) and the Netherlands Cancer Registry (NCR).
Population 290,000 patients, registered with 120 General Practitioners, from 50 practice centres in
the Utrecht area, the Netherlands in January 2013.
InterventionLinking the Electronic Health Records (EHRs) of all patients in the Julius General Practitioners’
Network (JGPN) database at an individual patient level to the full Netherlands Cancer
Registry (NCR) (approx. 1.7 mln tumours between 1989 and 2011), to determine the
proportion of matching cancer diagnoses. Full-text EHR extraction and manual analysis
for non-matching diagnoses.
Main Outcome MeasuresProportions of matching and non-matching breast, lung, colorectal and prostate cancer
diagnoses between 2007 and 2011, stratified by age category, cancer type and EHR
system. Differences in year of diagnosis between EHR and NCR. Reasons for non-matching
diagnoses.
ResultsIn the Primary Care EHR, 60.6% of cancer cases were registered and coded in accordance
with the NCR. Of the EHR diagnoses 48.9% was potentially false positive (not registered
in the NCR). Results differed between EHR systems but not between age-categories or
cancer types. The year of diagnosis corresponded in 80.6% of matching coded diagnoses.
Adding full-text EHR analysis improved results substantially. A national disease registry
(the NCR) proved incomplete.
ConclusionsEven though GPs do know their cancer patients, only 60.6% are coded in concordance
with the NCR. Re-users of coded EHR data should be aware that 40% of cases can be
missed and almost half can be false-positive. The type of EHR system influences registration
quality. If full-text manual EHRs analysis is used, only 10% of cases will be missed and
20% of cases found will be wrong. EHR data should only be re-used with care.
60
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
IntRoDuCtIon
Ask General Practitioners (GPs) if they know their cancer patients and they will most
likely answer with an outspoken “yes”! Ask them if these patients are registered with
this cancer diagnosis in their Electronic Health Record (EHR) system and the answer will
be “yes, probably”. Most GPs will acknowledge the importance of adequate disease
registry in EHR records, certainly for a serious disease such as cancer, since these records
are used for information exchange between care providers.
Since re-use of EHR records for other purposes such as chronic disease management,
[1] research [2–4] and quality assessment [5,6] is becoming commonplace, not only in
hospitals [7,8] but also in Primary Care, [2–4] correct and complete registry of diagnoses
using coding systems is pivotal [9–11]. Disease registry using coding systems has been
common practice in Primary Care EHRs for almost two decades in western countries. In
several countries including the Netherlands, guidelines have been developed for correct
registration in the EHR that do address adequate disease coding.[12]
Despite these developments, there are indications that (coded) disease registry in
Primary Care is still suboptimal, [13,14] even for important diagnoses such as cancer.
[15] Literature describing the quality of data in Primary Care however is limited. In order
to assess quality and subsequent (re-) usability of EHR data, it is important to quantify
this quality and to determine which variables influence quality of disease registry. We
will assess various aspects of the quality of disease registry in primary care for re-use
purposes. We focus on cancer since supposedly reliable and elaborate information
concerning cancer diagnoses is available from the Netherlands Cancer Registry (NCR),
thereby providing a potential reference standard required for our study.
We aim to answer the following questions:
1. What are the proportions of matching, missing (potentially false-negative) and wrong
(potentially false-positive) cancer diagnoses in the Primary Care EHR using the NCR
as a reference standard?
2. How accurate is the year of diagnosis registered in the EHR for the matching cancer
cases, when compared to the NCR as a reference standard?
3. Do age of the patient, cancer type and EHR system influence the quality of cancer
diagnosis registry?
4. What are the causes of suboptimal cancer diagnosis registry in the EHR and subsequent
opportunities for improvement?
61
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
mEtHoDS
Design Using a cross-sectional validation study, we assessed the proportion of matching, missing
(potentially false negative) and wrong (potentially false positive) breast, lung, colorectal
and prostate cancer diagnoses in primary care EHR data between 2007 and 2011, using
the NCR as a reference standard. We linked the EHR to the NCR at an individual patient
level using a Trusted Third Party (TTP), to obtain an anonymous dataset containing both
the EHR and the NCR data. We defined coded diagnoses representing the same cancer
type in both databases as matching cases. We defined missing (potentially false-negative)
diagnoses as occurring with one or more of the four cancers under study in the NCR, but
not in the EHR in one of the years 2007 to 2011. We defined wrong (potentially false-
positive) diagnoses as registered with one of the four cancer types under study in the EHR,
but not registered with the same diagnosis in the NCR, in one of the years 2007 to 2011.
DataWe used the routine EHR data extracted from practice centres in the Utrecht area, the
Netherlands, that are a member of the Julius General Practitioners’ Network (JGPN;
120 GPs, 50 practice centres, 290,000 patients). Coded and free-text primary care data
from individual patients enlisted with these centres is periodically extracted to the central
anonymized EHR JGPN database. Data was included if GPs used one of the three most
frequently used EHR systems in the study region; Promedico®, Medicom® and Microhis®.
These systems cover 85% of the population registered with participating GPs. The systems
vary in design and user-interface but are all based on the Reference Model provided by the
Dutch College of General Practitioners. The JGPN population is considered representative
for the Dutch population [16] and its GPs and GP centres represent the average Dutch
GPs and GP centres. GPs were not aware of this study at the time of registry; neither did
they receive specific training on coding. Hence the data in the JGPN can be regarded as
true “routine care data”.
In the Netherlands GP medical encounters are registered according to the “SOAP-
system”.[17] A SOAP-journal consists of four data fields. The first is “Subjective” (S) and
is used to register in plain text what the patient describes, such as complaints and the
reason for the encounter. The second data field is called “Objective” (O) and includes
the GP’s findings and results from clinical examination and measurements, in plain text.
The third field is “Analysis” (A), which is used to register the (working) diagnosis, most
important symptom or hypothesis as plain text and is coded using the International
Classification of Primary Care version 1 (ICPC-1) coding system.[18] The final field is “Plan”
(P), comprising the GPs medication prescriptions, diagnostic tests, referrals to medical
specialists and follow-up appointments as plain text. The list “Episodes”, also coded using
ICPC-1, clusters consultations concerning the same diagnosis for an individual patient
62
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
with corresponding start and end dates. According to the Dutch College of General
Practitioners’ guideline [12] for correct registration, every cancer diagnosis should be
registered as an episode in the EHR and consultations concerning relevant complaints or
treatments should be added to this Episode. The guideline also states that it is mandatory
for GPs to update the EHR Episode with the final diagnosis.
ICPC diagnosis codes are available for the more common types of cancer, including the
cancers under study. There are no separate codes available for the recurrence of cancer, for
suspected cancer or for treatments of cancer. The GP manually enters the ICPC code for a
‘cancer’ diagnosis during consultation or after receiving secondary care correspondence.
A diagnosis code should only be used in the EHR after confirmation of the diagnosis and
not if a diagnosis is suspected. The GP decides if and when a new Episode is created for
the cancer diagnosis and which consultations are added to this Episode.
Reference StandardElaborate information on cancer diagnoses and treatment is available in the NCR.[19]
Specially trained staff members enter relevant data about all Dutch cancer diagnoses
in the NCR database, triggered by hospital pathology reports of newly found cancers.
In addition, cancer diagnoses reported in hospital patient discharge files, for which no
pathological investigation is being performed, are also included in the NCR as clinical
diagnoses for most hospitals. The NCR claims to be almost complete (>95% of all cancers)
for the population of the Netherlands and without false-positive records since 1989.
There is a registration delay reported at the NCR of 3 to 9 months after the pathologist
confirmed the cancer and the delay is claimed to be decreasing. There is some evidence
that the quality of the NCR data is complete and accurate.[20,21] Theoretically, cancer
patients missing in the NCR could include those who are diagnosed with cancer in primary
care based on clinical signs and symptoms, but are unable or refuse to go to a hospital.
Therefore, for these patients, no pathology report or hospital admission is registered,
which would be needed to enter the NCR.
Data collection and analysis
Step 1: Identification of Cancer Cases in Primary Care EHR
We identified all breast, lung, colorectal and prostate cancer cases diagnosed at ages 20
to 90 between 2007 and 2011 in the EHR, using a three-step search strategy. First we
searched for patients with one or more cancer Episodes in the database with ICPC codes
X76 (breast cancer), R84 (lung cancer), D75 (colorectal cancer), and/or Y77 (prostate
cancer). Next we selected all cases without a coded cancer related Episode but with one
or more encounters (home visits, correspondence, consultation) coded with X76, R84, D75
and/or Y77. Finally we selected patients without an Episode or medical encounter coded
for the types of cancer under study, but with whom any prescription for cancer specific
medication was registered during observation time. After identifying these patients, a
63
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
subset of data including all the required information was extracted from the EHR: ICPC
code, year of diagnosis, age at diagnosis, year of birth, sex and type of EHR system used.
Step 2: Linkage of data
Parallel to step 1, the TTP linked the entire JGPN database with the entire NCR database.
The linking was performed after encryption of the data using a mixed algorithm with
deterministic and probabilistic parts based on the following variables: date of birth,
gender, zip code, last name, initials and first name. If date of birth and gender matched
(deterministic part), the probabilistic part of the algorithm started based on the Fellegi-
Sunter [22] model. This means the other variables were compared, yielding scores, which
were totalled and evaluated using weights.
The TTP provided a list with pseudonymised patient numbers of all patients that were
successfully linked and added the NCR data. JGPN datamanagement added EHR data to
every patient number on the list.
Step 3: Matching diagnoses
Dutch GP’s use the ICPC-1 coding system. The NCR uses the International Classification of
Diseases for Oncology (ICD-O). Diagnoses were counted as matching when their ICPC-1
code and ICD-O code represented the same cancer (Table 1).
Diagnoses were stratified based on cancer type, age category (<50, 50 to 75 and
>75) and EHR system. Differences in the year of diagnosis between EHR and NCR were
determined by subtracting the year of diagnosis at the GP from the year of diagnosis as
registered in the NCR.
Step 4: Non-matching (missing and wrong) diagnoses
Non-matching diagnoses were assessed using the NCR as a reference. We determined
the proportion of missing (potentially false-negative) diagnoses in the EHR and stratified
these per cancer type and per age category. We also determined the proportion of
wrong (potentially false-positive) diagnoses stratified by cancer type, age category and
EHR systems used. In case of repeated entries of a new diagnosis for the same person in
the EHR, we counted one as being correct and the other(s) as false-positive.
We extracted and studied the full EHR of a random sample of 120 of the 1,644 (7.3%)
potentially false-positive EHR diagnoses and 120 of the 1,720 (7.0%) potentially false-
negative EHR diagnoses to determine reasons for inaccurate and incomplete GP disease
registry. To ensure representativeness of our sample, we used the following sampling
method: the sample of false positive cases consisted of 4 x 30 cases per cancer type
equally distributed over the years of interest and the EHR system used. The sample of
false-negative case consisted of 5 cases per cancer type per year. The EHR system could
not be taken into account in the selection of false-negative cases since this could only
be determined after extracting the full EHR of the sample.
All data analysis and calculations were carried out using SPSS Statistics version 21.
64
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
Patient InvolvementPatients were not involved in the design, development of outcome measures or conduct
of this study. Since this study uses anonymized EHR data from an existing network
database only, no patients were recruited and thanking patients or disseminating results
directly is not applicable.
RESuLtS
The linkage by the TTP of the full JGPN database to the full NCR database yielded 12,930
JGPN subjects with a registered cancer at the NCR (data from January 2013), of whom 12,526
could be included in our analysis. The remaining 404 (3.1%) records belonged to 202 patients
who were matched twice. These records were considered incorrectly identified. We had to
remove another 14 (0.1%) records for suspected wrong linkage before starting our analyses.
The extraction of breast, lung, colorectal and prostate cancer diagnoses yielded 3,364
cases from the EHR data and 2,839 from the NCR (Table 2).
table 1. Cross-linking of ICPC-1 with ICD-O codes used to define matching cancers in the Electronic Health Records and the Netherlands Cancer Registry
Cancer typeICPC -1 codes used in Electronic
Health RecordICD-9/10- o codes used in national
Cancer Registry
Colorectal cancer D75 incl subtypes: C18 #
D75.01 C19#
D75.02 C20#
D75.03 153*
154 *
Lung cancer R84 C34
162*
163*
Breast cancer X76 incl. subtypes: C50#
X76.1 174*
X76.01
X76.02
X76.03
Prostate cancer Y77 C61#
158*
ICPC-1: International Classification of Primary Care version 1ICD-O: International Classification of Diseases for Oncology (ICD-O)* Codes from ICD-9 (since 1978)# Codes from ICD-10 (since 1990)
65
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
Overall 60.6% (1,720 of 2,839) of cases matched (“sensitivity” of the EHR), which
means that 39.4% (1,644 of 2,839) of cancer cases seem to be missing in the EHR
(potentially false-negative). Furthermore, 1,644 (48.9% of 3,364) of EHR cases were not
found in the NCR, thus should be qualified as potentially false-positive. Consequently,
the “positive predictive value” of a cancer diagnosis in the EHR is 51.1%. The two by
two table illustrates these findings, including a negative predictive value of 99.5% and
a specificity of 99.3 % (Table 3).
We found no substantial differences in proportion of matching, potentially false-
negative and false-positive cases between cancer types and age categories (Table 2).
However, there are differences between EHR systems used; Microhis® has the highest
proportion of matching cases (64.5%, 534 out of 828) and the lowest proportion of
potentially false-positive cases (35.5% 294 of 828). Promedico® has the lowest proportion
of matching cases (44.7%, 925 out of 2,068) and the highest proportion of potentially
false-positive cases (55.3%, 1,143 of 2,068).
The year of diagnosis in the EHR is registered in accordance with the NCR for 80.6%
(1,386 out of 1,720) of cancer cases. For 75.5% (252 of 334) of cases with a differing year
of diagnoses, the deviation from the NCR incidence year is less than 2 years (figure 1).
15
Figure 1: Deviation in year of registered cancer diagnosis in the Electronic Health Record from
reference standard (National Cancer Registry)
Manual analysis of the full EHR text in a random sample of 120 unregistered NCR
cases (potentially false-negative), shows that, even though these cases were not coded with
an Episode or Journal consultation, for 29% (n=35) information about the cancer diagnosis is
present in the EHR plain text, indicating the GPs awareness of the diagnosis. For another 23%
(n=27) the cancer diagnosis is also mentioned, available in plain text, but the coding is based
on cancer related symptoms and not on the final cancer diagnosis (for example breast cancer
coded as “ lump in the breast” X19 and not recoded after confirmation of the diagnosis). In
17% (n=20) of cases the GP registered a coded cancer diagnosis but added the date of registry
figure 1. Deviation in year of registered cancer diagnosis in the Electronic Health Record from reference standard (National Cancer Registry)
66
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
table 2. Results of record linking between the Electronic Health Records and the Netherlands Cancer Registry
total cancers JGPn
total cancers nCR
number matching
Proportion matching
95%CI
number false-Postives
Proportion false-Positives 95%CI
number false-negatives
Proportion false-negatives
95%CI n m n n/m (%) m m/n (%) k k/m (%)
4 types combined 3364 2839 1720 60.6 (58.8; 62.4) 1644 48.9 (47.2; 50.6) 1119 39.4 (37.6; 41.2)
Age < 50 451 412 246 59.7 (55.0; 64.4) 205 45.5 (40.9; 50.1) 166 40.3 (35.5; 45.0)
Age 50-75 2109 1824 1139 62.4 (60.2; 64.7) 970 46.0 (43.9; 48.1) 685 37.6 (35.3; 39.8)
Age > 75 804 603 335 55.6 (51.6; 59.5) 469 58.3 (54.9; 61.7) 268 44.4 (40.5; 48.4)
EHR system
Promedico 2068 x 925 44.7 1143 55.3 (53.1; 57.4) x
Medicom 468 x 261 55.8 207 44.2 (39.7; 48.7) x
MicroHis 828 x 534 64.5 294 35.5 (32.2; 38.8) x
Cancer Type
Breast Cancer 1144 1008 622 61.7 (58.7; 64.7) 522 45.6 (42.7; 48.5) 386 38.3 (35.3; 41.3)
Age < 50 290 267 156 58.4 (52.5; 64.3) 134 46.2 (40.5; 51.9) 111 41.6 (35.7; 47.5)
Age 50-75 681 598 381 63.7 (59.7; 67.6) 300 44.1 (40.3; 47.8) 217 36.3 (32.4; 40.1)
Age > 75 173 143 85 59.4 (51.4; 67.5) 88 50.9 (43.4; 58.3) 58 40.6 (32.5; 48.6)
Prostate Cancer 662 547 336 61.4 (57.3; 65.5) 326 49.2 (45.4; 53.1) 211 38.6 (34.5; 42.7)
Age < 50 6 5 4 80.0 (44.9; 115.1) 2 33.3 (-4.4; 71.1) 1 20.0 (-15.1; 55.1)
Age 50-75 469 425 270 63.5 (59.0; 68.1) 199 42.4 (38.0; 46.9) 155 36.5 (31.9; 41.0)
Age > 75 187 117 62 53.0 (43.9; 62.0) 125 66.8 (60.1; 73.6) 55 47.0 (38.0; 56.1)
Lung Cancer 731 600 331 55.2 (51.2; 59.1) 400 54.7 (51.1; 58.3) 269 44.8 (40.9; 48.8)
Age < 50 46 36 18 50.0 (33.7; 66.3) 28 60.9 (46.8; 75.0) 18 50.0 (33.7; 66.3)
Age 50-75 492 438 248 56.6 (52.0; 61.3) 244 49.6 (45.2; 54.0) 190 43.4 (38.7; 48.0)
Age > 75 193 126 65 51.6 (42.9; 60.3) 128 66.3 (59.7; 73.0) 61 48.4 (39.7; 57.1)
Colon Cancer 827 684 431 63.0 (59.4; 66.6) 396 47.9 (44.5; 51.3) 253 37.0 (33.4; 40.6)
Age < 50 74 58 43 74.1 (62.9; 85.4) 31 41.9 (30.7; 53.1) 15 25.9 (14.6; 37.1)
Age 50-70 502 409 265 64.8 (60.2; 69.4) 237 47.2 (42.8; 51.6) 144 35.2 (30.6; 39.8)
Age > 75 251 217 123 56.7 (50.1; 63.3) 128 51.0 (44.8; 57.2) 94 43.3 (36.7; 49.9)
67
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
table 2. Results of record linking between the Electronic Health Records and the Netherlands Cancer Registry
total cancers JGPn
total cancers nCR
number matching
Proportion matching
95%CI
number false-Postives
Proportion false-Positives 95%CI
number false-negatives
Proportion false-negatives
95%CI n m n n/m (%) m m/n (%) k k/m (%)
4 types combined 3364 2839 1720 60.6 (58.8; 62.4) 1644 48.9 (47.2; 50.6) 1119 39.4 (37.6; 41.2)
Age < 50 451 412 246 59.7 (55.0; 64.4) 205 45.5 (40.9; 50.1) 166 40.3 (35.5; 45.0)
Age 50-75 2109 1824 1139 62.4 (60.2; 64.7) 970 46.0 (43.9; 48.1) 685 37.6 (35.3; 39.8)
Age > 75 804 603 335 55.6 (51.6; 59.5) 469 58.3 (54.9; 61.7) 268 44.4 (40.5; 48.4)
EHR system
Promedico 2068 x 925 44.7 1143 55.3 (53.1; 57.4) x
Medicom 468 x 261 55.8 207 44.2 (39.7; 48.7) x
MicroHis 828 x 534 64.5 294 35.5 (32.2; 38.8) x
Cancer Type
Breast Cancer 1144 1008 622 61.7 (58.7; 64.7) 522 45.6 (42.7; 48.5) 386 38.3 (35.3; 41.3)
Age < 50 290 267 156 58.4 (52.5; 64.3) 134 46.2 (40.5; 51.9) 111 41.6 (35.7; 47.5)
Age 50-75 681 598 381 63.7 (59.7; 67.6) 300 44.1 (40.3; 47.8) 217 36.3 (32.4; 40.1)
Age > 75 173 143 85 59.4 (51.4; 67.5) 88 50.9 (43.4; 58.3) 58 40.6 (32.5; 48.6)
Prostate Cancer 662 547 336 61.4 (57.3; 65.5) 326 49.2 (45.4; 53.1) 211 38.6 (34.5; 42.7)
Age < 50 6 5 4 80.0 (44.9; 115.1) 2 33.3 (-4.4; 71.1) 1 20.0 (-15.1; 55.1)
Age 50-75 469 425 270 63.5 (59.0; 68.1) 199 42.4 (38.0; 46.9) 155 36.5 (31.9; 41.0)
Age > 75 187 117 62 53.0 (43.9; 62.0) 125 66.8 (60.1; 73.6) 55 47.0 (38.0; 56.1)
Lung Cancer 731 600 331 55.2 (51.2; 59.1) 400 54.7 (51.1; 58.3) 269 44.8 (40.9; 48.8)
Age < 50 46 36 18 50.0 (33.7; 66.3) 28 60.9 (46.8; 75.0) 18 50.0 (33.7; 66.3)
Age 50-75 492 438 248 56.6 (52.0; 61.3) 244 49.6 (45.2; 54.0) 190 43.4 (38.7; 48.0)
Age > 75 193 126 65 51.6 (42.9; 60.3) 128 66.3 (59.7; 73.0) 61 48.4 (39.7; 57.1)
Colon Cancer 827 684 431 63.0 (59.4; 66.6) 396 47.9 (44.5; 51.3) 253 37.0 (33.4; 40.6)
Age < 50 74 58 43 74.1 (62.9; 85.4) 31 41.9 (30.7; 53.1) 15 25.9 (14.6; 37.1)
Age 50-70 502 409 265 64.8 (60.2; 69.4) 237 47.2 (42.8; 51.6) 144 35.2 (30.6; 39.8)
Age > 75 251 217 123 56.7 (50.1; 63.3) 128 51.0 (44.8; 57.2) 94 43.3 (36.7; 49.9)
68
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
Manual analysis of the full EHR text in a random sample of 120 unregistered NCR cases
(potentially false-negative), shows that, even though these cases were not coded with an
Episode or Journal consultation, for 29% (n=35) information about the cancer diagnosis is
present in the EHR plain text, indicating the GPs awareness of the diagnosis. For another
23% (n=27) the cancer diagnosis is also mentioned, available in plain text, but the coding
is based on cancer related symptoms and not on the final cancer diagnosis (for example
breast cancer coded as “ lump in the breast” X19 and not recoded after confirmation of
the diagnosis). In 17% (n=20) of cases the GP registered a coded cancer diagnosis but
added the date of registry (later than 2011 so not included in primary analysis) instead
of the date of diagnosis. For 10 cases (8%) the cancer was not found in our initial search
(step 1 methods) but appears to be registered correctly when extracting the full EHR. This
means that in 77% (29 + 23 + 17 + 8) of the cancer diagnoses qualified as potentially
false-negative in the EHR registries, the cancer diagnosis is actually known to the GP and
can be recognized in the EHR with adequate text-finding strategies.
For another 8% (n=10) no information could be traced indicating the presence of
cancer in the full text EHR. For 2% (n=2) another valid explanation for the missing coded
cancer diagnosis is present in the EHR text: one patient moved and unlisted with his GP
and for another patient the diagnosis was made after death and the GP did not add this
diagnosis to the EHR. For 13% (n=16) no written journal EHR data linked to the patient
could be retrieved in the JGPN database.
Analysis of the full EHR text of another 120 randomly selected potentially incorrectly
assigned cancer diagnoses shows that for 18% of these seemingly false-positive
diagnoses, clear and reliable indications of the presence of cancer in the EHR is found,
while no diagnosis is present in the NCR. For 49% (n=59) of false-positive diagnoses, no
reason can be traced in the EHR text (Table 4).
Reviewing these numbers, 90% of NCR confirmed cancer cases can be recognized
and found in primary care EHR systems, counting for two thirds of the EHR coded cancer
cases. An additional 10% of coded cancer cases in primary care EHR systems should be
considered correct, but stays unvalidated since the diagnosis did not reach the National
table 3. Two by two table for matching and non-matching cancer cases registered from 2007 to 2011 in the JGPN and NCR
Cancer status according to nCR
totalCancer present Cancer absent
Cancer status according to JGPN
Cancer present 1,720 1,644 3,364
Cancer absent 1,119 237,906 239,025
total 2,839 239,550 242,389*
* Population of the JGPN registered in the included EHR systems.
69
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
table 4. Causes of False-Negative and False-Positive registration of cancer diagnoses in the Electronic Health Records
false Positive diagnoses
Number % Comments
59 49 no Explanation
59 49 No logical reason can be traced in the full EHR text about the registration of a cancer code with this patient
15 13 Coding Error by GP
15 13 Coding error by GP (f.i. “R84” lung cancer when lung cancer is suspected by the GP)
46 38 Diagnosis correct in EHR
16 13 Year of diagnosis in the EHR is 2011, leaving a small chance that the histological confirmation of the diagnosis (NCR) was performed (and registered) in 2012, while the clinical diagnosis was made and registered in the EHR in 2011.
11 9 Year of diagnosis > 10 years before 2007; cancer not available in the NCR*
8 7 Cancer registered twice at GP
7 6 Patient has moved or diagnosis was made abroad, thus not in the NCR*
4 3 No tissue biopsy was performed in agreement with patient and/or family; thus not in the NCR*
120 100
false negative diagnoses
Number % Comments
92 77 Information about the cancer is available
35 29 Information about the cancer is available in plain text in the EHR but the cancer is not coded
27 23 GP assigned a wrong code / did not update existing code after diagnosis (f.i. “X19” lump in breast” instead of “X76” breast cancer)
20 17 The cancer is coded in the EHR but GP assigned a year of diagnosis > 2011
10 8 Coded cancer found in full EHR but patient was not in initial search due to time lapse between initial search and linkage (> 1 year)
16 13 EHR record cannot be retrieved
16 13 Patient EHR cannot be retrieved (probably linkage error)
10 8 No Explanation
10 8 No EHR text or codes about any cancer
2 2 Various
2 2 Remaining causes: patient has moved, diagnosis after death
120 100
* NCR = National Cancer Registry
70
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
Cancer Registry for various reasons. Approximately 20% of cancer cases found in EHR
systems should be considered as wrongly coded, false positive cases.
DISCuSSIon
Principal FindingsExtracting coded cancer diagnoses from a primary care EHR (JGPN) and linking these
to the Netherlands Cancer Registry demonstrates matching diagnoses in over 60% of
cases. Almost 40% of cancer cases registered in the NCR are missing in the EHR (Table 2).
However, for at least 77% of these false-negative coded diagnoses, un-coded information
indicating the GPs knowledge of the cancer can be found in the EHR (Table 4). Overall,
GPs seem to know the great majority of their cancer patients, since 90% of the NCR
validated cancers are also described in EHR systems.
Almost half of the coded cancer diagnoses in the EHR seem to be false-positive
(Table 2), of which only a minority can be explained by wrongly used diagnostic coding
such as coding symptoms as actual cancer.
There are differences up to 20% in proportions of correct (matching), missing (false-
negative) and wrong (false-positive) cancer diagnoses between EHR systems but not
between age-categories or cancer types. The year of diagnosis is in the EHR is confirmed
in over 80% of matching cases. For incorrectly registered diagnosis-years, 76% deviates
no more than 2 years from the NCR.
Strengths and WeaknessesTo the best of our knowledge, this is the first study to use record linkage to assess quality
of cancer registry in routine primary care data, combined with a search for actual causes
of inadequate registry and resulting opportunities for improvement. The major strengths
of this study are the size of the cohort, the extensive EHR data and the availability of a
reliable reference standard (Netherlands Cancer Registry). Furthermore, the JGPN database
comprises un-manipulated EHR data as available from routine care, hence not improved
or enriched in any way such as in the UK Clinical Practice Research Datalink, formerly
known as General Practice Research Database.[23]
Our study has some restrictions. Since our study was performed in the Netherlands,
the results are indicative for settings similar to ours, which means countries with the GP
in a gatekeeper role, which have adapted to the use of EHRs in primary care resulting
in relatively “mature” EHR systems. We were able to analyse only a sample of false-
positive and false-negative cases. In our study a number of missing (false-negative)
cases (21% (13 + 8) of our sample of 120) could not be traced in the JGPN, and no
explanation regarding wrongfully coded cancer diagnoses could be found in 49% of
false-positive cases.
71
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
Since no unique identifiers could be used for linkage, we used the commonly used
alternative method of probabilistic linkage. Consequently, discrepancies in databases could
in part be a consequence of linkage errors, which could have biased our results in either
direction. The primary problem that may occur, is the rare occurrence of matching two
different patients with identical characteristics. This would result in the false assumption
that a cancer diagnosis registered in the NCR is “missing” in the matched patient in the
EHR. This is expected to occur in < 1% of cases. Another linkage error which may occur
is “no match” for a patient which is registered in both databases, but not by the same
characteristics used for linkage. Since we used: date of birth, gender, zip code, last name,
initials and first name, these characteristics are unlikely to be registered differently in the
registries. Only in case of typing errors, moving out of the zip code area or changing last
names within the time frame between dates on which the data were extracted from the
different databases, or in case of not registering such a previous change in one of these
databases, such linkage error will occur. We estimate the chance of such an occurrence
to be below 1%.
To calculate the concordance in year of diagnosis, we used the calendar year in which
the diagnosis was registered. Consequently, both in the EHR and the NCR, a diagnosis
registered in January of a calendar year, e.g. 2011, and a diagnosis registered in December
of the same year are considered as, for this example, “registered in 2011”. This means
that a difference in registration of “one year”, could be either two days (registered in
JGPN on 31rd of December 2010 and in NCR on first of January 2011) or nearly two
years (registered in JGPN on first of January 2010 and in NCR on 31rd of December
2011). Because we used this rough measure, we only showed the absolute numbers and
refrained from providing statistical testing for concordance.
Comparison with Existing LiteratureTwo Dutch [24,25] [25,24]studies and one Swiss [26] study investigating the coding of
diagnoses in primary care EHRs concluded that the quality of coding in general was
fairly good but varied widely between general practices. None of these studies used
record linkage and the largest study (1.1mln patients [24]) assessed only the presence of
‘meaningful’ ICPC diagnostic codes in the EHR. In another Dutch study [13] the diagnosis
inflammatory arthritis in Dutch EHRs could be validated in 71-78% of 219 patients by
comparison with correspondence from a medical specialist.
In a UK CPRD database study,[14] a high (93%) Positive Predictive Value of Read
Codes for congenital cardiac malformations registered between 1996 – 2010 was found.
However, 31% of cases had a different event date, including 10% that differed no more
than 30 days. These results cannot readily be generalized because practices contributing
data to the CPRD are accepted only when they meet standards of data completeness.
[27] Even though the five studies described so far assessed quality of non-cancer disease
registry, they all present outcomes that are in line with our findings.
72
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
There are several studies that assess the quality of disease registry in primary care
for cancer. Boggon[15] et al reported a concordance level of 83.3% between CRPD
records from a Diabetes cohort and the UK National Cancer Data Repository. This is
higher than the proportion of matching cases (concordance level) at first sight of 60.6%
that we found, but might be comparable to the 80% we found when using additional
search techniques. Also the proportion of false-positives (17%, 967 of 5.797) and false-
negatives (6%, 341 of 5.676) are much lower than the proportions we found, but again,
only high-quality data is imported in the CRPD. Boggon also found relevant differences
between cancer types and age-categories (less concordance with increasing age), which
we did not.
Pascoe et al [10] recruited five GP centres for a retrospective analysis of EHR records
on registration of cancer diagnoses compared to a regional cancer registry in the UK.
One in five (20%) of all primary care cancer patients was not identified when a search for
all patients with cancer was conducted using codes for malignancy. Also 20% of patient
records with a code for malignancy that was confirmed in the cancer registry lacked the
necessary documentation to verify the cancer documented in the EHR. Overall, codes for
cancer in these EHRs had a poor level of completeness (29.4%) and correctness (65.6%),
if compared to the UK Cancer Registry as the reference standard.
The California Kaiser Permanent study [28] aimed to assess variability in date of
prostate cancer diagnosis between 2000 and 2010 by comparing Cancer Registry,
pathology reports and EHR data. Variability in date of diagnosis was found: from 9.6 years
earlier to 10 years later but the vast majority of deviations was small. These results are
comparable with the results in our study, although our deviations ranged from -10 to +
4 years. A recent study by Kearney et al,[29] validating the completeness and accuracy of
the Northern Ireland Cancer Registry (NICR), found a high level of completeness (99.9%)
within the NICR compared to the GP registries. The authors suggest that these excellent
results could be induced by the introduction of the National Health Service (NHS) unique
identifier in 2008, which enables matching and data-enrichment, but also by financially
rewarding GPs that maintain a high quality up-to-date record of patients with chronic
diseases, including cancer.[29,30]
Meaning and Implications for Research and PracticeDo GPs know their cancer patients? Yes, our data show they do know the vast majority.
Does this mean re-users of data can retrieve all these patients using coded diagnoses?
No, because the proportion of wrong and missing diagnoses is too high. If re-users
have access to full-text, they would be able to identify op to 90% of cases reliably using
labor-intensive manual exploration or text-mining techniques of EHRs. Also, 20% of cases
identified will have to be excluded after re-assessment. For some purposes this will be
acceptable, for other purposes it will not be.
73
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
Although we have shown that GPs seem to know their own cancer patients, locums
and doctors working at out-of-hours clinics do rely heavily on EHR data, including
coded diagnoses. Missing (false-negative) and wrong (false-positive) cancer diagnoses
on this list could have adverse effects on clinical practice, including medical decisions
made elsewhere. Also, patients could perceive errors in diagnosis lists as unprofessional
and unreliable. From a research perspective, erroneously including non-cancer patients
(false-positives) and missing real cancer cases (false-negatives), may introduce bias. If
textmining techniques are used, these results improve substantially, as was shown in this
study as well as in others.[31] However, the possibility of residual confounding cannot
be completely excluded.
A number of causes for suboptimal registry have been demonstrated in our study
that might be used as a starting point for improving data quality at the source, hence at
data entry. Improvements could be made (1) through education for practicing and future
GPs, by improving usability of (2) EHR systems and (3) coding systems.
First of all, GPs awareness and coding skills could possibly be improved through
education in order to decrease coding errors and errors in registered year of diagnosis.
Although we have not found any studies proving education actually improves data quality,
we do know that financial incentives as well as feedback using data quality reports does
improve data quality.[30,32,33] This shows that improving registration quality is feasible
and can be learned. Furthermore, GPs could evaluate and update working processes at
the GP practice to integrate diagnosis registry after a letter from a hospital or diagnostic
laboratory is received.
Second, EHR systems could be improved by facilitating user-friendly and accurately
coded diagnosis registry. Some systems are subject to less false-positive diagnoses and
a higher number of accurate cancer diagnosis. Since we do not expect these differences
to be caused by confounding resulting from different types of GPs, choosing certain
types of EHRs, this implies that differences in system design lead to varying data quality.
Adding options which are now missing, e.g. to directly register date of diagnosis,
suspected, recurring and metastasizing disease, treatment, increased markers (f.i. for
Prostate Specific Antigen or PSA) and a positive family history could also improve quality.
Thirdly, improvements could also be realized by adding codes to the coding system
(f.i. for recurring and metastasizing disease), by providing adequate synonyms to improve
findability of codes and by adding relevant crosslinks to facilitate data-sharing between
sources.
Another strategy, which might improve clinical practice, was used in our study.
Linkage has its benefits, since [23] multiple data sources are often complementary and
taken together have added value, as was also demonstrated in our study. For current
GP practice, the effort that has to be put into requesting and performing the actual
linkage process (including patient informed consent) is not worth the benefit. However,
74
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
if structural linkage of EHRs to accurate medical data sources (such as the NCR) should
become available, this information could be used to pro-actively alert and inform the GP
to check and adjust recordings in routine care data.
For research purposes and quality assessment we feel that validation of disease cases
and/or improvement of EHR data is necessary before working with the data, particularly
if only coded data are used. Research using EHR data provides access to a very rich data
source, but its interpretation and use should only be performed in cooperation with
experienced clinicians who can judge the meaning of the information in its context.
Linkage could be one of the tools to decrease the number of false-negative coded cancer
diagnoses. For a lot of diseases however no reference standard is available. Linkage to
hospital records could improve data also by decreasing false-negative cases, but the
quality of these data is also likely to be suboptimal. Studying the full EHR, which is time-
consuming, might also be supported by automatized textmining techniques which will
help identifying false-positive records. If in the future these techniques can be made
more advanced and more reliable, these might replace the need for manual searching.
Unanswered Questions and Future researchFuture research should, beside correctness and completeness, evaluate other dimensions
of data quality; concordance, plausibility and currency.[9] Also, quality of other key data
items besides the diagnosis should be studied, for instance risk factors, treatment and
allergies. In cooperation with the NCR, further exploration of the cases where the EHR
seems to provide reliable indications of the presence of cancer, whereas there is no record
in the NCR, is needed in the near future.
Evaluating the user interface of the various EHR systems and determining how these
explain the differences in data quality in this study, would be a worthwhile exercise.
Furthermore in this study we investigated a serious and relatively common disease;
investigating the quality of registry for more rare or less serious diseases may provide
different results, which would be of additional value.
Last but not least, the design, implementation and evaluation of actual interventions
in the GP practice to improve disease registry would provide a necessary next step in
improving EHR data quality. Improving text-mining software and strategies could be
part of this.
Concluding remarksYes, GPs do know the vast majority of their cancer patients. However, re-users of coded
Electronic Health Care data should be aware that they are at risk of missing 40% of cancer
cases and that almost half of the cancer cases found could be wrongfully registered.
Analysing the full-text EHRs improves these numbers: only 10% of cases will be missed and
20% of cases found will be wrong. Particularly in non-clinical circumstances like research,
when high accuracy is needed, Primary Care EHR data should only be re-used with care.
75
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
aCknoWLEDGEmEntS
The authors thank the registration teams of the Comprehensive Cancer Centers for the
collection of data for the Netherlands Cancer Registry and the scientific staff of the
Netherlands Cancer Registry. We thank the GPs in the Utrecht area participating in the
Julius General Practitioners’ Network for sharing their anonymized EHR data with us for
this study and Julia Velikopolskaia for her assistance in extracting data.
76
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
REfEREnCES1. Bleijenberg N, Drubbel I, Ten Dam VH, et al. Proactive and integrated primary care for frail older
people: design and methodological challenges of the Utrecht primary care PROactive frailty intervention trial (U-PROFIT). BMC Geriatr 2012;12:16. doi:10.1186/1471-2318-12-16 [doi]
2. Dolor RJ, Fagnan LJ. PBRN conference summary and update. Ann Fam Med 2014;12:176–7. doi:10.1370/afm.1631 [doi]
3. Peterson KA, Lipman PD, Lange CJ, et al. Supporting better science in primary care: a description of practice-based research networks (PBRNs) in 2011. J Am Board Fam Med 2012;25:565–71. doi:10.3122/jabfm.2012.05.120100 [doi]
4. Terry AL, Chevendra V, Thind A, et al. Using your electronic medical record for research: a primer for avoiding pitfalls. Fam Pract 2010;27:121–6. doi:10.1093/fampra/cmp068
5. Dentler K, Numans ME, ten Teije A, et al. Formalization and computation of quality measures based on electronic medical records. J Am Med Inform Assoc 2014;21:285–91. doi:10.1136/amiajnl-2013-001921 [doi]
6. Blumenthal D. Meaningful use: an assessment. An interview with David Blumenthal, M.D., National Coordinator for Health Information Technology, Office of the National Coordinator. Interview by Mark Hagland. Healthc Inform 2011;28:40,44.
7. Coorevits P, Sundgren M, Klein GO, et al. Electronic health records: new opportunities for clinical research. J Intern Med 2013;274:547–60. doi:10.1111/joim.12119
8. Danciu I, Cowan JD, Basford M, et al. Secondary use of clinical data: the Vanderbilt approach. J Biomed Inform 2014;52:28–35. doi:10.1016/j.jbi.2014.02.003
9. Weiskopf NG, Weng C. Methods and dimensions of electronic health record data quality assessment: enabling reuse for clinical research. J Am Med Inform Assoc 2013;20:144–51. doi:10.1136/amiajnl-2011-000681
10. Pascoe SW, Neal RD, Heywood PL, et al. Identifying patients with a cancer diagnosis using general practice medical records and Cancer Registry data. Fam Pract 2008;25:215–20. doi:10.1093/fampra/cmn023 [doi]
11. Muller S. Electronic medical records: the way forward for primary care research? Fam Pract 2014;31:127–9. doi:10.1093/fampra/cmu009
12. The Dutch College of General Practitoners. Guideline adequate EHR registry. Revised version 2013. Available at: https://www.nhg.org/themas/publicaties/richtlijn-adequate-dossiervorming-met-het-epd.
13. Nielen MMJ, Ursum J, Schellevis FG, et al. The validity of the diagnosis of inflammatory arthritis in a large population-based primary care database. BMC Fam Pract 2013;14:79. doi:10.1186/1471-2296-14-79
14. Hammad TA, Margulis A V, Ding Y, et al. Determining the predictive value of Read codes to identify congenital cardiac malformations in the UK Clinical Practice Research Datalink. Pharmacoepidemiol Drug Saf 2013;22:1233–8. doi:10.1002/pds.3511
15. Boggon R, van Staa TP, Chapman M, et al. Cancer recording and mortality in the General Practice Research Database and linked cancer registries. Pharmacoepidemiol Drug Saf 2013;22:168–75. doi:10.1002/pds.3374
16. Hak E, Rovers MM, Sachs APE, et al. Is asthma in 2-12 year-old children associated with physician-attended recurrent upper respiratory tract infections? Eur J Epidemiol 2003;18:899–902.http://www.ncbi.nlm.nih.gov/pubmed/14561050 (accessed 4 Aug2014).
17. Van der Zanden G. Quality Assessment of Medical Health Records using Information Extraction, Master’s thesis. 2010.
18. Bentsen BG. International classification of primary care. Scand J Prim Health Care 1986;4:43–50.
19. Netherlands Cancer Registry, 2015 www. cijfersoverkanker. nl/?languag. = en. Netherlands Comprehesive Cancer Registration.
77
4
Assessin
g th
e qu
Alit
y o
f cA
nc
er r
egistr
Atio
n in
Du
tch
Prim
Ary
cA
re
20. Van Leersum NJ, Snijders HS, Henneman D, et al. The Dutch surgical colorectal audit. Eur J Surg Oncol 2013;39:1063–70. doi:10.1016/j.ejso.2013.05.008
21. Schouten LJ, Jager JJ, van den Brandt PA. Quality of cancer registry data: a comparison of data provided by clinicians with those of registration personnel. Br J Cancer 1993;68:974–7.http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=1968711&tool=pmcentrez&rendertype=abstract (accessed 20 Aug2015).
22. Xinran Li, Guttmann A, Demongeot J, et al. An empiric weight computation for record linkage using linearly combined fields’ similarity scores. Conf Proc. Annu Int Conf IEEE Eng Med Biol Soc IEEE Eng Med Biol Soc Annu Conf 2014;2014:1346–9. doi:10.1109/EMBC.2014.6943848
23. Herrett E, Shah AD, Boggon R, et al. Completeness and diagnostic validity of recording acute myocardial infarction events in primary care, hospital care, disease registry, and national mortality records: cohort study. BMJ 2013;346:f2350.http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=3898411&tool=pmcentrez&rendertype=abstract (accessed 18 Feb2015).
24. Visscher, S. ten Veen, P. Verheij PR. Kwaliteit van ICPC-Codering (‘Quality of ICPC Coding’. Huisarts en Wet j;10:459–459.
25. Akkerman A, Verheij T, Veen R, et al. Interactieve medische informatie van het Huisartsen Netwerk Utrecht en de Almere Zorggroep (‘Interacive medical information from the General Practitioners Network Utrecht and the Almere Group of Care’). Huisarts Wet 2008;51:90–5. doi:10.1007/BF03086658
26. Chmiel C, Bhend H, Senn O, et al. The FIRE project: a milestone for research in primary care in Switzerland. Swiss Med Wkly 2011;140:w13142. doi:10.4414/smw.2011.13142 [doi]
27. Clinical Practice Research Datalink, 2015, www.cprd.com. 28. Porter KR, Chao C, Quinn VP, et al. Variability in date of prostate cancer diagnosis: a
comparison of cancer registry, pathology report, and electronic health data sources. Ann Epidemiol 2014;24:855–60. doi:10.1016/j.annepidem.2014.09.004
29. Kearney TM, Donnelly C, Kelly JM, et al. Validation of the completeness and accuracy of the Northern Ireland Cancer Registry. Cancer Epidemiol 2015;39:401–4. doi:10.1016/j.canep.2015.02.005
30. Investing in General Practice; The new general medical services contract. 2004.www.legislation.gov.uk/uksi/2004/291/pdfs/uksi_20040291_en.pdf
31. Hoogendoorn M, Szolovits P, Moons LMG, et al. Utilizing uncoded consultation notes from electronic medical records for predictive modeling of colorectal cancer. Artif Intell Med 2016;69:53–61. doi:10.1016/j.artmed.2016.03.003
32. van der Bij S, Khan N, Ten Veen P, et al. Improving the quality of EHR recording in primary care: a data quality feedback tool. J Am Med Inform Assoc 2016;356:2527–34. doi:10.1093/jamia/ocw054
33. Taggart J, Liaw S-T, Yu H. Structured data quality reports to improve EHR data quality. Int J Med Inform 2015;84:1094–8. doi:10.1016/j.ijmedinf.2015.09.008
CHaPtERPrimary Care management of
women with breast cancer related concerns - a dynamic cohort study
using a Network Database
Annet Sollie MD, MSc, Charles W Helsper MD PhD, Rosanne J.M. Ader MD, Margreet G.E.M. Ausems MD, PhD, Johannes C van der Wouden, PhD,
Mattijs E Numans MD, PhD
5
Published in the European Journal of Cancer Care as, June 2016 as: Sollie A, Helsper CW, Ader RJ, Ausems MG, van der Wouden JC, Numans ME. Primary care management of women with breast cancer-related concerns-a dynamic cohort study using a network database. Eur J Cancer Care (Engl). 2016 Jun 15. doi: 10.1111/ecc.12526.
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
81
abStRaCt
The aim of this study was to determine the incidence, management and diagnostic
outcomes of breast cancer related concerns presented in primary care.
A dynamic cohort study was performed in the anonymised routine electronic medical
records (EMR’s) extracted from 49 General Practices in The Netherlands (163,471 person-
years, women aged 18 - 75). Main Outcome Measures were: 1) incidence rates for breast
cancer related concerns in Primary Care, 2) proportions of these women with and without
symptoms of the breast referred for further investigation, 3) proportions of referrals
(not) according to the guideline and 4) proportions of women with breast cancer related
concerns diagnosed with breast cancer during follow-up.
Breast cancer related concerns are presented frequently in Primary Care (incidence
rate 25.9 per 1,000 women annually). About half these women are referred for further
investigation. There is room to improve GP management, mainly for women with an
increased lifetime risk of developing breast cancer. Information concerning family history
of cancer is often missing in the EMR. Since cancer is rarely diagnosed during follow-up,
particularly when symptoms are absent, reduction of unnecessary concerns is plausible
if identification of those without an increased risk is improved.
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
82
baCkGRounD
Breast cancer is by far the most frequently diagnosed cancer among women worldwide
and is still the second cause of cancer death among women in more developed regions[1].
This is the most plausible explanation why many women have concerns about breast
cancer [2]. Breast cancer related concerns can be caused by physical complaints of the
breast, by fear of breast cancer in general or by experiences with breast cancer, e.g. in
the family[3,4]. In the Netherlands, as in other countries with a gatekeeper health care
system such as the UK, women with potential cancer concerns present first to their
General Practitioner (GP) if they decide to consult a health care worker. GP management
should be focussed on identifying women with an increased risk of having or developing
breast cancer, with as little delay as possible. Women without an increased risk of breast
cancer should be reassured to prevent further anxiety and fear.
Research has shown that delays in cancer diagnosis can be caused in primary
care (first presentation to referral) [5,6]. Furthermore, studies show that genetic risk
assessment with respect to cancer in primary care is still inadequate [7–11]. This means
there are indications that GP management for a number of women presenting with
breast cancer related concerns, is not optimal [12–15]. However, this problem has not
recently been quantified. Also, no recent data is available on the frequency with which
GPs are consulted by women with breast cancer related concerns and neither have
breast cancer outcomes among these women been related to the initial reason for
encounter at the GP recently.
Therefore we aim to determine the incidence, management, and diagnostic outcomes
of breast cancer related concerns presented by women in primary care, by answering
the following research questions:
1. What is the incidence of GP consultations primarily focussed on breast cancer related
concerns, for women with and without symptoms of the breast?
2. What proportion of these symptomatic and asymptomatic women are referred by
their GP for further investigation (breast clinic or radiology department) and/or for
cancer genetic counseling?
3. What proportion of women, from subgroups that present with fear of breast cancer
or a positive family history of breast cancer, are identified by their GP as having an
increased lifetime breast cancer risk?
4. What proportion of women identified by their GPs as having an increased lifetime
risk of developing breast cancer (from question 3) are referred for annual screening
or cancer genetic counseling in accordance to the guideline?
5. What proportion of women presenting in primary care with breast cancer related
concerns are diagnosed with breast cancer during follow-up in relation to the initial
reason for the encounter?
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
83
mEtHoDS
A flowchart of the methods used in this study is depicted in Figure 1.
35
Methods A flowchart of the methods used in this study is depicted in Figure 1.
Study Population Women 18-75y, registered in 2008, 2009 and/or 2010 in Julius Network Database (JGPN)
Step 1 - Selection Selection of all women with one or more coded consultations concerning the mamma/breast,
grouping according to ICPC code assigned at first consultation
Physical complaints X18, X19, X20, X20.1,
X79 X88
(Known) Breast cancerX76, X76.1
Fear of Breast cancerX26
Positive Family History A29.2
Step 4 - Determine Breast cancer patients Number of women with coded diagnosis of breast cancer at final consultation
Step 3 - Determine GP management 3.1 - Referral rates to breast clinic, mammogram/ultrasound and genetic counseling
3.3 – Calculate number of adequate, inadequate and missed referrals for annual screening and
genetic counseling
Step 2 – Symptoms or not Determine if selected women (X26 and A29.2) are
symptomatic or asymptomatic
3.2 - Determine individual life time risk of breast cancer based on documented family history in
EMR text
Data were obtained from the Julius General Practitioners’ Network (JGPN) Database[16–
19]. This database comprises anonymized, coded data that are periodically extracted
from the GPs routine Electronic Medical Records (EMR). The data concern approximately
250,000 patients enlisted with 49 sentinel General Practice centres (120 GPs) in the region
of Utrecht in 2010. Since every Dutch citizen is enlisted with a GP and the practices sharing
their data are randomly spread around the city of Utrecht and surrounding villages, the
involved population is considered representative of the Dutch primary care population
[20]. The available data include ICPC-coded consultations and episodes with diagnoses,
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
84
ATC-coded prescribed medication, laboratory test results, and (for a large proportion of
the patients) also coded referrals and coded letters from medical specialists.
GP consultations in The Netherlands are registered according to the “SoaP-system”
[21]. A SOAP journal consists of four data fields: Subjective: patient complaint, reason
for consultation; objective: clinical examination; analysis (possible) diagnosis; and Plan.
The A-line of a SOAP journal is coded using the International Classification of Primary
Care version 1 (ICPC-1) of the 2009 coding system, published by the WHO [22]. The S, O,
and P lines are usually registered using free text only. The Dutch College of General
Practitioners published a Guideline for Adequate Registry [23]. According to this guideline,
the General Practitioner decides how and when to code a diagnosis or symptom in the
A line. For breast cancer this will usually be done after a letter has been received from a
hospital where the cancer has been diagnosed.
GP’s were not aware of this study at the time of registering their consultations neither
did they receive specific training on coding.
Study population: selection of patients and extraction of data from DatabaseIncluded were all female patients enlisted in participating JGPN practice centres that use
the Dutch EMR system Promedico-ASP®. Relevant parts of the EMR records of all female
patients between ages 18 and 75 years with one or more ICPC consultation codes that
indicate breast cancer related concerns in 2008, 2009, and/or 2010 were extracted from
the database (Table 1).
table 1. Relevant ICPC-Codes used for identification of patients from database
Physical Complaints of the breast
X18 Breast pain female
X19 Breast lump / mass female
X20 Nipple symptom / complaint female
X20.1 Nipple discharge
X21 Breast symptom / complaint female other
X79 Benign neoplasm breast female
X88 Fibrocystic disease breast
breast cancer
X76 Malignant neoplasm breast female
X76.01 Adenocarcinoma breast female
fear of breast cancer
X26 Fear of breast cancer female
Positive family history of breast cancer
A29.2 Breast cancer in family history
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
85
Age at the time of the first visit to the GP that was registered with the selected ICPC
code was determined. Data that were extracted include year of birth; all dates and EMR
records assigned to the selected ICPC-codes; ATC prescribed medication; and referrals
to departments of radiology, surgery, or clinical genetics.
The follow-up time for each patient was defined as the time period between the
date of the first registered consultation with one of the selected ICPC codes and that of
closure of the database for this study (31 December, 2010).
For two subgroups of consultations registered with having fear of breast cancer (X26)
and having a positive family history of breast cancer (A29.2), the complete, anonymized
Electronic Medical Record data were extracted.
Data analysisThe data were analysed in four consecutive steps (Figure 1):
Step 1 (research question 1)
All women with one or more registered ICPC coded consultations that indicate concerns
related to breast cancer in one of the study years were identified in the JGPN database
using the 11 ICPC codes for either complaints of the breast, breast cancer, fear of
breast cancer or a positive family history for breast cancer (table 1). The results were
grouped according to ICPC-code at first consultation and to age category (18–50 years
or 50–75 years). Women between 50–75 years of age are invited to participate in the
Dutch national breast cancer screening program[24]. These women are advised to consult
their GP in case of an abnormal mammogram; also the GP receives a notification. It was
not possible to extract these notifications from the EMR.
Multiple consultations with the same code were counted as one; for women who
had multiple consultations with different codes, the code at the first consultation was
chosen as the reason for encounter. Corresponding incidence rates were calculated per
1000 person years.
Step 2 (research question 1)
For women that were registered with an ICPC code X26 (fear of breast cancer) or A29.2
(family history of breast cancer), the EMR texts in 2008, 2009, and 2010 were studied
to determine whether or not they also reported symptoms of the breast.
Step 3 (research questions 2, 3 and 4)
GP management of these women was determined in three sub-steps:
Step 3.1. First, the number of women that were referred to a breast clinic, for a mammogram
or ultrasound, and to a genetics department were counted by checking coded referrals in
the EMR but also by manually checking all A and P lines in the SOAP journal of the selected
women. Multidisciplinary breast clinics and radiology departments use referral forms, but
the actual referral is written in free text. The text registered by the GP in the EMR or a
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
86
referral letter was used in the analysis. Referral rates were calculated as percentages of the
total population of women that consulted their GPs with concerns about breast cancer.
Step 3.2. For the subgroups of women with codes X26 (fear of breast cancer) and A29.2
(positive family history), the presence of an increased risk of developing breast cancer
was determined, using the complete EMR, including free text. Increased risk was defined
according to the primary care guideline “Diagnosing Breast Cancer” from the Dutch
college of general practitioners, summarized in table 2 ([25].
table 2. Referral policies for screening and genetic counseling, according to Dutch guideline “Diagnosing breast cancer”
SCREEnInG - moderately increased lifetime risk (20 – 30%): mammography requested by GP from age 40 to 50 possibly supplemented by clinical breast examination
- 1 first and 1 second degree relative diagnosed with breast cancer < 50 years old.
- 2 first degree relatives with breast cancer regardless of age
- ≥ 3 first or second degree relatives with breast cancer, regardless of age
- 1 first degree relative with bilateral or multifocal breast cancer, with first tumor diagnosed at age < 50.
- 1 first or second degree relative with ovarian cancer regardless of age and 1 first or second degree relative with breast cancer.
GEnEtIC CounSELInG - Strongly increased lifetime risk of developing breast cancer (>30%), referral to a genetics department
- 1 first degree relative diagnosed with breast cancer < 35 years old.
- ≥ 2 first degree relatives diagnosed with breast cancer regardless of age
- ≥ 3 first or second degree relatives with breast cancer of which at least one tumor diagnosed before age of 50.
Step 3.3. Based on the same guideline it was determined whether women at risk should
be referred for either annual screening, or genetic counseling (table 2). These results
were compared with the registered referral practice of the GP based on the EMR text
and it was determined whether or not referral was done according to the guideline. If
the available information in the EMR was insufficient to determine whether or not referral
was indicated, this was also marked. Subsequently these numbers were converted to
proportions of referrals for annual screening or genetic counseling (not) according to
the guideline.
Step 4 (research question 5)
The proportions of women with a coded diagnosis of breast cancer (X76 or X76.1) during
follow-up were determined for each subgroup of initially coded consultations.
All statistical analyses were performed using SPSS (version 20); EMR texts were
extracted as .csv files and analysed using Microsoft Excel 2010.
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
87
RESuLtS
Incidence rates The total number of women between the age of 18–75 registered in the database was
54,947 in 2008, 52,885 in 2009, and 55,639 in 2010, giving a total of 163,471 person
years. As demonstrated in Figure 2 and Table 3, we found 4,240 women with one or
more contacts with breast cancer related ICPC codes within these years.
The mean age of the women at first consultation was 42.1 ± 13.7 years, and 72.4%
were aged <50. Mean duration of follow-up in the dataset was 1.6 ± 0.9 years. The
overall incidence rate for women consulting their GP with breast cancer related concerns
was 25.9 per 1,000 women per year (4,240/163,471). This means that a Dutch GP with
an average list size of 2,350 patients[26] and an age distribution comparable to the
whole country, will be consulted by women concerned about breast cancer or having
complaints concerning the breast about 22 times a year.
Of the 4,240 unique women with a breast cancer related ICPC found, 3,619 (85.3%
or 22.1 per 1,000 per year) were coded as presenting with physical symptoms and
signs of the breast(s). Most of them reported pain (983 = 23.2% or 6.0 per 1,000 per
year) or a lump (1,080 = 25.5% or 6.6 per 1,000 per year). Fear of breast cancer was
registered with 340 women (8.0% or 2.0 per 1,000 per year), and breast cancer in the
family history was registered as the first reason for encounter with 281 women (6.6%
or 1.7 per 1,000 per year). Among the women registered with fear of breast cancer at
first consultation, 138 of the 340 (41%) also reported having one or more symptoms.
Among the 281 women initially coded with having a family history of breast cancer, 37
(13%) were symptomatic at the time of coding (Table 4).
In summary, the incidence rates are 22.1 per 1,000 per year for women presenting
with physical signs and symptoms of the breast at the first consultation, and 3.8 women
per 1,000 per year for women presenting with fear of breast cancer or a positive family
history. The overall incidence rate for women consulting their GP with breast cancer
related concerns is 25.9 per 1,000 per year.
ManagementAs indicated in Tables 4 and 5, the overall referral rate for further investigation (breast
clinic or radiology department) was 53.2% (N=2,257 out of 4,240 consulting their
GP). A quarter of these women were referred to a breast clinic, which is 11.7% of
total referrals (N = 495, 19.7% aged 50–75 years, 8.6% aged 18–50 years); 1,762
women (41.6%) were referred for mammography and/or ultrasound without further
diagnostic facilities, equally spread over the age categories. Only 47 women (1.1 %)
were referred to a genetics department, of whom the majority were in the age group
18–50 (Table 5).
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
88 42
Figu
re 2
: Num
ber o
f wom
en w
ith re
leva
nt IC
PC c
ode
at fi
rst G
P co
nsul
tatio
n; n
umbe
r dev
elop
ing
brea
st c
ance
r
fig
ure
2. N
umb
er o
f w
omen
wit
h re
leva
nt IC
PC c
od
e at
firs
t G
P co
nsul
tati
on;
num
ber
dev
elo
ping
bre
ast
canc
er
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
89
table 4. Presenting number and proportion of (a)symptomatic women referred for further investigation with A29 or X26 codes at first consultation
a29.2 at first consultationfamily history of bC
x26 at first consultationfear of bC
Consulting GP n (% of total)
Referred m (% of
subgroup)Consulting GP n (% of total
Referred m (% of
subgroup)
Symptomatic 50-75 yr 8 28
18-50 yr 29 110
Total 37 (13%) 27 (73%) 138 (41%) 72 (52%)
Asymptomatic 50-75 yr 40 19
18-50 yr 163 88
Total 203 (72%) 109 (54%) 107 (31%) 56 (52%)
EMR info insufficient 50-75 yr 8 28
18-50 yr 33 67
Total 41 (15%) 32 (78%) 95 (28%) 57 (60%)
Total 50-75 yr 56 75
18-50 yr 225 265
total 281 168 (60%) 340 185 (54%)
table 3. Subgroups of women with ICPC codes at first GP consultation per 1,00 women per year (163,471 person years)
ICPC-Code at first consultation
18-50 yr n (inc per 1,000
women per yr)
50-75yr n (inc per 1,000
women per yr)
18–75yr n (inc per 1,000
women per yr)
A29.2 Breast cancer in family history 225 (1.4) 56 (0.3) 281 (1.7)
X18 Breast pain female 777 (4.8) 206 (1.3) 983 (6.0)
X19 Breast lump /mass female 812 (5.0) 268 (1.6) 1,080 (6.6)
X20 Nipple symptom / complaint female 201 (1.2) 32 (0.2) 233 (1.4)
X20.1 Nipple discharge 67 (0.4) 11 (0.1) 78 (0.5)
X21 Breast symptom / complaint female other
330 (2.0) 169 (1.0) 499 (3.1)
X26 Fear of breast cancer female 264 (1.6) 76 (0.5) 340 (2.1)
X76 Malignant neoplasm breast female 56 (0.3) 243 (1.5) 299 (1.8)
X76.1 Adenocarcinoma breast female 7 (0.0) 35 (0.2) 42 (0.3)
X79 Benign neoplasm breast female 34 (0.2) 27 (0.2) 61(0.4)
X88 Fibrocystic disease breast 299 (1.8) 45 (0.3) 344 (2.1)
total 3,070 1,170 4,240
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
90
table 5. Proportion of women referred to a multidisciplinary breast clinic, radiology department and genetics department
breast Clinicn (% of 4240
consulting GP)
Radiology (mammogram or ultrasound)n (% of 4240
consulting GP)
Genetics Departmentn (% of 4240
consulting GP)
total number of women per age group
Referrals 50-75 yr 231 (19.7%) 430 (36.8%) 6 (0.5%) 1,170
18-50 yr 264 (8.6%) 1,332 (43.4%) 41 (1.3%) 3,070
total 495 (11.7%) 1,762 (41.6%) 47 (1.1%) 4,240
Of the women registered with fear of breast cancer at first consultation but without
physical symptoms according to the EMR text, more than half (52%) were referred
for further investigation. For the group of women initially coded as having a family
history of breast cancer, 109 (54%) were referred for further investigation, despite being
asymptomatic (Table 4).
Identification and management of women with increased lifetime risk Increased
lifetime risk (subgroups X26 and A29.2 at first consultation) was established and
registered by GP’s in 46 and 128 women, in subgroups X26 and A29.2, respectively. For
324 of the 621 EMR files studied (52%), the information regarding family history was
insufficient to determine whether women have an increased lifetime risk and should be
referred for annual screening or genetic counseling (table 6). This includes 89 (32 + 57)
EMR files from patients that were actually referred.
Moderately increased lifetime risk (20–30%). A total of 34 (10%) women out of 340
with an ICPC code X26 (fear of breast cancer) at first consultation and a total of 105
(37%) women out of 281 with an ICPC A29.2 (positive family history of breast cancer)
at first consultation were referred for annual screening and, thus, had been identified by
the GP as having a moderately increased lifetime risk (table 6). Of these referrals, 18–25%
were done according to the guideline, 13–15% not according to the guideline, and for
62– 68% of referrals, the information in the EMR was insufficient to determine whether
or not the referrals were done according to the guideline (table 6). Three women who
should have been referred for annual screening were missed.
Strongly increased lifetime risk (>30%). For genetic counseling, 12 (4%) women who
were coded as presenting with fear of breast cancer and 23 (8%) coded with a family
history of breast cancer were referred and, thus, had been identified by the GP as having
a strongly increased lifetime risk. Referral was according to guideline for 13%–33%,
and not according to the guideline for 26%–42%. For 25%– 61% of referrals it was
impossible to determine whether or not they were done according to the guideline due to
insufficient data in the EMR text (table 6). Seven women who should have been referred
for genetic counseling were missed.
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
91
Outcomes A total of 450 women (10.6% of N = 4,240 consulting their GP with breast cancer related
concerns) were registered as having breast cancer (ICPC code X76 and/or X76.1 at final
consultation in our population during follow-up time). Of these women, 109 (2.6% of
N = 4,240) developed breast cancer after first consulting their GP with complaints of the
breast or concerns about developing breast cancer (table 3, figure 2). Of the remaining
341 women (8.0% of N = 4,240) that the GPs registered breast cancer as the reason for
encounter, 278 (6.5 % of N = 4,240) of them were within the age category 50–75 years
and, thus, had been invited to participate in the national breast cancer screening program.
Leaving out the patients in this age category gives a percentage of 4% ((450 - 341
+ 278)/4,240) developing breast cancer during follow up time (1.6 ± 0.9 years).
Women consulting their GP with a lump in the breast, coded as benign neoplasm of
the breast (X79) or breast lump (X19), were registered with breast cancer during follow
up in 11.5% (n = 7 of 61) and 5.9% (n=64 of 1,080) of cases, respectively. Of the women
presenting with fear of breast cancer or with a family history of breast cancer, only 0.9%
(N = 3 of 340) and 2.1% (N = 6 of 281) were diagnosed with breast cancer during follow up.
table 6. Proportion of referrals for annual screening and genetic counseling in accordance with guideline
x26 fear of breast Cancer a29.2 family history of breast Cancer
moderately increased lifetime
risk (20-30%)Referrals for
annual Screening n (%)
Stongly increased lifetime risk
(>30%)Referrals
for Genetic Counseling n (%)
moderately increased lifetime
risk (20-30%)Referrals for
annual Screening n (%)
Stongly increased lifetime risk
(>30%)Referrals
for Genetic Counseling n (%)
Referral in accordance with guideline1 incl. bRCa positive family2
6 (18%)0
4 (33%)5
26 (25%)3
3 (13%)2
Referral not according to guideline3
5 (15%) 3 (42%) 14 (13%) 6 (26%)
Information not assessable4
23 (68%) 3 (25%) 65 (62%) 14 (61%)
total Referrals 34 12 105 23
1Based on the EMR data the patient was considered at risk and was referred as instructed according to the Guideline “Diagnosing Breast Cancer” from the Dutch college of general practitioners (NHG)2Family with proven mutation in BRCA1 or BRCA2 gene3Based on the EMR data the patient had an increased risk but was not referred according to the guideline or did not have an increased risk but had a referral 4Information in the EMR was ambiguous making it impossible to decide whether or not a patient had an increased risk
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
92
DISCuSSIon
SummaryThis dynamic cohort study in the Julius GP’s Network database was performed to
determine incidence rates, management, and diagnostic outcomes for women in primary
care with concerns related to breast cancer. The overall incidence rate for breast cancer
related concerns is 25.9 per 1,000 women per year, the majority of them presenting
with physical signs and symptoms of the breast (85.3% or 22.1 per 1,000 per year).
The referral rate for further investigation is just over 50%, for symptomatic as well as
for asymptomatic women. About a quarter of referrals for annual screening or genetic
counseling were determined as performed according to the guideline, about a quarter
were not, and the information in the EMR was insufficient for the remainder. Because a
large proportion of cases could not be assessed, it is unclear how many women that should
have been referred for annual screening or genetic counseling because of an increased
lifetime risk were missed. Among the women presenting with complaints of the breast
or concerns about developing breast cancer, 2.6% were diagnosed with breast cancer
during follow-up. Women presenting with a lump in the breast were most frequently
diagnosed with breast cancer during follow up.
Strengths and limitationsMajor strengths of the present study are the size of the cohort and the quality of the data.
The JGPN database is comprised of well-documented information of enlisted patients.
Characteristics of these patients did not differ from the overall Dutch population, and the
main characteristics of the GPs were comparable with total Dutch GPs with respect to
age, gender, part-time and fulltime workers, and practice in urban and rural areas [20]. To
the best of the present authors’ knowledge, this is the first report on the actual referral
practice of GPs for this patient group and the first report where GP referral behaviour is
compared with the national guidelines.
The use of routine care data has limitations that should be kept in mind when
interpreting results. These limitations relate to the patient, the GP, coding, and the EMR.
The patient decides whether and when she consults her GP with a breast cancer related
concern. Others might have the same concerns without visiting the doctor. Secondly, GPs
decide what they consider to be the most important symptom presented by their patients
during consultation to be registered in the EMR. This might result in one GP registering
a symptom as pain in the breast, while another registers worry about breast cancer. A
third uncertainty in the present data is the ICPC-1 coding system. Misclassification, lack of
available codes, and differences in classification between years and GPs cannot be ruled
out. This may have resulted in either an overestimation or underestimation of the true
incidence rates. However, we know that in larger populations and among larger groups
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
93
of GPs sharing their data, differences in presentation, prioritization, and lack of detail
become less significant [27]. Furthermore, we know that variation in registration rules in
different EMR systems may be a cause of heterogeneity in extracted data. To minimise
the effect of heterogeneity in the present study, we restricted data collection to only one
type of EMR. Also, patient preferences discussed during the consultation for or against
referral are not captured using ICPC codes. Finally, restrictions in registration possibilities
in existing EMR systems should be taken into account (e.g., the limited possibilities in
Dutch EMR systems to register family history information, leading to possible registration
of this information as plain text). In the present study, this information (in plain text)
appeared to be insufficient to determine increased lifetime risk for a large proportion
of cases.
However, the fact that information concerning family history is missing in a large part
of EMR files could also mean that the information is available but not registered by the
GP, that the GP lacks knowledge in this area or the guidelines are not readily available or
unclear, or that the GP has other reasons to deviate from the guidelines. Previous studies
have shown that GPs currently lack knowledge and confidence in this area [28] and that
there is an urgent need for a genetics curriculum for postgraduate and continuing general
practice education in this area [29].
Another limitation might be that our study population was comprised also of
women who participate in the national breast cancer screening program (age category
50–75 years). Within the screening programme, abnormal mammography results are
reported to GPs together with advice to refer to a multidisciplinary breast clinic. Inclusion
of these women has probably resulted in an overestimation of the “true” rate of women
that would consult their GP with complaints of the breast or concerns about developing
breast cancer. However, an average Dutch general practice is faced with only three
positive screening referrals per year, only one of whom will be diagnosed with breast
cancer [30]. Note that these women may also be less inclined to visit their GP with
concerns about breast cancer in between screening mammograms.
Finally, in the present study, the follow-up time was restricted to a maximum of
three years after the first consultation. This is likely to be too short to determine the
total number of women developing breast cancer, leading to an underestimation of
this outcome.
Comparison with existing literatureThe incidence rates for breast cancer related concerns found in the present study are
considerably higher than those in older data presented in two UK studies (data from
1995–1996) but slightly lower than in another Dutch study (data from 1985–2003).
However, we included patients that the GP registered as having breast cancer at first
consultation (X76 and X76.01), in contrast to the other studies described here.
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
94
The Bridge Study group recorded presentation rates of breast symptoms in 34 general
practices in South Wales in 1995–1996. These presentation rates ranged from 1.9–14.8
patients per GP per year (median = 6.5): 46.4% with breast lump, 28.2% with breast
pain, 16.2% with lumpiness, and 5.5% with nipple discharge (The BRIDGE study group
1999). This translates to 2.2–17.0 per 1,000 women per year (median 7.4), calculating with
an average general practice size of 1,712 in 1997 in the UK, (Royal College of General
Practitioners 2004) and 51% of the population being female in 2001 [32]. Note that
these numbers only include women with physical complaints of the breast (in our study
22.1 per 1,000 women per year). In a study by Newton [33] 257 GPs from Sheffield (UK)
participated and recorded a mean number of 2.05 consultations over a 4-week period
in 1995. If annual figures are extrapolated from these data, they suggest that each GP
sees 15.8 women with new breast problems per year, or 18 per 1,000 women per year.
These numbers include all “breast problems,” including women that present with a
positive family history. This study does not explicitly present women with fear of breast
cancer as a separate group.
The incidence rates found in the present study are slightly lower than the findings of
Eberl, [34] who studied routine family practice data from Dutch GP practices between
1985 and 2003 on breast symptoms. Breast symptoms were reported in 29.7 consultations
per 1,000 active female patients per year, with breast pain (13 per 1,000 per year) and
breast mass (9 per 1,000 per year) being the most common breast-related complaints.
Note that this excludes 3.6 per 1,000 per year who consult their GP for fear of breast
cancer (compared with 2.0 per 1,000 in the present study). This means that Eberl found an
overall incidence rate of 33.3 per 1,000 women per year, compared to 25.9 per 1,000 per
year in the present study. Incidence rates found in her study were 2.5 per 1,000 women
per year for nipple complaints and 4.3 per 1,000 per year with other breast complaints.
A possible explanation for the differences in incidence rates found could be that these
are a reflection of the incidence trends of breast cancer. Toriola [35] summarizes these
for the US as having four distinct patterns; an increase over 45 years (1943–1979), a
more rapid increase over 20 years (1980–1999) – punctuated by a gradual (2000–2002),
then sharp decline (2002–2003) –, and a post 2003 period of stable incidence rates.
Another explanation may be that there is a growing awareness in recent years among
women concerning breast cancer, which prompts more GP visits. Furthermore there may
be slight differences in ICPC codes used in the different studies, resulting in differing
inclusion conditions.
Concerning referral patterns, limited literature was available. The Bridge Study group,
mentioned above, reported that, in 1995–1996, 55% of all patients were referred[36].
This is in line with the results found in the present study, with an overall referral rate
54.3%. Newton [33] found that, in 1995, at an initial consultation for breast symptoms,
GPs referred approximately one-third of women to secondary care.
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
95
Eberl [34] who studied routine family practice data from Dutch GP practices between
1985 and 2003 on breast symptoms (mean registration time 5.6 years), found that of the
women complaining of breast symptoms, 81 (3.2%) had breast cancer diagnosed, which is
much lower than the 10% found in the present study. Note that we included patients who
were coded as having breast cancer at first consultation, while Eberl probably did not (no
information available in the published material). Leaving out these patients means that only
2.6% developed breast cancer ((450 - 341)/4,240); however, these numbers include women
with a positive screening advise from the national breast cancer screening program. Leaving
out the patients in this age category (278 women aged 50–75 years) gives a percentage of 4%
((450 - 341 + 278)/4,240) as developing breast cancer, which closely resembles the Eberl data.
A recent UK CPRD study by Walker[37] revealed that the PPV of breast cancer
(diagnosed between 2000 and 2009) with a breast lump presented at the GP was 4.8%
in women aged 40–49 years, rising to 48% in women aged >70 years. PPVs were lower
in women who also reported breast pain. Hippesley-Cox[38] found, also using UK Primary
Care data (2000 – 2012) that a breast lump was associated with a 51-fold increased risk
of breast cancer. These studies confirm our finding that GPs should be aware of women
presenting with a lump in the breast because they are most frequently diagnosed with
breast cancer (5.9%, figure 2) during follow up.
ConCLuSIonS
This study demonstrates breast cancer related concerns are presented in Primary Care
frequently. The referral rate for further investigation is over 50%. There is room to improve
GP Management, mainly for women with an increased lifetime risk of developing breast
cancer. Information regarding the family history is often missing in the EMR. Furthermore,
since only 2.6% of women with breast cancer related concerns were diagnosed with
breast cancer during follow-up time, substantial reduction of unnecessary concerns is
plausible by improving identification of those without an increased risk of breast cancer.
Unanswered Questions and Future ResearchFuture research aimed at the underlying mechanism of low adherence to guidelines for
referral by GPs in case of an increased lifetime risk, and poor registration of the family
history in the EMR, could assist in the development of effective interventions to improve
referral practice of patients at increased genetic risk.
Strategies to be evaluated include: (1)increasing awareness of the importance of
registering an increased lifetime risk (2) optimizing the availability of [online] up-to-date
and easy to use referral guidelines, for example by integrating them into the EMR; (3)
developing training for GPs in taking, assessing and registering a family history and (4)
enabling the registry of a family history within the context of the EMR.
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
96
REfEREnCES1. Ferlay J, Soerjomataram I I, Dikshit R, et al. Cancer incidence and mortality worldwide:
sources, methods and major patterns in GLOBOCAN 2012. Int J Cancer 2014;136:E359–86. doi:10.1002/ijc.29210
2. Consedine NS, Magai C, Krivoshekova YS, et al. Fear, anxiety, worry, and breast cancer screening behavior: a critical review. Cancer Epidemiol Biomarkers Prev 2004;13:501–10.http://www.ncbi.nlm.nih.gov/pubmed/15066912 (accessed 13 Dec2015).
3. Bennett, P.Parsons, E.Brain, K.Hood K. Long-term cohort study of women at intermediate risk of familial breast cancer: experiences of living at risk. 2010.
4. Gibbons A, Groarke A. Can risk and illness perceptions predict breast cancer worry in healthy women? J Health Psychol Published Online First: 23 February 2015. doi:10.1177/1359105315570984
5. Neal RD. Do diagnostic delays in cancer matter? Br J Cancer 2009;101 Suppl :S9–12. doi:10.1038/sj.bjc.6605384
6. Maclean R, Jeffreys M, Ives A, et al. Primary care characteristics and stage of cancer at diagnosis using data from the national cancer registration service, quality outcomes framework and general practice information. BMC Cancer 2015;15:500. doi:10.1186/s12885-015-1497-1
7. Ardern-Jones A, Kenen R, Eeles R. Too much, too soon? Patients and health professionals’ views concerning the impact of genetic testing at the time of breast cancer diagnosis in women under the age of 40. Eur J Cancer Care (Engl) 2005;14:272–81. doi:10.1111/j.1365-2354.2005.00574.x
8. Carroll JC, Rideout AL, Wilson BJ, et al. Genetic education for primary care providers: improving attitudes, knowledge, and confidence. Can Fam Physician 2009;55:e92–9.http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=2793208&tool=pmcentrez&rendertype=abstract (accessed 17 Jan2015).
9. McCann S, MacAuley D, Barnett Y. Genetic consultations in primary care: GPs’ responses to three scenarios. Scand J Prim Health Care 2005;23:109–14. doi:10.1080/02813430510015259
10. Nippert I, Harris HJ, Julian-Reynier C, et al. Confidence of primary care physicians in their ability to carry out basic medical genetic tasks-a European survey in five countries-Part 1. J Community Genet 2011;2:1–11. doi:10.1007/s12687-010-0030-0
11. Vasen HFA, Möslein G, Alonso A, et al. Recommendations to improve identification of hereditary and familial colorectal cancer in Europe. Fam Cancer 2010;9:109–15. doi:10.1007/s10689-009-9291-3
12. Burke W, Culver J, Pinsky L, et al. Genetic assessment of breast cancer risk in primary care practice. Am J Med Genet A 2009;149a:349–56. doi:10.1002/ajmg.a.32643
13. Campbell H, Holloway S, Cetnarskyj R, et al. Referrals of women with a family history of breast cancer from primary care to cancer genetics services in South East Scotland. Br J Cancer 2003;89:1650–6. doi:10.1038/sj.bjc.6601348
14. Febbraro T, Robison K, Wilbur JS, et al. Adherence patterns to National Comprehensive Cancer Network (NCCN) guidelines for referral to cancer genetic professionals. Gynecol Oncol 2015;138:109–14. doi:10.1016/j.ygyno.2015.04.029
15. Bell RA, McDermott H, Fancher TL, et al. Impact of a randomized controlled educational trial to improve physician practice behaviors around screening for inherited breast cancer. J Gen Intern Med 2015;30:334–41. doi:10.1007/s11606-014-3113-5
16. Dentler K, Numans ME, ten Teije A, et al. Formalization and computation of quality measures based on electronic medical records. J Am Med Inform Assoc 2014;21:285–91. doi:10.1136/amiajnl-2013-001921 [doi]
17. Hamoen EH, Reukers DF, Numans ME, et al. Discrepancies between guidelines and clinical practice regarding prostate-specific antigen testing. Fam Pract 2013;30:648–54. doi:10.1093/fampra/cmt045
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
97
18. Kasteleyn MJ, Wezendonk A, Vos RC, et al. Repeat prescriptions of guideline-based secondary prevention medication in patients with type 2 diabetes and previous myocardial infarction in Dutch primary care. Fam Pract 2014;31:688–93. doi:10.1093/fampra/cmu042
19. Lacourt TE, Houtveen JH, Smeets HM, et al. Infection load as a predisposing factor for somatoform disorders: evidence from a dutch general practice registry. Psychosom Med 2013;75:759–64. doi:10.1097/PSY.0b013e3182a3d91f
20. Hak E, Rovers MM, Sachs APE, et al. Is asthma in 2-12 year-old children associated with physician-attended recurrent upper respiratory tract infections? Eur J Epidemiol 2003;18:899–902.http://www.ncbi.nlm.nih.gov/pubmed/14561050 (accessed 4 Aug2014).
21. Van der Zanden G. Quality Assessment of Medical Health Records using Information Extraction, Master’s thesis. 2010.
22. Bentsen BG. International classification of primary care. Scand J Prim Health Care 1986;4:43–50.23. Dutch College of General Practitoners. Guideline adequate EHR registry. Revised version 2013.
Available at: https://www.nhg.org/themas/publicaties/richtlijn-adequate-dossiervorming-met-het-epd.
24. Dutch breast cancer screening programme by the National Institute for Public Health and the Environment. http://rivm.nl/en/Topics/B/Breast_cancer_screening_programme
25. Zonderland HM, Tuut MK, den Heeten GJ, et al. [Revised practice guideline ‘Screening and diagnosis of breast cancer’]. Ned Tijdschr Geneeskd 2008;152:2336–9.http://www.ncbi.nlm.nih.gov/pubmed/19024064 (accessed 17 Jan2015).
26. Schäfer, W.L.A., Berg, M.J. van den, Groenewegen PP. De werkbelasting van huisartsen in internationaal perspectief. Huisarts Wet;3:94–101.http://www.henw.org/archief/volledig/id12123-de-werkbelasting-van-huisartsen-in-internationaal-perspectief.html.
27. van Bommel MJ, Numans ME, de Wit NJ, et al. Consultations and referrals for dyspepsia in general practice--a one year database survey. Postgrad Med J 2001;77:514–8.http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=1742094&tool=pmcentrez&rendertype=abstract (accessed 18 Jan2015).
28. Watson E, Clements A, Yudkin P, et al. Evaluation of the impact of two educational interventions on GP management of familial breast/ovarian cancer cases: a cluster randomised controlled trial. Br J Gen Pract 2001;51:817–21.http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=1314127&tool=pmcentrez&rendertype=abstract (accessed 17 Jan2015).
29. Houwink EJF, Henneman L, Westerneng M, et al. Prioritization of future genetics education for general practitioners: a Delphi study. Genet Med 2012;14:323–9. doi:10.1038/gim.2011.15
30. Verbeek ALM, van Dijck JAAM, Kiemeney LALM, et al. [Responsible cancer screening]. Ned Tijdschr Geneeskd 2011;155:A3934.http://www.ncbi.nlm.nih.gov/pubmed/22085573 (accessed 17 Jan2015).
31. The presentation and management of breast symptoms in general practice in South Wales. The BRIDGE Study Group. Br J Gen Pract 1999;49:811–2.http://www.pubmedcentral.nih.gov/articlerender.fcgi?artid=1313533&tool=pmcentrez&rendertype=abstract (accessed 17 Jan2015).
32. Palgrave Macmillan. Social Trends. Available at: http://www.palgrave-journals.com.proxy.library.uu.nl.st/journal/v39/n1/full/st20093a.html. 2009.
33. Newton P, Hannay DR, Laver R. The presentation and management of female breast symptoms in general practice in Sheffield. Fam Pract 1999;16:360–5.http://www.ncbi.nlm.nih.gov/pubmed/10493705 (accessed 17 Jan2015).
34. Eberl MM, Phillips RL, Lamberts H, et al. Characterizing breast symptoms in family practice. Ann Fam Med 2008;6:528–33. doi:10.1370/afm.905
35. Toriola AT, Colditz GA. Trends in breast cancer incidence and mortality in the United States: implications for prevention. Breast Cancer Res Treat 2013;138:665–73. doi:10.1007/s10549-013-2500-7
5
Prim
ary
Ca
re m
an
ag
emen
t of w
om
en w
ith b
rea
st Ca
nC
er r
elated
Co
nC
ern
s
98
36. The BRIDGE Study Group. The presentation and management of breast symptoms in general practice in South Wales. Br J Gen Pract 1999;:811–2.
37. Walker S, Hyde C, Hamilton W. Risk of breast cancer in symptomatic women in primary care: a case-control study using electronic records. Br J Gen Pract 2014;64:e788–93. doi:10.3399/bjgp14X682873
38. Hippisley-Cox J, Coupland C. Symptoms and risk factors to identify women with suspected cancer in primary care: derivation and validation of an algorithm. Br J Gen Pract 2013;63:e11–21. doi:10.3399/bjgp13X660733
Strategies & Solutions for improving data quality and enabling data
reuse and sharing
CHaPtERA new coding system for metabolic disorders demonstrates gaps in the international disease classifications ICD-10 and SNOMED-CT which can
be barriers to genotype-phenotype data sharing
Annet Sollie, Rolf H. Sijmons, Dick Lindhout, Ans T. van der Ploeg, M. Estela Rubio Gozalbo, G. Peter A. Smit, Frans Verheijen, Hans R. Waterham,
Sonja van Weely, Frits A. Wijburg, Rudolph Wijburg, and Gepke Visser
6
Published in Human mutation, July 2013 as: Sollie A, Sijmons RH, Lindhout D, van der Ploeg AT, Rubio Gozalbo ME, Smit GP, Verheijen F, Waterham HR, van Weely S, Wijburg FA, Wijburg R, Visser G. A new coding system for metabolic disorders demonstrates gaps in the international disease classifications ICD-10 and SNOMED-CT, which can be barriers to genotype-phenotype data sharing. Hum Mutat. 2013 Jul;34(7):967-73. doi: 10.1002/humu.22316.
105
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
abStRaCt
Data sharing is essential for a better understanding of genetic disorders. Good phenotype
coding plays a key role in this process. Unfortunately, the two most widely used coding
systems in medicine, ICD-10 and SNOMED-CT, lack information necessary for the detailed
classification and annotation of rare and genetic disorders. This prevents the optimal
registration of such patients in databases and thus data-sharing efforts. In order to improve
care and to facilitate research for patients with metabolic disorders we developed a new
coding system for metabolic diseases with a dedicated group of clinical specialists. Next,
we compared the resulting codes with those in ICD and SNOMED-CT. No matches were
found in 76% of cases in ICD-10 and in 54% in SNOMED-CT. We conclude that there
are sizable gaps in the SNOMED-CT and ICD coding systems for metabolic disorders.
There may be similar gaps for other classes of rare and genetic disorders. We have
demonstrated that expert groups can help in addressing such coding issues. Our coding
system has been made available to the ICD and SNOMED-CT organizations as well as to
the Orphanet and HPO organizations for further public application and updates will be
published online (www.ddrmd.nl and www.cineas.org).
106
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
IntRoDuCtIon
Data sharing is essential for a better understanding of rare genetic disorders and the
underlying genetic defects. Good phenotype coding plays a key role in this process and
also in general in processes where phenotype data needs to be entered into clinical
registries, genotype-phenotype databases, and biobanks, and shared between them.
Such initiatives to register, combine and exchange clinical and research data are pivotal
in supporting research and improving health care [Jones et al., 2011; Richesson and
Vehik, 2010].
Rare diseases are life threatening or chronically debilitating diseases with a
prevalence of up to 5 per 10,000 inhabitants in the European Union (EU). It is
estimated that there are at least 5,000 rare diseases, many of them genetic, affecting
6-8% of the total population in the EU, which implies a minimum 27 million people in
the EU are affected [European Medicines Agency (EMA), 2007]. In the United States,
the Rare Disease Act of 2002 also defined rare disease according to prevalence,
specifically “any disease or condition that affects less than 200,000 persons in
the United States”, or about 1 in 1,500 people. Although there are no disease-
modifying therapies for most rare diseases, the passing of the 1983 U.S. Orphan
Drug Act (ODA) [Food and Drug Administration] and European legislation in 2000
[European Parliament. Regulation (EC) No. 141/2000 of the European Parliament
and the Council of 16 December 1999 on orphan medicinal products] stimulated
new research lines by creating financial incentives and other supportive measures
for developers of new drugs to treat people with rare diseases [Talele et al., 2010].
It is expected that many more rare diseases will become amenable to treatment
within the next few decades.
The need to improve research and care in the field of rare disorders, which can be
strongly supported by the sharing and combining of data on these rare patients, has
also been recognized by the Council of the European Union. Through their European
Action in the Field of Rare Diseases [Official Journal of the European Union, Council
recommendation of 8 June 2009 on an action in the field of rare disease], signed in 2009,
the EU member states committed themselves to establishing and implementing a national
rare disease action plan and to cooperating at a European level on this health issue. The
European Action stated that member states should “aim to ensure that rare diseases
are adequately coded and traceable in all health information systems, encouraging
an adequate recognition of the disease in the national healthcare and reimbursement
systems based on the ICD.”
Unfortunately, the two most widely used coding systems in medicine– ICD (the WHO’s
International Classification of Diseases, www.who.int/classifications/icd) and SNOMED-CT
(Systematized Nomenclature of Medicine Clinical Terminology, www.ihtsdo.org) – lack
107
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
essential details for classifying and annotating rare and hereditary disorders. This is a
barrier to the optimal registration of patients with these disorders in databases, and
to much needed data-sharing efforts, such as those in the Human Variome Project
(http://www.humanvariomeproject.org).
Our study addresses this problem for metabolic diseases, a particular hereditary
subgroup of rare disorders. Metabolic diseases, also referred to as inborn errors of
metabolism, are generally monogenic defects resulting in a deficient activity in an
enzyme or a transporter in a pathway of cellular metabolism [Scriver et al., 2001]. The
number of recognized metabolic diseases is continually increasing due to advances
in knowledge and diagnostic laboratory techniques. Most metabolic diseases are
extremely rare (< 1 per 50,000 inhabitants), although all metabolic diseases combined
have an estimated, relatively high birth prevalence of up to 1 per 800 newborns
[Sanderson et al., 2006]. In the Netherlands, we decided to build a registry for
patients with metabolic disorders and also to optimize the codes for national use
in medical and clinical genetics. With these purposes in mind, we developed with
a dedicated group of clinical specialists a clinically oriented annotation system
for metabolic disorders based on two existing national coding systems. To assess
the potential value of adding our annotation system to ICD and SNOMED-CT, we
compared the three systems and identified large gaps in both ICD and SNOMED-CT.
To the best of our knowledge, we are the first to actually quantify these gaps for a
specific field of rare diseases.
matERIaLS & mEtHoDS
Study OverviewWe combined and expanded two existing coding systems for metabolic diseases, the
DDRMD (Dutch Diagnosis and Registration of Metabolic Diseases, www.ddrmd.nl)
and a subset of CINEAS (Dutch center for disease code development and distribution
to the clinical genetics community, www.cineas.org) to develop a more detailed and
strongly clinically oriented coding system. The DDRMD was set up by specialists
in metabolic disorders, whereas CINEAS was initiated by clinical geneticists. Both
systems were originally developed independent of each other and born out of the
need to have more extensive coding system available than the ones offered by
SMOMED and ICD. The primary purpose of each of our original coding systems was
improving patient classification and retrieval. We used the DDRMD as a starting
point for extending the coding system of metabolic diseases, because this system
had already been used for more than ten years by metabolic specialists in clinical
practice. We matched and enriched these systems in a three-step process, exemplified
in a flowchart (Figure 1).
108
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
A list of criteria for including codes in the coding system was drawn up for the
matching process (Table 1).
table 1. List of Criteria for Including Codes in our Coding System
no. Criteria
1 The disease has to be a separate clinical entity
2 Is must be likely that the disease is a separate clinical entity; just one case report in the literature is not enough, unless an enzyme deficiency or transport defect was demonstrated
3 No separate entries for gene defects; a gene can, however, be connected to a disease (no specific mutation is mentioned)
4 No specific entries for groups of diseases
5 One enzyme defect leads to only one separate code
To facilitate cross-linking, but also to investigate the extent to which codes were
lacking in ICD and SNOMED-CT, we checked and updated existing mappings to these
two international systems.
6
practice. We matched and enriched these systems in a three-step process, exemplified in a
flowchart (Figure 1).
A list of criteria for including codes in the coding system was drawn up for the
matching process (Table 1).
Table 1. List of Criteria for Including Codes in our Coding System
No. Criteria
1 The disease has to be a separate clinical entity
2 Is must be likely that the disease is a separate clinical entity; just one case report in the
literature is not enough, unless an enzyme deficiency or transport defect was
demonstrated
3 No separate entries for gene defects; a gene can, however, be connected to a disease
(no specific mutation is mentioned)
figure 1. Flowchart of building the metabolic disease coding system and quantifying the gaps in ICD-10 and SNOMED-CT
109
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
DDRMD (background, origin, objective)The Dutch Diagnosis and Registration of Metabolic Diseases is a collaborative project of
all the clinical metabolic centers in the Netherlands. It was started in 2001 and over 5,000
patients have been registered so far, with almost 300 different metabolic diseases. The main
reason for initiating the DDRMD was that despite the various diagnosis registration systems
used in hospitals, it was proving difficult to retrieve patients with metabolic diseases from
these registers. Since there was no disease-specific registration for metabolic diseases, it
was impossible to analyze relevant patient data, either for research or for care purposes.
In the DDRMD, patient data are registered by one metabolic specialist per metabolic
center (see Figure 2)
via a secure web server (SSL). In addition, relevant data on newborns referred
because of an abnormal neonatal screening result indicative for metabolic disease are
also included. The data are used to facilitate research on metabolic diseases and to
provide information on the outcome of the Dutch newborn screening procedure for
metabolic diseases.
CINEAS (background, origin, objective)CINEAS is the Dutch center for disease code development and distribution to the clinical
genetics community. It was initiated by the eight clinical genetics centers responsible for
genetic counseling and diagnostics in the Netherlands in 1992 [Zwamborn-Hanssen et al.,
1997]. It is used in daily practice by the Dutch clinical geneticists and genetic counselors
to assign diseases to patients. Presently, the 55th edition of CINEAS lists more than 5,500
diseases, most of them rare, and the metabolic diseases form a distinct subset (Figure 3).
figure 2. Data model for DDRMD**Dutch Diagnosis % Registration of Metabolic Diseases; bsn = dutch national identification number;nbs = newborn screening
110
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps 9
via a secure web server (SSL). In addition, relevant data on newborns referred because of an
abnormal neonatal screening result indicative for metabolic disease are also included. The
data are used to facilitate research on metabolic diseases and to provide information on the
outcome of the Dutch newborn screening procedure for metabolic diseases.
CINEAS (background, origin, objective) CINEAS is the Dutch center for disease code development and distribution to the
clinical genetics community. It was initiated by the eight clinical genetics centers responsible
for genetic counseling and diagnostics in the Netherlands in 1992 [Zwamborn-Hanssen et al.,
1997]. It is used in daily practice by the Dutch clinical geneticists and genetic counselors to
assign diseases to patients. Presently, the 55th edition of CINEAS lists more than 5,500
diseases, most of them rare, and the metabolic diseases form a distinct subset (Figure 3).
figure 3. Data model CINEAS*- only core tables.*Dutch national disease code development distribution center for the clinical genetics community
A number of Dutch diagnostic DNA laboratories use the CINEAS system as well, and
recently the Danish genetics centers have decided to adopt CINEAS. Each new edition of
the list contains new disease entries submitted by users, after they have been discussed
and approved by a group of experts. The entire process of submitting and adding new
entries to the database is supported by a website (www.cineas.nl or www.cineas.org),
a paid professional curator and a quickly responding national expert panel, which has
reduced throughput time to an average of two weeks. Local system administrators
upload new editions to their own patient information systems, and the website facilitates
searching of the CINEAS database and is used to publish the new editions. Codes are
never removed from the system, but can be made obsolete and thus no longer assigned
to patients. Entry, modifications and obsoletion of codes including dates are saved in
the Diagnosis History. Crosslinks are provided to OMIM, Online Mendelian Inheritance
in Man, a catalogue of hereditary disorders and their genes (www.omim.org), and to
SNOMED and ICD. Although the codes in CINEAS, including the metabolic codes, are
non-hierarchical (see Discussion), individual codes can easily be found in the system using
the onboard search engine.
111
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
Existing Coding SystemsThe most widely used system in medical practice is the ICD, published by the World
Health Organization (version 9 published in 1977, or version 10 in 1999). It is categorized
by the affected organ system, which makes it difficult to use for diseases in which
more than one organ is affected, as is the case for many rare genetic diseases. The
WHO is working on the revision of ICD-10, with the aim of publishing ICD-11 in 2015
(www.who.int/classifications/icd/revision).
SNOMED-CT is a coding system which has been adopted by many hospital information
systems and standards organizations worldwide as a key coding system. Since 1974,
SNOMED-CT has evolved from a pathology-specific nomenclature into a healthcare
terminology system. There are many studies that have shown the value of SNOMED-CT
in theory, but studies on its use in clinical practice are relatively rare [Cornet and de
Keizer, 2008].
Assessment of gaps for metabolic diesease in ICD and SNOMED-CTDuring the final steps of our matching process (Figure 1), for each code in our system we
chose the most appropriate ICD-10 code as a cross-link, using the online WHO browser
(http://apps.who.int/classifications/icd10/browse/2010/en) and searching with disease
names and synonyms. When no specific disease code was available, we chose a group
name or non-specific code based on the etiology, for example “E79.8 Other disorders
of purine and pyrimidine metabolism” or “E88.8 other specified metabolic disorder”
(Table 2).
112
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
table 2. Example from Our Coding System
Disease Synonyms Identifier omIm ICD-10 SnomED
adenylosuccinase deficiency
adenylosuccinate lyase 1573 103050 E79.8 73843004
aldolase-B deficiency hereditary fructose intolerance 1318 229600 E74.1 20052008
alpha-aminoadipic aciduria
2-amino-/ 2-oxoadipic aciduria 2-aminoadipic 2-oxoadipic
1599 204750 E88.8 not available
alpha-aminoadipic semialdehyde dehydrogenase deficiency
pyridoxine dependent epilepsy AASA folinic acid responsive convulsions antiquitine gene
2286 266100 E88.8 not available
alpha-N-acetylgalactosaminidase deficiency
NAGAneuroaxonal dystrophiaSchindler disease
1539 609241 E88.8 double codes: 238048001
and 230365004 with 3 subcodes
In addition, we mapped as many entries as possible to the SNOMED-CT International
Edition of January 2011, using CliniClue Explore (http://www.cliniclue.com/) and by
searching SNOMED-CT using disease names and synonyms. We recorded all the
unambiguous mappings and all the possible mappings if more than one SNOMED-CT
code was available. Finally we calculated the gaps for metabolic diseases in ICD-10 and
SNOMED-CT as percentages of codes with matches in our coding system.
RESuLtS
We have developed a specific coding system for metabolic diseases, currently containing
almost 300 different disorders. Every item in our system has a unique identifier and
includes a disease name, existing synonyms, and mappings to the OMIM catalogue,
ICD-10 and SNOMED-CT (example in Table 2). For the unique identifiers we used the
existing CINEAS codes. The mappings to the other coding systems can be used for data
exchange with other databases and provide extra search possibilities.
Note that we deviated from our inclusion criteria (Table 1) for the group of mitochondrial
diseases. Apart from the separate respiratory chain disorders, we added two general
codes for diseases caused by mitochondrial DNA variations. This is because this particular
area is evolving rapidly and clear classification is not yet possible in all cases. We expect
to be able to create more specific entries for these diseases in the coming years.
For 214 (76%) of the diseases in our coding system, there was no specific
matching ICD-10 code and only an ICD-10 group name that was too general for our
113
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
clinical classification purposes was available (e.g. ICD-10 code E88.8 “other specified
metabolic disorders”).
For 155 (54%) of our codes, it was not possible to map unambiguously to SNOMED-CT,
because for 81 codes (29%), there was no SNOMED-CT code available and for 72 codes
(25%) SNOMED-CT contained double codes. These duplicates were counted mostly
because the disease and enzyme deficiency were given separate codes in SNOMED-CT,
but also because too much detail in SNOMED-CT made it difficult to distinguish between
group codes and subcodes. An example is the disease “alpha-N-acetylgalactosaminidase
deficiency” for which SNOMED provided mistakingly two codes, one with three subcodes
(Table 2). This shows that, despite the size of the SNOMED system, the unambiguous
detail needed in clinical practice for metabolic diseases is often not available.
We aim to publish incidences and prevalence of individual metabolic diseases in our
coding system online in the spring of 2013.Our coding system is being continuously
updated and is published on www.ddrmd.nl and www.cineas.org in pdf and xml formats.
CINEAS and DDRMD keep existing as two different organizations, each with a different
purpose, now sharing the code system for metabolic disorders. Requests for additions to
and alterations of DDRMD and/or CINEAS users are submitted by email to [email protected]
or [email protected] or by using an online web-form on the CINEAS website for
registered organizations. These requests are subsequently discussed in the national
CINEAS online expert panel for approval. The national coordinator of DDRMD is now a
member of the CINEAS expert panel. The coding system has already been updated using
these procedures and now contains 285 diseases.
Continued funding for the classification efforts is provided by CINEAS (Dutch national
disease code development and distribution center for the clinical genetics community).
Our novel coding system has recently been donated to ICD, SNOMED-CT, and also to
the Human Phenotype Ontology (www.human-phenotype-ontology.org), a promising
emerging ontology for phenotypic abnormalities, and to Orphanet (www.orpha.net)
an important reference portal for information on rare diseases and orphan drugs, for
further public application. Continuous updates of our system will be published online.
DISCuSSIon
The most widely used classification and coding systems in medical databases are ICD
and SNOMED-CT. Historically, the focus in the development of these systems has been
directed towards classifying common disorders. The development and updating process
for international broad medical coding systems is a highly demanding task and we
acknowledge the important contribution of ICD and SNOMED-CT to the annotation
of common disorders. However, annotation for rare disorders has been left behind.
Collectively, this group is large and growing steadily due to the identification of new
114
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
diseases and improved clinician awareness. Our study demonstrates large gaps in both ICD
(76%) and SNOMED-CT (54%) for metabolic disorders. Based on our clinical experience,
we suspect that there may be similar gaps for other types of rare disorders. Such gaps
are a barrier to database- and data-sharing efforts. We have shown that with the help
of dedicated clinicians and code development agencies, the problem of coding gaps for
rare disorders can be successfully addressed.
Developing codes for a rare field of medicine has special challenges. We observed
during the development of our system that existing hierarchical, ‘tree’, classification
structures, such as those used in SNOMED-CT and in ICD-10, were proving inconvenient
for our purpose. Such structures, when well developed for the particular branches, allow
for the selection of patients from groups of disorders rather than those with particular
individual disorders. However, in the rare field of metabolic disorders, these existing tree
structures turned out to be problematic and we dropped our initial hierarchical approach
for several reasons. Firstly, several diseases did not fit into any specific group or category
leading to a risk of misclassification. Secondly, several diseases fitted into more than one
group or category leading to significant risk of double entries for the same disorder.
Furthermore, given the explosion of knowledge in this field of rare genetic diseases,
extensive and continuous expertise is needed to update the accuracy of a specialist tree
structure. Given the aim of our coding system to assign diagnostic end codes to patients
and to obtain incidence and prevalence data from our registry on individual metabolic
diseases, a non- hierarchical design turned out to be functional. With growing knowledge
on underlying molecular pathways, well fitting metabolic branches of the coding trees are
likely to be developed in the future in the international community and this will support
better data handling on the level of groups of metabolic disorders.
The World Health Organization has signaled the need to improve ICD-10 for use in the
field of rare diseases. A special Topic Advisory Group (http://www.who.int/classifications/
icd/TAGs/en/index.html) has been assigned to the subject of rare diseases to advise the
WHO on the current updating and revision process from ICD-10 to ICD-11 (anticipated
publication in 2015). We recently donated our work to both the ICD and SNOMED-CT
communities to support further code development, and to the Orphanet and HPO
organizations as well. These organizations are also contributing to solving annotation
problems. The Orphanet organization (www.orpha.net) has stressed the need to provide
well-designed codes for rare diseases, especially for the purposes of data sharing and
it puts much effort into this field [Rath et al., 2012]. The Human Phenotype Ontology
(HPO, http://www.human-phenotype-ontology.org/) is another important international
initiative to support the annotation of genetic disorders and we are presently collaborating
with HPO in order to further enrich both coding systems.
We are convinced that the approach we adopted – of code development driven by
particular clinical and epidemiological needs, and support for that development from
115
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
experts working in the clinical and medical fields of interest – can contribute to the quality
of annotation for rare diseases, and thus to healthcare for patients with these diseases.
aCknoWLEDGmEntS
This work was supported by CINEAS (Dutch national disease code development and
distribution center for the clinical genetics community) and DDRMD (Dutch Diagnosis
& Registration of Metabolic Diseases). DDRMD was initially funded by Metakids and later
by Top Institute Pharma, Leiden, the Netherlands, as part of projects T6-208 (2008-2011)
and T6-505 (2012-2013). The public funding organizations were not involved in the
design or conduct of the study reported in this article, nor in the collection, analysis,
and interpretation of the data or preparation, review, or approval of the manuscript.
The corporate sponsors of the DDRMD only reviewed the manuscript for intellectual
property issues.
We thank Jan Smeitink, Franc Jan van Spronsen, Annet M. Bosch, Maaike de Vries,
Monique Williams, Margot F. Mulder and Jolanda Huizer for their contributions, and
Jackie Senior for editing the manuscript.
116
6
A n
ew c
od
ing
system
for
metA
bo
lic d
isor
der
s dem
on
strA
tes gA
ps
REfEREnCES1. Official Journal of the European Union. Council recommendation of 8 June 2009 on an
action in the field of rare disease (). 2. Cornet R, de Keizer N. 2008. Forty years of SNOMED: A literature review. BMC Med Inform
Decis Mak 8, Suppl 1:S2. 3. European Medicines Agency (EMA). Orphan drugs and rare diseases at a glance. http://
www.ema.europa.eu/docs/en_GB/document_library/Other/2010/01/WC500069805.pdf accessed October 20, 2011. Document ref. EMEA/290072/2007.
4. European Parliament. Regulation (EC) No. 141/2000 of the European Parliament and the Council of 16 December 1999 on Orphan Medicinal Products. Official Journal of the European Union. http://eur-lex.europa.eu.proxy-ub.rug.nl/LexUriServ/LexUriServ.do?uri=OJ: L:2000:018:0001:0005:En:PDF, January 22, 2000.
5. Food and Drug Administration (USA). Orphan Drug Act. Available at: Http://www.fda.gov/RegulatoryInformation/Legislation/FederalFoodDrugandCosmeticActFDCAct/SignificantAmendmentstotheFDCAct/OrphanDrugAct/default.htm accessed October 20, 2011.
6. Jones S, James E, Prasad S. 2011. Disease registries and outcomes research in children: focus on lysosomal storage disorders. Paediatr Drugs 13:33-47.
7. Rath A, Olry A, Dhombres F, Brandt MM, Urbero B, Ayme S. 2012. Representation of rare diseases in health information systems: The Orphanet approach to serve a wide range of end users. Hum Mutat 33:803-808.
8. Richesson R, Vehik K. 2010. Patient registries: Utility, validity and inference. Adv Exp Med Biol 686:87-104.
9. Sanderson S, Green A, Preece MA, Burton H. 2006. The incidence of inherited metabolic disorders in the West Midlands, UK. Arch Dis Child 91:896-899.
10. Scriver CR, Beaudet AL, Sly WS, Valle MD. 2001. The Metabolic and Molecular Bases of Inherited Disease. New York: McGraw-Hill, Medical Publishing Division. page xliiii.
11. Talele SS, Xu K, Pariser AR, Braun MM, Farag-El-Massah S, Phillips MI, Thompson BH, Cote TR. 2010. Therapies for inborn errors of metabolism: What has the orphan drug act delivered? Pediatrics 126:101-106.
12. Zwamborn-Hanssen AM, Bijlsma JB, Hennekam EF, Lindhout D, Beemer FA, Bakker E, Kleijer WJ, de France HF, de Die-Smulders CE, Duran M, van Gennip AH, van Mens JT, Pearson PL, Mantel G, Verhage RE, Geraedts JP. 1997. The Dutch uniform multicenter registration system for genetic disorders and malformation syndromes. Am J Med Genet 70:444-447.
CHaPtERSORTA: a System for Ontology-based Recoding and Technical
Annotation of biomedical phenotype data
Chao Pang, Annet Sollie, Anna Sijtsma, Dennis Hendriksen, Bart Charbon, Mark de Haan, Tommy de Boer, Fleur Kelpin, Jonathan Jetten, K. Joeri van der Velde,
Nynke Smidt, Rolf Sijmons, Hans Hillege, Morris A. Swertz
7
Published in Database, the journal of biological databases and curation, September 2015 as: Pang C, Sollie A, Sijtsma A, Hendriksen D, Charbon B, de Haan M, de Boer T, Kelpin F, Jetten J, van der Velde JK, Smidt N, Sijmons R, Hillege H, Swertz MA. SORTA: a system for ontology-based re-coding and technical annotation of biomedical phenotype data. Database (Oxford). 2015 Sep 18;2015. pii: bav089. doi: 10.1093/database/bav089.
121
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
abStRaCt
There is an urgent need to standardize the semantics of biomedical data values, such
as phenotypes, to enable comparative and integrative analyses. However, it is unlikely
that all studies will use the same data collection protocols. As a result, retrospective
standardization is often required, which involves matching of original (unstructured
or locally coded) data to widely used coding or ontology systems such as SNOMED
CT (clinical terms), ICD-10 (International Classification of Disease), and HPO (Human
Phenotype Ontology). This data curation process is usually a time-consuming process
performed by a human expert.
To help mechanize this process, we have developed SORTA, a computer-aided system
for rapidly encoding free text or locally coded values to a formal coding system or
ontology. SORTA matches original data values (uploaded in semicolon delimited format)
to a target coding system (uploaded in Excel spreadsheet, OWL ontology web language or
OBO open biomedical ontologies format). It then semi-automatically shortlists candidate
codes for each data value using Lucene and n-gram based matching algorithms, and can
also learn from matches chosen by human experts.
We evaluated SORTA’s applicability in two use cases. For the LifeLines biobank, we
used SORTA to recode 90,000 free text values (including 5,211 unique values) about
physical exercise to MET (Metabolic Equivalent of Task) codes. For the CINEAS clinical
symptom coding system, we used SORTA to map to HPO, enriching HPO when necessary
(315 terms matched so far). Out of the shortlists at rank 1, we found a precision/recall
of 0.97/0.98 in LifeLines and of 0.58/0.45 in CINEAS. More importantly, users found the
tool both a major time saver and a quality improvement because SORTA reduced the
chances of human mistakes. Thus, SORTA can dramatically ease data (re)coding tasks
and we believe it will prove useful for many more projects.
Database URL: http://molgenis.org/sorta or as an open source download from
http://www.molgenis.org/wiki/SORTA
122
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
IntRoDuCtIon
Biobank and translational research can benefit from the massive amounts of phenotype
data now being collected by hospitals and via questionnaires. However, heterogeneity
between data sets remains a barrier to integrated analysis. For the BioSHaRE(1) biobank
data integration project, we previously developed BiobankConnect(2), a tool to overcome
heterogeneity in data structure by mapping data elements from the source database
onto a target scheme. Here, we address the need to overcome heterogeneity of data
contents by coding and/or recoding data values, i.e. mapping free text descriptions or
locally coded data values onto a widely used coding system. In this ‘knowledge-based
data access’, data is collected and stored according to local requirements while information
extracted from the data is revealed using standard representations, such as ontologies,
to provide a unified view(3).
The (re)coding process is essential for the performance of three different kinds of
functions:
1. Search and query. The data collected in a research and/or clinical setting can be
described in numerous ways with the same concept often associated with multiple
synonyms, making it difficult to query distributed database systems in a federated
fashion. For example, using standard terminologies, the occurrence of ‘cancer’ written
in different languages can be easily mapped between databases if they have been
annotated with same ontology term.
2. Reasoning with data. Ontologies are the formal representation of knowledge and
all of the concepts in an ontology have been related to each other using different
relationships, e.g. ‘A is a subclass of B’. Based on these relationships, the computer can be
programmed to reason and infer the knowledge(4). For example, when querying cancer
patients’ records from hospitals, those annotated with ‘Melanoma’ will be retrieved
because ‘Melanoma’ is specifically defined as a descendant of ‘Cancer’ in the ontology.
3. Exchange or pooling of data across systems. Ontologies can also be used to
describe the information model, such as the MGED (Microarray Gene Expression Data)
ontology describing microarray experiments or hospital information coded using the
ICD-10 (International Classification of Diseases) coding system, so that the data can
easily flow across systems that use the same model(4).
The data (re)coding task is essentially a matching problem between a list of free text
data values to a coding system, or from one coding system to another. Unfortunately,
as far as we know, there are only a few software tools available that can assist in this
(re)coding process. Researchers still mostly have to evaluate and recode each data value
by hand, matching values to concepts from the terminology to find the most suitable
candidates. Not surprisingly, this is a time-consuming and error-prone task. Based on
123
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
our previous success in BioSHaRE, we were inspired to approach this problem using
ontology matching and lexical matching(2). We evaluated how these techniques can aid
and speed-up the (re)coding process in the context of phenotypic data. In particular, we
used our newly developed system, SORTA, to recode 5,210 unique entries for ‘physical
exercise’ in the LifeLines biobank(5) and 315 unique entries for ‘physical symptoms’
(including terms that are similar, but not the same) in the Dutch CINEAS (www.cineas.org)
(6) and HPO (Human Phenotype Ontology) coding systems for metabolic diseases.
RequirementsSeveral iterations of SORTA-user interviews resulted in the identification of the following
user requirements:
1. Comparable similarity scores, e.g. scores expressed as a percentage, so users can
easily assess how close a suggested match is to their data, and decide on a cut-off
to automatically accept matches.
2. Support import of commonly used ontology formats (OWL/OBO) for specialists and
Excel spread sheets for less technical users.
3. Fast matching algorithm to accommodate large input datasets and coding systems.
4. Online availability so users can recode/code data directly and share with colleagues
without need to download/install the tool.
5. Maximize the sensitivity to find candidate matches and let users decide on which one
of them is the ‘best’ match.
6. Enable complex matching in which not only a text string is provided but also associated
attributes such as labels, synonyms and annotations, e.g. [label: Hearing impairment,
synonyms:(Deafness, Hearing defect)].
ApproachesTwo types of matching approaches have been reported in the literature: lexical matching
and semantic matching. Lexical matching is a process that measures the similarity
between two strings(7). Edit-distance(8), n-gram(9) and Levenshtein distance(10) are
examples of string-based algorithms that focus on string constituents and are often
useful for short strings, but they do not scale up for matching large numbers of entity
pairs. Token-based techniques focus on word constituents by treating each string as a
bag of words. An example of these techniques is the vector space model algorithm(11), in
which each word is represented as a dimension in space and a cosine function is used to
calculate the similarity between two string vectors. Lexical matching is usually implemented
in combination with a normalization procedure such as lowering case, removing stop
words (e.g. ‘and’, ‘or’, ‘the’) and defining word stems (e.g. ‘smoking’ à ‘smoke’).
Semantic matching techniques search for correspondences based not only on the
textual information associated to a concept (e.g. description) but also on the associative
relationships between concepts (e.g. subclass, ‘is-a’)(7). In these techniques, for example,
124
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
‘melanoma’ is a good partial match for the concept called ‘cancer’. Because our goal is to
find the most likely concepts matching data values based on their similarity in description,
lexical-based approaches seem most suitable.
One of the challenges in the (re)coding task is the vast number of data values that need
to be compared, which means that the matcher has to find correspondences between the
Cartesian product of the original data values and the codes in the desired coding system.
High-throughput algorithms are needed to address this challenge and two methods have
been developed to deal with the matching problem on a large scale. The Early Pruning
Matching Technique(12) reduces search space by omitting irrelevant concepts from the
matching process, e.g. the ontology concept (label:hearing impairment, synonyms[deafness,
hearing defect, congenital hearing loss]) that does not contain any words from the search
query ‘protruding eye ball’ are eliminated. The Parallel Matching Technique(12) divides the
whole matching task into small jobs and the matcher then runs them in parallel, e.g. 100 data
values are divided into 10 partitions that are matched in parallel with ontologies.
Existing toolsWe found several existing tools that offered partial solutions, see table 1.
table 1. Comparison of existing tools with SORTA. ZOOMA and BioPortal Annotator were the closest to our needs.
SoRtabioPortal annotator Zooma Shiva
agreement maker Logmap Peregrine
Comparable similarity score
Y N N N Y Y N
Import code system in ontology format
Y Y Y Y Y Y Y
Import code system in excel format
Y N N N N N N
Uses lexical index to improve performance
Y Y Y N N Y Y
Code/Recode data directly in the tool
Y N N N Y N N
Tool available as online service
Y Y Y N/A N/A N/A N
Support partial matches
Y N N Y Y Y N
Match complex data values
Y N N Y Y Y N
Learns from curated dataset
Y N Y N N N N
Y represents Yes; N represents No; N/A represents unknown
125
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
Mathur and Joshi (13) described an ontology matcher, Shiva, that incorporates four
string-matching algorithms (Levenshtein distance, Q-grams, Smith Waterman and Jaccard),
any of which could be selected by users for particular matching tasks. They used general
resources like WordNet and Online Dictionary to expand the semantics of the entities being
matched. Cruz (14) described a matcher, Agreement Maker, in which lexical and semantic
matchers were applied to ontologies in a sequential order and the results were combined
to obtain the final matches. At the lexical matching stage, Cruz (14) applied several
different kinds of matchers, string-based matches (e.g. edit distance and Jar-Winkler) and
an internally revised token-based matcher, then combined the similarity metrics from these
multiple matchers. Moreover the philosophy behind this tool is that users can help make
better matches in a semi-automatic fashion that are not possible in automatic matching
(14). Jiménez-Ruiz and Cuenca Gra (15) described an approach where: I) they used lexical
matching to compute an initial set of matches; II) based on these initial matches, they took
advantage of semantic reasoning methods to discover more matches in the class hierarchy,
and III) they used indexing technology to increase the efficiency of computing the match
correspondences between ontologies. Peregrine (16) is an indexing engine or tagger that
recognizes concepts within human readable text, and if terms match multiple concepts
it tries to disambiguate BioPortal(17), the leading search portal for ontologies, provides
the BioPortal Annotator that allows users to annotate a list of terms with pre-selected
ontologies. While it was useful for our use cases, it was limited because it only retrieves
perfect matches and terms with slightly different spellings cannot be easily matched
(e.g. ‘hearing impaired’ vs. ‘hearing impairment’)(18). In addition, BioPortal Annotator’s
500-word limit reduces its practical use when annotating thousands of data values.
Finally, ZOOMA(19) enables semi-automatic annotation of biological data with selected
ontologies and was closest to our needs. ZOOMA classifies matches as ‘Automatic’ or
‘Curation required’ based on whether or not there is manually curated knowledge that
supports the suggested matches. ZOOMA does not meet our requirements in that it
does not provide similarity scores for the matches, does not prioritize recall over precision
(i.e. ZOOMA matches are too strict for our needs), and does not handle partial/complex
matches. For example, in ZOOMA, the OMIM (Online Mendelian Inheritance in Man)
term ‘Angular Cheilitis’ could not be partially matched to the HPO term ‘Cheilitis’ and
‘Extra-Adrenal Pheochromocytoma’ could not be matched to the HPO term ‘Extraadrenal
pheochromocytoma’ because of the hyphen character.
mEtHoD
Based on our evaluation of existing tools, we decided to combine a token-based algorithm,
Lucene(20), with an n-gram-based algorithm. Lucene is a high-performance search engine
that works similarly to the Early Pruning Matching Technique. Lucene only retrieves
126
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
concepts relevant to the query, which greatly improves the speed of matching. This
enables us to only recall suitable codes for each value and sort them based on their
match. However, the Lucene matching scores are not comparable across different queries
making it unsuitable for human evaluation. Therefore, we added an n-gram-based
algorithm as a second matcher, which allows us to standardize the similarity scores as
percentages (0-100%) to help users understand the quality of the match and to enable
a uniform cut-off value.
We implemented the following three steps. First, coding systems or ontologies are
uploaded and indexed in Lucene to enable fast searches (once for each ontology). Second,
users create their own coding/recoding project by uploading a list of data values. What
users get back is a shortlist of matching concepts for each value that has been retrieved
from the selected coding system based on their lexical relevance. In addition, the concepts
retrieved are matched with the same data values using the second matcher, the n-gram-
based algorithm, to normalize the similarity scores to values from 0-100%. Finally, users
apply a %-similarity-cut-off to automatically accept matches and/or manually curates
the remaining codes that are assigned to the source values. Finally, users download the
result for use in their own research. An overview of the strategy is shown in Figure 1.
We provide a detailed summary below.
Users upload coding sources such as ontologies or terminology lists to establish the
knowledge base. Ontologies are the most frequently used source for matching data
values, but some of the standard terminology systems are not yet available in ontology
formats. Therefore, we allow users to not only upload ontologies in OWL and OBO
figure 1. Sorta Overview.
METHOD Based on our evaluation of existing tools, we decided to combine a token-based algorithm,
Lucene(20), with an n-gram-based algorithm. Lucene is a high-performance search engine
that works similarly to the Early Pruning Matching Technique. Lucene only retrieves
concepts relevant to the query, which greatly improves the speed of matching. This enables us
to only recall suitable codes for each value and sort them based on their match. However, the
Lucene matching scores are not comparable across different queries making it unsuitable for
human evaluation. Therefore, we added an n-gram-based algorithm as a second matcher,
which allows us to standardize the similarity scores as percentages (0-100%) to help users
understand the quality of the match and to enable a uniform cut-off value.
We implemented the following three steps. First, coding systems or ontologies are uploaded
and indexed in Lucene to enable fast searches (once for each ontology). Second, users create
their own coding/recoding project by uploading a list of data values. What users get back is a
shortlist of matching concepts for each value that has been retrieved from the selected coding
system based on their lexical relevance. In addition, the concepts retrieved are matched with
the same data values using the second matcher, the n-gram-based algorithm, to normalize the
similarity scores to values from 0-100%. Finally, users apply a %-similarity-cut-off to
automatically accept matches and/or manually curates the remaining codes that are assigned
to the source values. Finally, users download the result for use in their own research. An
overview of the strategy is shown in Figure 1. Figure 1 – Sorta Overview.
127
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
formats, but also import a ‘raw knowledge base’ stored in a simple Excel format which
includes system ID, concept ID, and label (see table 2).
This example shows an Excel file with MET (Metabolic Equivalent of Task), a system
developed to standardize physical activity, in which each concept ID includes a list of
different sports representing specific amounts of energy consumption.
table 2. Example of how to upload a coding system and a coding/recoding target.
Concept ID Concept Label System ID
02060 cardio training MET
02020 bodypump MET
18310 swimming MET
15430 kung fu MET
15350 hockey MET
12150 running MET
The uploaded data is then indexed and stored locally to enable rapid matching.
To match data values efficiently, we used the Lucene search index with the default
snowball stemmer and a standard filter for stemming and removing stop words. A code/
ontology concept is evaluated as being a relevant match for the data value when it or its
corresponding synonyms (if available) contain at least one word from the data value. The
assumption in this strategy is that the more words a concept’s label or synonyms contain,
the more relevant Lucene will rank it, and therefore the top concepts on the list are most
likely to be the correct match. However, the snowball stemmer could not stem some of
the English words properly, e.g. the stemmed results for ‘placenta’ and ‘placental’ were
‘placenta’ and ‘placent’, respectively. To solve this problem, we enabled fuzzy matching
with 80% similarity and this allowed us to maximize the number of relevant concepts
retrieved by Lucene.
Lucene also provides matching scores that are calculated using a cosine similarity
between two weighted vectors (21), which takes the information content of words into
account, e.g. rarer words are weighted more than common ones. However, after our
first user evaluations we decided not to show Lucene scores to users for two reasons.
First, Lucene calculates similarity scores for any indexed document as long as it contains
at least one word from the query. Documents that have more words that match the
query, or contain words that are relatively rare, will get a higher score. Secondly, the
matching results produced by different queries are not comparable because the scales
are different (22) making it impossible to determine the ‘best’ cut-off value above which
the suggested matches can be assumed to be correct.
128
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
We therefore decided to provide an additional similarity score that ranges from
0-100% by using an n-gram calculation between the data value and the relevant
concepts retrieved by Lucene. In this n-gram-based algorithm, the similarity score
is calculated for two strings each time. The input string is lowercased and split by
whitespace to create a list of words, which are then stemmed by the default snowball
stemmer. For each of the stemmed words, it is appended with ‘̂ ’ at the beginning and
‘$’ at the end, from which the bigram tokens are generated, e.g. ^smoke$ à [^s, sm,
mo, ok, ke, e$]. All the bigram tokens are pushed to a list for the corresponding input
string with duplicated tokens allowed. The idea is that the more similar two strings are,
the more bigram tokens they can share. The similarity score is the product of number
of shared bigram tokens divided by the sum of total number of bigram tokens of two
input strings as follows,
respectively. To solve this problem, we enabled fuzzy matching with 80% similarity and this
allowed us to maximize the number of relevant concepts retrieved by Lucene.
Lucene also provides matching scores that are calculated using a cosine similarity between
two weighted vectors (21), which takes the information content of words into account, e.g.
rarer words are weighted more than common ones. However, after our first user evaluations
we decided not to show Lucene scores to users for two reasons. First, Lucene calculates
similarity scores for any indexed document as long as it contains at least one word from the
query. Documents that have more words that match the query, or contain words that are
relatively rare, will get a higher score. Secondly, the matching results produced by different
queries are not comparable because the scales are different (22) making it impossible to
determine the ‘best’ cut-off value above which the suggested matches can be assumed to be
correct.
We therefore decided to provide an additional similarity score that ranges from 0-100% by
using an n-gram calculation between the data value and the relevant concepts retrieved by
Lucene. In this n-gram-based algorithm, the similarity score is calculated for two strings each
time. The input string is lowercased and split by whitespace to create a list of words, which
are then stemmed by the default snowball stemmer. For each of the stemmed words, it is
appended with ‘^’ at the beginning and ‘$’ at the end, from which the bigram tokens are
generated, e.g. ^smoke$ [^s, sm, mo, ok, ke, e$]. All the bigram tokens are pushed to a list
for the corresponding input string with duplicated tokens allowed. The idea is that the more
similar two strings are, the more bigram tokens they can share. The similarity score is the
product of number of shared bigram tokens divided by the sum of total number of bigram
tokens of two input strings as follows,
Because we were only interested in the constituents of the strings being compared, the order
of the words in strings does not change the score. We also considered only using the n-gram
calculation, but that would require calculation of all possible pairwise comparisons between
all data values and codes, which would greatly slow down the process.
Ultimately both algorithms were combined because Lucene is very efficient in retrieving
relevant matches while our users preferred n-gram scores because they are easier to compare.
Combining Lucene with the n-gram-based algorithm is an optimal solution in which the
advantages of both methods complement each other while efficiency, accuracy and
comparability of scores are preserved.
Because we were only interested in the constituents of the strings being compared,
the order of the words in strings does not change the score. We also considered only
using the n-gram calculation, but that would require calculation of all possible pairwise
comparisons between all data values and codes, which would greatly slow down
the process.
Ultimately both algorithms were combined because Lucene is very efficient in retrieving
relevant matches while our users preferred n-gram scores because they are easier to
compare. Combining Lucene with the n-gram-based algorithm is an optimal solution in
which the advantages of both methods complement each other while efficiency, accuracy
and comparability of scores are preserved.
To code the data values, the data can be uploaded as a simple comma separate value
file or copy/pasted into the text area directly in SORTA. The uploaded data is usually a
list of simple string values, however in some cases it also can be complex data values
containing information other than a simple label.
For these cases, SORTA allows inclusion of descriptive information such as synonyms
and external database identifiers to improve the quality of the matched results shown
in table 3.
At minimum, one column of values should be provided: the first column with the
header ‘Name’. Additional optional columns that start with ‘Synonym_’ can contain the
synonyms for input values. Other optional column headers can contain other identifiers,
e.g. in this example OMIM.
129
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
table 3. Example of how to upload data values and coding/recoding source).
name (required) Synonym_1(optional) omIm (optional)
2,4-dienoyl-CoA reductase deficiency DER deficiency 222745
3-methylcrotonyl-CoA carboxylase deficiency 3MCC 210200
Acid sphingomyelinase deficiency ASM 607608
For each of the data values, a suggested list of matching concepts is retrieved and
sorted based on similarity. Users can then check the list from the top downwards and
decide which of the concepts should be selected as the final match. However, if the first
concept on the list is associated with a high similarity score, users can also choose not
to look at the list because they can confidently assume that a good match has been
found for that data value. By default, 90% similarity is the cut-off above which the first
concept on the retrieved list is automatically picked as the match for the data value and
stored in the system. Below 90% similarity, users are required to manually check the list
to choose the final match. The cut-off value can be changed according to the needs of
the project, e.g. a low cut-off of 70% can be used if the data value was collected using
free text because typos are inevitably introduced during data collection.
RESuLtS
We evaluated SORTA in various projects. Here we report two representative matching
scenarios where the original data values were either free text (case 1) or already coded, but
using a local coding system (case 2). In addition, as a benchmark, we generated matches
between HPO, NCIT (National Cancer Institute Thesaurus), OMIM (Online Mendelian
Inheritance in Man) and DO (Disease Ontology) and compared the matches with existing
cross references between these two (case 3)
Case 1: Coding unstructured data in the LifeLines biobank
Background
LifeLines is a large biobank and cohort study started by the University Medical Centre
Groningen, the Netherlands. Since 2006, it has recruited 167,729 participants from the
northern region of the Netherlands(5). LifeLines is involved in the EU BioSHaRE consortium
and one of the joint data analyses being conducted by BioSHaRE is the ‘Healthy Obese
Project’ (HOP) that examines why some obviously obese individuals are still metabolically
healthy(23). One of the variables needed for the HOP analysis is physical activity but,
unfortunately, this information was collected using a Dutch questionnaire containing free
text fields for types of sports. Researchers thus needed to match these to an existing
130
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
coding system: the Ainsworth compendium of physical activities(24). In this compendium
each code matches a metabolic equivalent task (MET) intensity level corresponding to
the energy cost of that physical activity and defined as the ratio of the metabolic rate for
performing that activity to the resting metabolic rate. One MET is equal to the metabolic
rate when a person is quietly sitting and can be equivalently expressed as:
using a local coding system (case 2). In addition, as a benchmark, we generated matches
between HPO, NCIT (National Cancer Institute Thesaurus), OMIM (Online Mendelian
Inheritance in Man) and DO (Disease Ontology) and compared the matches with existing
cross references between these two (case 3)
Case 1: Coding unstructured data in the LifeLines biobank
Background
LifeLines is a large biobank and cohort study started by the University Medical Centre
Groningen, the Netherlands. Since 2006, it has recruited 167,729 participants from the
northern region of the Netherlands(5). LifeLines is involved in the EU BioSHaRE consortium
and one of the joint data analyses being conducted by BioSHaRE is the ‘Healthy Obese
Project’ (HOP) that examines why some obviously obese individuals are still metabolically
healthy(23). One of the variables needed for the HOP analysis is physical activity but,
unfortunately, this information was collected using a Dutch questionnaire containing free text
fields for types of sports. Researchers thus needed to match these to an existing coding
system: the Ainsworth compendium of physical activities(24). In this compendium each code
matches a metabolic equivalent task (MET) intensity level corresponding to the energy cost of
that physical activity and defined as the ratio of the metabolic rate for performing that activity
to the resting metabolic rate. One MET is equal to the metabolic rate when a person is quietly
sitting and can be equivalently expressed as:
1 ��� ≡ 1 ������ � � ≡ 4.184 ��
�� � �
A list of 800 codes has been created to represent all kinds of daily activities with their
corresponding energy consumption(24). Code 1015, for example, represents ‘general
bicycling’ with a MET value of 7.5. The process of matching the physical activities of
LifeLines data with codes is referred to as coding.
Challenges and motivation
There were two challenges in this task. First, the physical activities were collected in Dutch
and therefore only researchers with a good level of Dutch could perform the coding task.
Second, there were data for more than 90,000 participants and each participant could report
up to four data values related to ‘Sport’ that could be used to calculate the MET value. In
total, there were 80,708 terms (including 5,211 unique terms) that needed to be coded. We
consulted with the researchers and learned that they typically coded data by hand in an Excel
A list of 800 codes has been created to represent all kinds of daily activities with their
corresponding energy consumption(24). Code 1015, for example, represents ‘general
bicycling’ with a MET value of 7.5. The process of matching the physical activities of
LifeLines data with codes is referred to as coding.
Challenges and motivation
There were two challenges in this task. First, the physical activities were collected in Dutch
and therefore only researchers with a good level of Dutch could perform the coding
task. Second, there were data for more than 90,000 participants and each participant
could report up to four data values related to ‘Sport’ that could be used to calculate the
MET value. In total, there were 80,708 terms (including 5,211 unique terms) that needed
to be coded. We consulted with the researchers and learned that they typically coded
data by hand in an Excel sheet or by syntax in SPSS, and for each entry they needed to
cross-check the coding table and look up the proper code. While this approach is feasible
on a small scale (<10,000 participants), it became clear it would be too much work to
manually code such a massive amount of data. Hence, we used our SORTA coding system.
To train SORTA, we reused a list of human-curated matches between physical
activities described in Dutch and the codes that were created for a previous project.
We used this as the basis to semi-automatically match the new data from LifeLines. An
example of the curated matches is shown in table 2 and the complete list can be found
at Supplementary material: Lifelines_mEt_mappings.xlsx. Moreover, we have
enhanced SORTA with an upload function to support multiple ‘Sport’-related columns
in one harmonization project. This can be done as long as the column headers comply
with the standard naming scheme, where the first column header is ‘Identifier’ and other
column headers start with string ‘Sport_’, e.g. ‘Sport_1’ and ‘Sport_2’. figure 2 shows
an example of manually coding the physical activity ‘ZWEMMEN’ (Swimming) with MET
codes, in which a shortlist of candidates were retrieved by SORTA and the first item of
the list selected as the true match.
Each time the manual curation process produced a new match, this new knowledge
could be added to the knowledge base to be applied to all future data values. This is an
optional action because data values (especially those filled in by participants of the study)
sometimes contain spelling errors that should not be added to the knowledge base.
131
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
Evaluation
With the assistance of SORTA, all of the data values have been coded by the researcher
who is responsible for releasing data about physical activity in the LifeLines project.
The coding result containing a list of matches was used as the gold standard for the
following analysis, in which we evaluated two main questions: I) How far could the
previous coding round improve the new matching results? II) What is the best cut-off
value above which the codes selected by SORTA can be confidently assumed to be
correct matches to a value?
SORTA’s goal is to shortlist good codes for the data values so we first evaluated the
rank of the correct manual matches because the higher they rank, the less manual work
the users need to perform. Our user evaluations suggested that as long as the correct
matches were captured in the top 10 codes, the researchers considered the tool useful.
Otherwise, based on their experience, users changed the query in the tool to update
the matching results.
Re-use of manually curated data from the previous coding round resulted in an
improvement in SORTA’s performance with recall/precision at rank 1st increasing from
0.59/0.65 to 0.97/0.98 and at rank 10th from 0.79/0.14 to 0.98/0.11 (see figure 3 and
table 4).
sheet or by syntax in SPSS, and for each entry they needed to cross-check the coding table
and look up the proper code. While this approach is feasible on a small scale (<10,000
participants), it became clear it would be too much work to manually code such a massive
amount of data. Hence, we used our SORTA coding system.
To train SORTA, we reused a list of human-curated matches between physical activities
described in Dutch and the codes that were created for a previous project. We used this as the
basis to semi-automatically match the new data from LifeLines. An example of the curated
matches is shown in Table 2 and the complete list can be found at Supplementary material: Lifelines_MET_mappings.xlsx. Moreover, we have enhanced SORTA with an upload
function to support multiple ‘Sport’-related columns in one harmonization project. This can
be done as long as the column headers comply with the standard naming scheme, where the
first column header is ‘Identifier’ and other column headers start with string ‘Sport_’, e.g.
‘Sport_1’ and ‘Sport_2’. Figure 2 shows an example of manually coding the physical activity
‘ZWEMMEN’ (Swimming) with MET codes, in which a shortlist of candidates were
retrieved by SORTA and the first item of the list selected as the true match.
Figure 2 Example of coding a physical activity
Each time the manual curation process produced a new match, this new knowledge could be
added to the knowledge base to be applied to all future data values. This is an optional action
because data values (especially those filled in by participants of the study) sometimes contain
spelling errors that should not be added to the knowledge base.
figure 2. Example of coding a physical activity
132
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
In total, 90,000 free text values (of which 5,211 were unique) were recoded to physical
exercise using MET coding system. The table shows recall and precision per position in
the SORTA result before coding (using only the MET score descriptions) and after coding
(when a human curator had already processed a large set of SORTA recommendations
by hand).
At the end of the coding task, about 97% of correct matches were captured at rank 1st
with users only needing to look at the first candidate match.
We included use of an n-gram-based algorithm to provide users with an easily
understood metric with which to judge the relevance of the proposed codes on
a scale of 1-100%, based on the n-gram match between value and code (or a
synonym thereof). Supplementary table 1 suggests that, in the LifeLines case,
82% similarity is a good cut-off for automatically accepting the recommended code
because 100% of the matches produced by the system were judged by the human
curator to be correct matches. Because LifeLines data is constantly being updated
(with new participants, and with new questionnaire data from existing participants
every 18 months), it would be really helpful to recalibrate the cut-off value when
the tool is applied anew.
figure 3. Receiver operating characteristic (ROC) curves evaluating performance on LifeLines data.
Evaluation
With the assistance of SORTA, all of the data values have been coded by the researcher who
is responsible for releasing data about physical activity in the LifeLines project. The coding
result containing a list of matches was used as the gold standard for the following analysis, in
which we evaluated two main questions: I) How far could the previous coding round improve
the new matching results? II) What is the best cut-off value above which the codes selected by
SORTA can be confidently assumed to be correct matches to a value?
SORTA’s goal is to shortlist good codes for the data values so we first evaluated the rank of
the correct manual matches because the higher they rank, the less manual work the users need
to perform. Our user evaluations suggested that as long as the correct matches were captured
in the top 10 codes, the researchers considered the tool useful. Otherwise, based on their
experience, users changed the query in the tool to update the matching results.
Re-use of manually curated data from the previous coding round resulted in an improvement
in SORTA’s performance with recall/precision at rank 1st increasing from 0.59/0.65 to
0.97/0.98 and at rank 10th from 0.79/0.14 to 0.98/0.11 (see Figure 3 and Table 4). Figure 3: Receiver operating characteristic (ROC) curves evaluating performance on LifeLines data.
Table 4 Precision and recall for the LifeLines case study.
In total, 90,000 free text values (of which 5,211 were unique) were recoded to physical
exercise using MET coding system. The table shows recall and precision per position in the
SORTA result before coding (using only the MET score descriptions) and after coding (when
a human curator had already processed a large set of SORTA recommendations by hand).
133
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
table 4. Precision and recall for the LifeLines case study.
Rank cut-off
before coding after coding
Recall Precision f-measure Recall Precision f-measure
1 0.59 0.65 0.62 0.97 0.98 0.97
2 0.66 0.39 0.49 0.97 0.50 0.66
3 0.71 0.29 0.41 0.97 0.34 0.50
4 0.74 0.24 0.36 0.97 0.26 0.41
5 0.76 0.21 0.33 0.97 0.21 0.35
6 0.77 0.19 0.30 0.97 0.18 0.30
7 0.78 0.17 0.28 0.97 0.15 0.26
8 0.78 0.16 0.27 0.98 0.14 0.25
9 0.78 0.14 0.24 0.98 0.12 0.21
10 0.79 0.14 0.24 0.98 0.11 0.20
11 0.79 0.13 0.22 0.98 0.10 0.18
12 0.79 0.12 0.21 0.98 0.09 0.16
13 0.79 0.12 0.21 0.98 0.09 0.16
14 0.79 0.12 0.21 0.98 0.08 0.15
15 0.79 0.11 0.19 0.98 0.08 0.15
16 0.79 0.11 0.19 0.98 0.07 0.13
17 0.79 0.11 0.19 0.98 0.07 0.13
18 0.80 0.11 0.19 0.98 0.06 0.11
19 0.80 0.10 0.18 0.98 0.06 0.11
20 0.80 0.10 0.18 0.98 0.06 0.11
30 0.80 0.10 0.18 0.98 0.04 0.08
50 0.80 0.09 0.16 0.98 0.03 0.06
Case 2: Recoding from CINEAS coding system to HPO ontology
Background
CINEAS is the Dutch centre for disease code development and its distribution to the
clinical genetics community (www.cineas.org)(6). This centre was initiated by the
eight clinical genetics centres responsible for genetic counselling and diagnostics
in the Netherlands in 1992(25). CINEAS codes are used in daily practice by Dutch
clinical geneticists and genetic counsellors to assign diseases and clinical symptoms
to patients. The 63rd edition of CINEAS now lists more than 5,600 diseases and
more than 2,800 clinical symptoms. The challenge was to match and integrate
134
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
(or recode) the CINEAS clinical symptom list with HPO in order to use one enriched
standardized coding system for future coding of patients’ symptoms and to obtain
interoperability for CINEAS codes already registered in local systems all over the
country. The metabolic diseases obtained from CINEAS disease list, which has
become an independent project called The Dutch Diagnosis Registration Metabolic
Diseases (DDRMD, https://ddrmd.nl/)(25), will be matched with Orphanet ontology
in the future.
Challenge and motivation
The previous strategy of CINEAS curators was to search HPO via BioPortal, however,
tracking possible candidate terms meant making written notes or keeping a digital
registry on the side, tracking methods that are time-consuming, prone to human errors
and demand a lot of switching between tools or screens. Therefore, SORTA was brought
into the project. figure 4 shows an example of a data value ‘external auditory canal
defect’ and a list of HPO ontology terms as candidate matches.
While none of them is a perfect match for the input term, the top three candidates
are the closest matches, but are too specific for the input. Scrutiny by experts revealed
that ‘Abnormality of auditory canal’ could be a good ‘partial’ match because of its
generality. Figure 4: Example of matching the input value ‘external auditory canal defect’ with HPO ontology
terms.
While none of them is a perfect match for the input term, the top three candidates are the
closest matches, but are too specific for the input. Scrutiny by experts revealed that
‘Abnormality of auditory canal’ could be a good ‘partial’ match because of its generality.
Evaluation
In an evaluation study, the first 315 clinical symptoms out of 2,800 were re-coded by a human
expert, in which 246 were matched with HPO terms while 69 could not be matched. In
addition, we performed the same matching task using BioPortal Annotator and ZOOMA
because these existing tools seemed most promising (see Table 5).
Table 5 Comparison of SORTA, BioPortal and ZOOMA. Evaluation based on the CINEAS case study in which 315 clinical symptoms were matched to
Human Phenotype Ontology. The table shows the recall/precision per position in SORTA,
BioPortal Annotator and ZOOMA. N.B. both BioPortal Annotator and ZOOMA have a
limitation that they can only find exact matches and return a maximum of three candidates. SORTA BioPortal ZOOMA
Rank
cut-off Recall
Precisi
on
F-
measurRecall
Precisi
on
F-
measurRecall
Precisi
on
F-
measur
figure 4. Example of matching the input value ‘external auditory canal defect’ with HPO ontology terms.
135
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
Evaluation
In an evaluation study, the first 315 clinical symptoms out of 2,800 were re-coded by
a human expert, in which 246 were matched with HPO terms while 69 could not be
matched. In addition, we performed the same matching task using BioPortal Annotator
and ZOOMA because these existing tools seemed most promising (see table 5). Evaluation based on the CINEAS case study in which 315 clinical symptoms were
matched to Human Phenotype Ontology. The table shows the recall/precision per
position in SORTA, BioPortal Annotator and ZOOMA. N.B. both BioPortal Annotator and
ZOOMA have a limitation that they can only find exact matches and return a maximum
of three candidates.
table 5. Comparison of SORTA, BioPortal and ZOOMA.
Rank cut-off
SoRta bioPortal Zooma
Recall Precision f-measure Recall Precision f-measure Recall Precision f-measure
1 0.58 0.45 0.51 0.34 0.54 0.42 0.17 0.63 0.27
2 0.69 0.27 0.39 0.35 0.44 0.39 0.17 0.60 0.26
3 0.73 0.19 0.30 0.35 0.44 0.39 0.18 0.60 0.28
4 0.76 0.15 0.25 N/A N/A N/A N/A N/A N/A
5 0.78 0.13 0.22 N/A N/A N/A N/A N/A N/A
6 0.81 0.11 0.19 N/A N/A N/A N/A N/A N/A
7 0.81 0.09 0.16 N/A N/A N/A N/A N/A N/A
8 0.83 0.08 0.15 N/A N/A N/A N/A N/A N/A
9 0.83 0.08 0.15 N/A N/A N/A N/A N/A N/A
10 0.85 0.07 0.13 N/A N/A N/A N/A N/A N/A
11 0.85 0.06 0.11 N/A N/A N/A N/A N/A N/A
12 0.85 0.06 0.11 N/A N/A N/A N/A N/A N/A
13 0.86 0.06 0.11 N/A N/A N/A N/A N/A N/A
14 0.86 0.05 0.09 N/A N/A N/A N/A N/A N/A
15 0.87 0.05 0.09 N/A N/A N/A N/A N/A N/A
16 0.87 0.05 0.09 N/A N/A N/A N/A N/A N/A
17 0.87 0.05 0.09 N/A N/A N/A N/A N/A N/A
18 0.88 0.04 0.08 N/A N/A N/A N/A N/A N/A
19 0.88 0.04 0.08 N/A N/A N/A N/A N/A N/A
20 0.88 0.04 0.08 N/A N/A N/A N/A N/A N/A
30 0.89 0.03 0.06 N/A N/A N/A N/A N/A N/A
50 0.92 0.02 0.04 N/A N/A N/A N/A N/A N/A
N/A not applicable
136
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
We further investigated which cut-off value can be confidently used to assume that
the automatic matches are correct by calculating precision and recall for all possible
n-gram cut-offs (0-100%). Supplementary table 2 shows 89% to be a good cut-off
value for future CINEAS matching tasks because above this value all of the suggested
matches are correct with 100% precision.
Case 3: Benchmark against existing matches between ontologiesWe downloaded 700 existing matches between HPO and DO concepts, 1148 matches
between HPO and NCIT concepts, and 3631 matches between HPO and OMIM concepts from
BioPortal. We used the matching terms from DO, NCIT and OMIM as the input values and HPO
as the target coding system and generated matches using SORTA, BioPortal Annotator and
ZOOMA. Supplementary table 3 shows that all three tools managed to reproduce most of
the existing ontology matches with SORTA slightly outperforming the other two by retrieving
all of the ontology matches. Scrutiny revealed that SORTA was able to find the complex
matches, where data values and ontology terms consist of multiple words, and some of which
are concatenated, e.g. matching ‘propionic acidemia’ from DO with ‘Propionicacidemia’ from
HPO. We also noticed that beyond the 1st rank, precision in SORTA is lower than the other
two (with the highest precision in ZOOMA). In addition, we investigated what proportion of
data values could be automatically matched at different cut-offs. Supplementary table 4
shows that at similarity score cut-off of 90%, SORTA recalled at least 99.6% of the existing
matches with 100% precision across all three matching experiments.
DISCuSSIon
In RESULTS section, we have evaluated SORTA in three different use cases. It has shown
that SORTA could indeed help human experts in performing the (re)coding tasks in
terms of improving the efficiency and user evaluations of SORTA were very positive,
but there was much debate among co-authors on the combination of Lucene-based
matching with n-gram post-processing. As mentioned in the Method section, Lucene
scores were not really informative for users, but the order in which the matching results
were sorted by Lucene seemed better thanks to the cosine similarity function that takes
information content into account. After applying the n-gram-based algorithm, this order
was sometimes changed. To evaluate this issue we performed the same matching tasks
using Lucene and Lucene + n-gram. In the case of coding LifeLines data, the performances
were quite similar and the inclusion of n-gram did not change the order of the matching
results, see Supplementary material: PrecisionRecallLifeLines.xlsx. However, in the
case of matching HPO terms, there was a large difference in precision and recall as shown
in figure 5 and Supplementary material PrecisionRecallCInEaS.xlsx.
137
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
Lucene alone outperformed the combination of the two algorithms. We hypothesize
that this may be caused by Lucene’s use of word inverse document frequency (IDF)
metrics, which are calculated for each term (t) using the following formula:
Figure 5: Performance comparison for matching HPO terms among three algorithms.
Lucene alone outperformed the combination of the two algorithms. We hypothesize that this
may be caused by Lucene’s use of word inverse document frequency (IDF) metrics, which are
calculated for each term (t) using the following formula:
where docFreq is the number of documents that contain the term.
We checked the IDFs for all the words from input values for the HPO use case and
Supplementary Figure 1 shows the large difference in the information carried by each word.
This suggested that, to improve the usability of the tool, we should allow users to choose
which algorithm they wish to use to sort the matching results, an option that we will add in
the near future. We also explored if we could simply add information content to the n-gram
scoring mechanism to make the ranks consistent by redistributing the contribution of each of
the query words in the n-gram score based on the IDF. For example, using n-gram the
contribution of the word ‘joint’ in the query string ‘hyperextensibility hand joint’ is about
18.5% because ‘joint’ is 5/27 letters. However, if this word is semantically more important,
results matching this word should have a higher score. We therefore adapted the n-gram
algorithm to calculate the IDF for each of the words separately, calculate the average, and
reallocate the scores to the more important words as follows:
where docFreq is the number of documents that contain the term.
We checked the IDFs for all the words from input values for the HPO use case and
Supplementary figure 1 shows the large difference in the information carried by each
word. This suggested that, to improve the usability of the tool, we should allow users
to choose which algorithm they wish to use to sort the matching results, an option that
we will add in the near future. We also explored if we could simply add information
content to the n-gram scoring mechanism to make the ranks consistent by redistributing
the contribution of each of the query words in the n-gram score based on the IDF.
For example, using n-gram the contribution of the word ‘joint’ in the query string
‘hyperextensibility hand joint’ is about 18.5% because ‘joint’ is 5/27 letters. However,
Figure 5: Performance comparison for matching HPO terms among three algorithms.
Lucene alone outperformed the combination of the two algorithms. We hypothesize that this
may be caused by Lucene’s use of word inverse document frequency (IDF) metrics, which are
calculated for each term (t) using the following formula:
where docFreq is the number of documents that contain the term.
We checked the IDFs for all the words from input values for the HPO use case and
Supplementary Figure 1 shows the large difference in the information carried by each word.
This suggested that, to improve the usability of the tool, we should allow users to choose
which algorithm they wish to use to sort the matching results, an option that we will add in
the near future. We also explored if we could simply add information content to the n-gram
scoring mechanism to make the ranks consistent by redistributing the contribution of each of
the query words in the n-gram score based on the IDF. For example, using n-gram the
contribution of the word ‘joint’ in the query string ‘hyperextensibility hand joint’ is about
18.5% because ‘joint’ is 5/27 letters. However, if this word is semantically more important,
results matching this word should have a higher score. We therefore adapted the n-gram
algorithm to calculate the IDF for each of the words separately, calculate the average, and
reallocate the scores to the more important words as follows:
figure 5. Performance comparison for matching HPO terms among three algorithms.
138
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
if this word is semantically more important, results matching this word should have a
higher score. We therefore adapted the n-gram algorithm to calculate the IDF for each
of the words separately, calculate the average, and reallocate the scores to the more
important words as follows:
��������������� = �����ℎ����������������ℎ�������� × ���������� � ��������������
����������
���������������� = �����ℎ����������������ℎ�������� � ���������������
������������������� = �����ℎ�������������������ℎ�������� � ���������������� ×
�����������������∑ ������������������
Common_word is defined as having an IDF that is lower than IDFaverage
Important_words is defined as the IDF that is higher than IDFaverage
This resulted in an improvement of recall compared to naive n-gram scoring at rank 10th from
0.79 to 0.84 (for details see Supplementary material: comparision_ngram_lucene.xlsx),
and the summarized comparison is provided via receiver operating characteristic (ROC) curve
in Figure 5. However, Lucene still outperforms this metric and we speculate that this can be
explained by the fundamental difference between the underlying scoring functions. The n-
gram score is more sensitive to the length of input strings than Lucene and it is quite possible
that two strings do not share any of the words but share similar bigram tokens, especially
when dealing with long strings. Consequently, the n-gram-based algorithm might find more
false positives than Lucene. However, in practice, the number of data values to be
coded/recoded is quite large and the benefit of using an n-gram score cut-off value above
which all the suggested matches are automatically selected outweighs this drawback.
Another issue was whether we could make better use of all the knowledge captured in
ontologies. We noticed in some matching examples that related terms that come from the
same ontological cluster tend to show up together in the matching results. For example,
Figure 4 shows that the input term ‘external auditory canal defect’ is not matched to any of
the top three candidates because they are too specific and hence we have to take the more
general ontology term ‘Auditory canal abnormality’, which is actually ranked 11th, as the
match even though this term is in fact the parent of the three top candidates. This indicates
that if the input value is not matched by any of the candidates with a high similarity score and
the candidates contain clusters of ontology terms, the parent ontology term should probably
be selected as the best match (which is similar to the way human curators make decisions on
such matches). However, translating this knowledge into an automatic adaptation of matching
a score is non-trivial and something we plan to work on in the future.
CONCLUSIONS We developed SORTA as a software system to ease data cleaning and coding/recoding by
automatically shortlisting standard codes for each value using lexical and ontological
This resulted in an improvement of recall compared to naive n-gram scoring at
rank 10th from 0.79 to 0.84 (for details see Supplementary material: comparision_ngram_lucene.xlsx), and the summarized comparison is provided via receiver
operating characteristic (ROC) curve in figure 5. However, Lucene still outperforms
this metric and we speculate that this can be explained by the fundamental difference
between the underlying scoring functions. The n-gram score is more sensitive to the
length of input strings than Lucene and it is quite possible that two strings do not
share any of the words but share similar bigram tokens, especially when dealing with
long strings. Consequently, the n-gram-based algorithm might find more false positives
than Lucene. However, in practice, the number of data values to be coded/recoded is
quite large and the benefit of using an n-gram score cut-off value above which all the
suggested matches are automatically selected outweighs this drawback.
Another issue was whether we could make better use of all the knowledge
captured in ontologies. We noticed in some matching examples that related terms
that come from the same ontological cluster tend to show up together in the
matching results. For example, figure 4 shows that the input term ‘external auditory
canal defect’ is not matched to any of the top three candidates because they are
too specific and hence we have to take the more general ontology term ‘Auditory
canal abnormality’, which is actually ranked 11th, as the match even though this
term is in fact the parent of the three top candidates. This indicates that if the
input value is not matched by any of the candidates with a high similarity score
and the candidates contain clusters of ontology terms, the parent ontology term
should probably be selected as the best match (which is similar to the way human
curators make decisions on such matches). However, translating this knowledge into
an automatic adaptation of matching a score is non-trivial and something we plan
to work on in the future.
139
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
ConCLuSIonS
We developed SORTA as a software system to ease data cleaning and coding/recoding
by automatically shortlisting standard codes for each value using lexical and ontological
matching. User and performance evaluations demonstrated that SORTA provided
significant speed and quality improvements compared to the earlier protocols used by
biomedical researchers to harmonize their data for pooling. With increasing use, we
plan to dynamically update the precision and recall metrics based on all users’ previous
selections so that users can start the matching tasks with confident cut-off values. In
addition, we plan to include additional resources such as WordNet for query expansion to
increase the chance of finding correct matches from ontologies or coding systems. Finally,
we also want to publish mappings as linked data, for example as nanopublications (26)
(http://nanopub.org), so they can be easily reused. SORTA is available as a service running
at http://molgenis.org/sorta. Documentation and source code can be downloaded from
http://www.molgenis.org/wiki/SORTA under open source LGPLv3 license.
aCknoWLEDGEmEntS
This work was supported by the European Union Seventh Framework Programme
(FP7/2007-2013) grant number 261433 (Biobank Standardisation and Harmonisation for
Research Excellence in the European Union - BioSHaRE-EU) and grant number 284209
(BioMedBridges). It was also supported by BBMRI-NL, a research infrastructure financed by
the Netherlands Organization for Scientific Research (NWO), grant number 184.021.007.
We thank Anthony Brookes of Leicester University who contributed the abbreviation
‘SORTA’, and Kate Mc Intyre and Jackie Senior for editing the manuscript.
140
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
REfEREnCES 1. BioShaRE (2011) BioSHaRE project. https://www.bioshare.eu/2. Pang,C., Hendriksen,D., Dijkstra,M. and Velde,K.J. Van Der (2015) BiobankConnect :
software to rapidly connect data elements for pooled analysis across biobanks using ontological and lexical indexing. 10.1136/amiajnl-2013-002577.
3. Poggi,A., Lembo,D., Calvanese,D., De Giacomo,G., Lenzerini,M. and Rosati,R. (2008) Linking data to ontologies. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 4900 LnCS, 133–173.
4. Rubin,D.L., Shah,N.H. and Noy,N.F. (2008) Biomedical ontologies: A functional perspective. Briefings in Bioinformatics, 9, 75–90.
5. Scholtens,S., Smidt,N., Swertz,M.A., Bakker,S.J.L., Dotinga,A., Vonk,J.M., van Dijk,F., van Zon,S.K.R., Wijmenga,C., Wolffenbuttel,B.H.R., et al. (2014) Cohort Profile: LifeLines, a three-generation cohort study and biobank. International Journal of Epidemiology, 10.1093/ije/dyu229.
6. Zwamborn-Hanssen,A.M.N., Bijlsma,J.B., Hennekam,E.F.A.M., Lindhout,D., Beemer,F.A., Bakker,E., Kleijer,W.J., De France,H.F., De Die-Smulders,C.E.M., Duran,M., et al. (1997) The Dutch uniform multicenter registration system for genetic disorders and malformation syndromes. American Journal of Medical Genetics, 70, 444–447.
7. Euzenat,J. and Shvaiko,P. (2013) Ontology Matching Second. Available at: http://www.springer.com/computer/database+management+&+information+retrieval/book/978-3-642-38720-3.
8. Navarro,G. (2001) A guided tour to approximate string matching. ACM Computing Surveys, 33, 31–88.
9. Brown,P.F., DeSouza,P. V., Mercer,R.L., Pietra,V.J. Della and Lai,J.C. (1992) Class-Based n-gram Models of Natural Language. Computational Linguistics, 18, 467–479.
10. Levenshtein,V.I. (1966) Binary codes capable of correcting deletions, insertions, and reversals. Soviet Physics Doklady, 10, 707–710.
11. Salton,G., Wong,A. and Yang,C.S. (1975) A vector space model for automatic indexing. Communications of the ACM, 18, 613–620.
12. Saruladha,K. (2011) A Comparative Analysis of Ontology and Schema Matching Systems. International Journal of Computer Applications, 34, 14–21.
13. Mathur,I. and Joshi,N. (2014) Shiva ++ : An Enhanced Graph based Ontology Matcher. International Journal of Computer Applications, 92, 30–34.
14. Cruz,I.F. (2009) AgreementMaker : Efficient Matching for Large Real-World Schemas and Ontologies. PVLDB, 2, 1586–1589.
15. Jiménez-Ruiz,E. and Cuenca Grau,B. (2011) LogMap: Logic-based and scalable ontology matching. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 7031 LnCS, 273–288.
16. Schuemie,M.J., Jelier,R. and Kors,J. (2007) Peregrine: Lightweight gene name normalization by dictionary lookup. In Proc of the Second BioCreative Challenge Evaluation Workshop.pp. 131–133.
17. Whetzel,P.L., Shah,N.H., Noy,N.F., Dai,B., Dorf,M., Griffith,N., Jonquet,C., Youn,C., Coulet,A., Callendar,C., et al. (2009) BioPortal : Ontologies and Integrated Data Resources at the Click of a Mouse. Nucleic acids research, 37, 170–3.
18. Funk,C., Baumgartner,W., Garcia,B., Roeder,C., Bada,M., Cohen,K.B., Hunter,L.E. and Verspoor,K. (2014) Large-scale biomedical concept recognition: an evaluation of current automatic annotators and their parameters. BMC bioinformatics, 15, 59.
19. Burdett,T., Jupp,S., Malone,J., Williams,E., Keays,M., Parkinson,H., Trust,W. and Campus,G. (2012) Zooma2 - A repository of annotation knowledge and curation API. In Intelligent Systems for Molecular Biology.
141
7
SOR
TA: A
SySTem
fOR
On
TOlO
gy
-bA
Sed R
ecO
din
g A
nd
Tech
nic
Al A
nn
OTA
TiOn
20. The Apache Software Foundation (2006) Apache Lucene. Agenda, 2009.21. Apache Software Foundation (2001) Lucene Similarity Score. https://lucene.apache.org/
core/4_6_0/core/overview-summary.html22. ElasticSearch (2015) ElasticSearch: Lucene’s Practical Scoring Function. https://www.elastic.
co/guide/en/elasticsearch/guide/master/practical-scoring-function.html#query-norm23. Van Vliet-Ostaptchouk,J. V, Nuotio,M.-L., Slagter,S.N., Doiron,D., Fischer,K., Foco,L.,
Gaye,A., Gögele,M., Heier,M., Hiekkalinna,T., et al. (2014) The prevalence of Metabolic Syndrome and metabolically healthy obesity in Europe: a collaborative analysis of ten large cohort studies. BMC Endocrine Disorders, 14, 1–13.
24. Ainsworth,B.E., Haskell,W.L., Leon,A.S., Jacobs,D.R., Montoye,H.J., Sallis,J.F. and Paffenbarger,R.S. (1993) Compendium of physical activities: Classification of energy costs of human physical activities. In Medicine and Science in Sports and Exercise.Vol. 25, pp. 71–80.
25. Sollie,A., Sijmons,R.H., Lindhout,D., van der Ploeg,A.T., Rubio Gozalbo,M.E., Smit,G.P. a, Verheijen,F., Waterham,H.R., van Weely,S., Wijburg,F. a., et al. (2013) A New Coding System for Metabolic Disorders Demonstrates Gaps in the International Disease Classifications ICD-10 and SNOMED-CT, Which Can Be Barriers to Genotype-Phenotype Data Sharing. Human Mutation, 34, 967–973.
26. Sernadela, Pedro and Horst, Eelke and Thompson, Mark and Lopes, Pedro and Roos, Marco and Oliveira,J. (2014) A Nanopublishing Architecture for Biomedical Data. In 8th International Conference on Practical Applications of Computational Biology & Bioinformatics (PACBB 2014).pp. 277–284.
CHaPtERProposed roadmap to stepwise
integration of genetics in family medicine and clinical research
Elisa JF Houwink#,, Annet W Sollie#, Mattijs E Numans, Martina C Cornel# contributed equally (shared first authors)
8
Published in Clinical and translational medicine, february 2013 as: Houwink EJ, Sollie AW, Numans ME, Cornel MC. Proposed roadmap to stepwise integration of genetics in family medicine and clinical research. Clin Transl Med. 2013 Feb 16;2(1):5. doi: 10.1186/2001-1326-2-5.
145
8
Pro
Posed
ro
ad
ma
P to stePw
ise integ
ra
tion
of g
enetic
s in fa
mily
med
icin
e
abStRaCt
We propose a step-by-step roadmap to integrate genetics in the Electronic Patient Record
in Family Medicine and clinical research. This could make urgent operationalization of
readily available genetic knowledge feasible in clinical research and consequently improved
medical care.
Improving genomic literacy by training and education is needed first. The second
step is the improvement of the possibilities to register the family history in such a way
that queries can identify patients at risk. Adding codes to the ICPC chapters “A21
Personal/family history of malignancy” and “A99 Disease carrier not described further”
is proposed. Multidisciplinary guidelines for referral must be unambiguous. Electronical
patient records need possibilities to add (new) family history information, including links
between individuals who are family members. Automatic alerts should help general
practitioners to recognize patients at risk who satisfy referral criteria. We present a familial
breast cancer case with a BRCA1 mutation as an example.
146
8
Pro
Posed
ro
ad
ma
P to stePw
ise integ
ra
tion
of g
enetic
s in fa
mily
med
icin
e
baCkGRounD
Public health benefits of advancements in understanding the human genome are still to
be realized for common chronic diseases such as cardiovascular disease, diabetes mellitus,
and cancer [1]. International attempts to integrate and operationalize such knowledge
into clinical practice are in the early stages, and as a result, many questions surround the
current state of this translation [1-3]. Most physicians lack genetic knowledge and skills
that might be relevant for decision support in daily practice. [4] Family history taking and
family tree drawing need to be introduced. Oversight of clinical utility of genetic testing
should be supported by e-Health facilities to bypass unfamiliarity with facts on genetic
testing. Shortcomings in registration systems and inadequate implementation of genetics
in existing guidelines are reported and result in inability to register genetic information
in Electronic Patient Records. Privacy and risk of discrimination cause concerns when
registration is considered. Consequently, inadequacy to deliver genetic services is reported
in literature [1]. We present a roadmap (Figure 1) to integrate actual genetic knowledge
into the Electronic Patient Record and into clinical research in Family medicine, which
would enable urgent operationalization of readily available knowledge feasible in daily
genetic medical care.
Evidence for necessary changeThe clinical relevance of integrating genetics in clinical practice was demonstrated
for several familial diseases such as colorectal cancer and breast cancer. Dove-Edwin
et al. calculated mortality risk reduction up to 80% by identifying and subsequently
screening individuals with an increased familial colorectal cancer (CRC) risk [5]. Cancer
risk management options through genetic testing for BRCA mutations and subsequent
options for preventive surgery after testing positive can empower women and can also
reduce morbidity and mortality [6]. Currently, a large number of patients in whom
screening would be beneficial, are out of sight or being missed by their physicians [7,8].
Barriers to changeScheuner et al. identified deficiencies in primary care workers’ basic genetic knowledge
and ability to interpret familial patterns [1]. This is in line with our prioritized educational
topics, including knowledge of basic genetic principles, the most common genetic
disorders and family history communication skills [9]. Taylor and Edwards stated primary
care should be encouraged to invest more time and energy in questioning and registering
family history data [10]. However, they also stressed identified barriers such as time
constraints should be encountered. They identified the need to develop strategies to
overcome difficulties as well as strategies to support accurate record keeping in the
electronic medical record (EMR) [10]. Another identified barrier is the presence of
147
8
Pro
Posed
ro
ad
ma
P to stePw
ise integ
ra
tion
of g
enetic
s in fa
mily
med
icin
e
ambiguous referral guidelines to clinical genetics and other medical specialists for patients
with a possible high risk at familial disease, such as cancer [9]. Computerised decision
support might be helpful in familial risk assessment for common cancers (e.g. breast,
ovarian and colon cancers) and would render timely genetic risk assessments and
consequently support referrals more consistent with guidelines. These results support the
implementation of genetics education aimed at enhancing effective referral indications
and options.
A roadmap for translationIn order to be able to truly turn useful genetic discoveries from the laboratory bench
to daily clinical practice, a roadmap is crucial to make urgent translation feasible. First,
advances in the genomic literacy of health care providers are indispensable. Secondly,
innovative and practical ICT tools to apply these newly acquired knowledge and skills
are needed, such as registration of family history and registry alerts supporting this. We
Background Public health benefits of advancements in understanding the human genome are still to be
realized for common chronic diseases such as cardiovascular disease, diabetes mellitus, and
cancer [1]. International attempts to integrate and operationalize such knowledge into clinical
practice are in the early stages, and as a result, many questions surround the current state of
this translation [1-3]. Most physicians lack genetic knowledge and skills that might be
relevant for decision support in daily practice. [4] Family history taking and family tree
drawing need to be introduced. Oversight of clinical utility of genetic testing should be
supported by e-Health facilities to bypass unfamiliarity with facts on genetic testing.
Shortcomings in registration systems and inadequate implementation of genetics in existing
guidelines are reported and result in inability to register genetic information in Electronic
Patient Records. Privacy and risk of discrimination cause concerns when registration is
considered. Consequently, inadequacy to deliver genetic services is reported in literature [1].
We present a roadmap (Figure 1) to integrate actual genetic knowledge into the Electronic
Patient Record and into clinical research in Family medicine, which would enable urgent
operationalization of readily available knowledge feasible in daily genetic medical care. Figure 1 Proposed roadmap to stepwise integration of genetics in the family medicine.
figure 1. Proposed roadmap to stepwise integration of genetics in the family medicine.
148
8
Pro
Posed
ro
ad
ma
P to stePw
ise integ
ra
tion
of g
enetic
s in fa
mily
med
icin
e
propose a step-by-step roadmap (Figure 1) to effectively integrate genetics in daily family
medicine to its full potential:
1. Improve basic knowledge of genetics in clinicians and develop skills and attitude to
obtain and interpret a family history through effective education;
For example, training on oncogenetics for GPs was recently developed and
evaluated in collaboration with The Dutch College of Family Physicians. Also, a
website on genetics targeted to GPs was developed to easily obtain information on,
amongst other topics, genetic diseases, referral guidelines and family history taking
(huisartsengenetica.nl, translated “GP and genetics”). Oncogenetic knowledge,
skills and attitude were effectively transmitted through an accredited online and
live interactive training and could internationally serve as an example for other
common topics (i.e. reproductive medicine, familial coronary heart disease and
diabetes) and possibly other medical specialties provided that they are translated
to its medical systems.
2. Add relevant International Classification of Primary Care (ICPC) codes and other
coding strategies for simple registry of family history and develop and support
coding skills;
In order to identify and track persons and/or families at risk for hereditary diseases
adequate coding is a starting point. We propose to add a number of codes for
simple registration of family history. This will enable and support adequate case-
finding and decision strategies [8]. We propose to add a number of codes in order
to enable simple but structured registry of a family history. In ICPC-2, which is the
most frequently used coding system for GPs in Western countries, these codes
should be included in Chapter A (General and Unspecified), under A21 “Risk
factor for malignancy”. ICPC-2 was developed by the WHO and classifies patient
data and clinical activity in the domains of General/Family Practice and primary
care, taking into account the frequency distribution of problems seen in these
domains. It allows classification of the patient’s reason for encounter (RFE), the
problems/diagnosis managed, interventions, and the ordering of these data in an
episode of care structure. ICPC-2 has a biaxial structure and consists of 17 chapters,
each divided into 7 components (comp.) dealing with symptoms and complaints
(comp. 1), diagnostic, screening and preventive procedures (comp. 2), medication,
treatment and procedures (comp. 3), test results (comp. 4), administrative (comp. 5),
referrals and other reasons for encounter (comp. 6) and diseases (comp. 7). (see
http://www.who.int/classifications/icd/adaptations/icpc2/en/index.html) Mapping
is available between ICPC and ICD-10, which was also developed by the WHO for
broad application in healthcare registries. The codes suggested below should suit
other coding systems such as SNOMED such as SNOMED and should also be added
for other cancer types.
149
8
Pro
Posed
ro
ad
ma
P to stePw
ise integ
ra
tion
of g
enetic
s in fa
mily
med
icin
e
A21 Personal/family history of malignancy (Existing code)
A21.1 One or more 1st degree family member(s) with breast cancer
A21.2 One or more 2nd degree family member(s) with breast cancer
A21.3 One or more family member(s) with bilateral or multifocal breast cancer
A21.4 Breast cancer in the family in one or more men
A99 Disease carrier not described further (Existing code)
A99.1 BRCA-1 mutation carrier
A99.2 BRCA-2 mutation carrier
A99.3 TP53 mutation carrier
A99.99 Carrier of mutation in other specified gene
3. Improve access to up-to-date and unambiguous referral guidelines;
For example, in the Netherlands multiple referral guidelines for hereditary cancers
were developed independently (Oncoline, Foundation for detection of hereditary
tumors (In Dutch STOET), clinical genetics centres in University hospitals and The
Dutch College of Family physicians (NHG)). Limited usable information however is
available for General Practitioners, i.e. only for Diagnostics of Breast Cancer and
Rectal Bleeding. The guidelines are heterogeneous and difficult to interpret We
propose to improve this by agreeing on national multi-disciplinary referral guidelines
and provide synchronised online access to up-to-date and easy to interpret versions.
Provide service or online app to (self) register family history including family
relations, that can be coupled with routine healthcare registries and the EMR
used in primary care;
The best way to re-use and expand previously recorded family history information and
to view this history from the perspective of a different family member is by recording
parent–child relations and diagnoses directly with each correct family member. This
would require functionality to be added to the EMR. In order to overcome privacy
issues an online app or website to register family history is recommended (for example:
myfamilyhistory.com or familyhealthware.com).
4. Pro-active genetic services integrated in clinical practice facilitated by ICT (for example
family history registry and registry alerts);
For example, the GP or nurse practitioners should be able to (periodically) register
or consult family history information directly into the EMR. Accurate and up-to-date
treatment and referral guidelines and subsequent automatic alerts should pop up
when certain combinations of symptoms and familial risk factors indicate referral to
a clinical geneticist or other medical specialist.
150
8
Pro
Posed
ro
ad
ma
P to stePw
ise integ
ra
tion
of g
enetic
s in fa
mily
med
icin
e
Illustration of the proposed roadmap with a familial breast cancer case in clinical
research and family medicine.
Patient name: Angela B., Female, age 35.
Step 1: Angela lives in the city with her husband and two daughters aged 13 and 10. She
works as a hair dresser, has been happily married for a decade and the family just bought
a new home in the suburbs. She consults the GP on a busy Monday morning with the
following complaints: lump in left breast which she noticed during the weekend. The 4
cm irregular swelling is not painful but rather sensitive. The skin on the swelling is a little
red and dimpled. Angela has no medical history, but since you followed the oncogenetic
training for GPs a few weeks ago you are aware of the possible familial risks of breast
cancer and decide to take her family history. Angela’s mother died of breast cancer when
she was only 50 years of age 10 years ago. Her mother’s father had an unknown cancer
and died at age 55. Angela tells you, when you further ask her for her family history, her
sister had bilateral breast cancer at age 30 and died of ovarian cancer at age 33, two years
ago. Her two other and younger sisters seem healthy. On father’s side of the family no one
has been diagnosed with cancer yet.
Step 2: If proposed codes would be added the following could be registered:
Two first-degree family members with breast cancer at an early age: mother (died at
age 50) and sister (age 30, died 33, bilateral breast cancer). : A21.1 and A21.3 One first-
degree family member with ovarian cancer at an early age (sister age 30, died age 33).
Step 3: You are alarmed by the family history and the medical complaints of Angela. After
checking the referral guidelines for cancer online, you talk with Angela about referral to
the closest hospital as soon as possible for further diagnostics and possibly necessary
surgical treatment. You also inform her of the chance that she might be a carrier of a
DNA mutation which could be further analysed by a clinical geneticist. You promise to
call the clinical geneticist and discuss the problem. The clinical geneticist agrees Angela
needs further genetic DNA testing based on this positive family history and will invite
her this week to quickly start DNA testing, which may inform further treatment. You call
Angela afterwards and she is grateful for taking her case so seriously.
Step 4: Angela is alarmed by the fact that her positive family history for breast and ovary
cancer could mean an added risk to her and her daughters to develop breast or ovarian
cancer and decides to use the online tool to easily register her family history together
with her family members during the upcoming family reunion. Although it was a little
awkward at first to ask her family members for their medical history, they agreed to
do so anonymously online and repeat this every 5 years. Angela shows her family tree
151
8
Pro
Posed
ro
ad
ma
P to stePw
ise integ
ra
tion
of g
enetic
s in fa
mily
med
icin
e
online to her GP who registers relevant information in his EPD and uses this information
to build a pdf with only initials and years of birth of family members and adds this to her
record. Not only is she now able to take her family history to her GP, the other family
members who used the online tool are also able to do so. The whole family is enabled
to operationalize their family history through a snowball effect.
Step 5: Five years later Angela’s daughter Stephany, then aged 18, visits the GP with
gynaecological problems. She feels a painful swelling. She started to study law in a
different city and her new GP uploaded her medical and family history into his EPD. The
EPD has alarmed Stephany’s new GP with a pop-up that Stephany is carrier of a BRCA2
mutation since the clinical geneticist not only diagnosed Angela with a mutation, but
unfortunately also her two daughters. Angela’s daughter is frequently checked with
a physical and MRI by a surgeon familiar with familial breast- and ovarian cancer who
follows the national guidelines for familial cancer. Now that she has these complaints
you decide to call the surgeon and after careful deliberation you refer her the same day
to the clinic for further diagnostics. Fortunately, no abnormalities are found through the
gynaecological and vaginal ultrasound examination.
Extending translational genetic competences We offered our conceptual framework for stepwise integration of genetics into family
medicine and clinical research by adding codes to the ICPC-2 list and took oncogenetics as
an example. Of course this list could be further improved by adding codes in case of other
diseases commonly seen in family medicine such as diabetes, cardiovascular diseases and
monogenic subtypes (Maturity Onset Diabetes of the Young (MODY), BRCA 1/2, familial
hypercholesterolemia (FH) and long QT syndrome) in particular, are expected to come
increasingly to the forefront in primary care. Translational health education research is our
guiding principle to improve our translational efforts and ultimately improve (genetic) medical
care. Engaging colleagues in health education, clinical and biomedical research and medicine
in collaboration will enhance our collective ability to move research from the “data generated
from research projects” phase to the “changes in practice and policy” phase, which will
then bring us full circle to finally translate genetics in to primary care. As advances both in
genetic discoveries and health education research evolve, it will generate interdisciplinary
collaborative endeavors within the broader scope of public health and medicine. Impact
of these advances will only become manifest in better decision-making, better advocacy,
better health policy and finally improved health if GPs could play a key role in translating
potentially life-saving advancements in genetic technologies to patient care. If GPs are to
make an effective contribution in this area, not only their competencies need to be upgraded
by offering suitable and effective genetics training, but performance in real practice needs to
be facilitated as well by operationalizing integration of genetics in Electronic Patient Records.
152
8
Pro
Posed
ro
ad
ma
P to stePw
ise integ
ra
tion
of g
enetic
s in fa
mily
med
icin
e
REfEREnCES1. Scheuner MT, Sieverding P, Shekelle PG: Delivery of genomic medicine for common chronic
adult diseases: a systematic review. Jama 2008, 299(11):1320–1334.2. Kemper AR, Trotter TL, Lloyd-Puryear MA, Kyler P, Feero WG, Howell RR: A blueprint for
maternal and child health primary care physician education in medical genetics and genomic medicine: recommendations of the United States secretary for health and human services advisory committee on heritable disorders in newborns and children. Genet Med 2010, 12(2):77–80.
3. Khoury MJ, Gwinn M, Yoon PW, Dowling N, Moore CA, Bradley L: The continuum of translation research in genomic medicine: how can we accelerate the appropriate integration of human genome discoveries into health care and disease prevention? Genet Med 2007, 9(10):665–674.
4. Baars MJ, Scherpbier AJ, Schuwirth LW, Henneman L, Beemer FA, Cobben JM, et al: Deficient knowledge of genetics relevant for daily practice among medical students nearing graduation. Genet Med 2005, 7(5):295–301.
5. Dove-Edwin I, Sasieni P, Adams J, Thomas HJ: Prevention of colorectal cancer by colonoscopic surveillance in individuals with a family history of colorectal cancer: 16 year, prospective, follow-up study. Bmj 2005, 331(7524):1047.
6. Domchek SM, Friebel TM, Singer CF, Evans DG, Lynch HT, Isaacs C, et al: Association of risk-reducing surgery in BRCA1 or BRCA2 mutation carriers with cancer risk and mortality. Jama 2010, 304(9):967–975.
7. Burke W, Culver J, Pinsky L, Hall S, Reynolds SE, Yasui Y, et al: Genetic assessment of breast cancer risk in primary care practice. Am J Med Genet A 2009, 149A(3):349–356.
8. Rose PW, Watson E, Yudkin P, Emery J, Murphy M, Fuller A, et al: Referral of patients with a family history of breast/ovarian cancer–GPs’ knowledge and expectations. Fam Pract 2001, 18(5):487–490.
9. Houwink EJ, Henneman L, Westerneng M, van Luijk SJ, Cornel MC, Dinant JG, et al: Prioritization of future genetics education for general practitioners: a Delphi study. Genet Med 2012, 14(3):323–329.
10. Taylor MR, Edwards JG, Ku L: Lost in transition: challenges in the expanding field of adult genetics. Am J Med Genet C Semin Med Genet 2006, 142C(4):294–303.
Summary
CHaPtERSummarizing Discussion 9
159
9
Sum
ma
rizin
g D
iScu
SSion
SummaRIZInG DISCuSSIon
ReflectionWhen we, as patients, visit our General Practitioner (GP) for a persistent cough, when we
are worried about a lump we seem to feel in our breast, or about a little blood in our stool,
we assume our GP will record this. We expect he or she will also register measurements
and findings as well as a hypothesis or diagnosis for the cause of our complaints and
a plan to reach a diagnosis and treatment somewhere in the Electronic Health Record
(EHR). We also assume the GP will know about our worries when we visit her the next
time. We are surprised when we visit an out-of-ours clinic or emergency department in
the local hospital during the weekend and the doctors do Not seem to have access to our
EHR or don’t know anything about our medical history. We are annoyed when we have
to repeat the serious diagnoses we have had in the past to every next doctor we meet.
Diagnoses such as asthma, myocardial infarction or even cancer should be available to
the locum or the surgeon when our sprained ankle turns out to be broken.
From a patient perspective, despite the anxieties about privacy issues that are often
expressed, electronic health record (EHR) data reuse and sharing for purposes of care is
rational and desirable. In fact nowadays patients do expect high quality data recording and
sharing between doctors. Most patients do not object and are even happy to contribute
when researchers want to re-use their data (even genomic data) for research, as we know
from studies in the field of rare diseases[1]. Many patients also welcome early detection
of (genetic) risk at serious disease as is illustrated by studies but also by the popularity of
new e-health tools such as www.yourdiseaserisk.wustl.edu or www.testuwrisico.nl We
do not know how patients feel about the fact that their EHR records are being used to
assess quality of care but we know patients do expect high quality healthcare[2].
As we have shown in the introduction (chapter 1) data reuse and sharing is highly
desirable not only from a patient’s perspective but also from the perspective of the
researcher, the quality assessor and the GP. We know that reuse and sharing of EHR
data is already becoming commonplace despite serious concerns about data-quality and
subsequent reusability. We felt there was a need to quantify this problem in Primary Care,
with a special focus on diagnosis registry and diagnosis coding and on exploring novel
ways for obtaining complete data. Furthermore we wanted to explore ways to enable
EHR data reuse and sharing in a sensible way. We decided to broaden our horizons by
working with rare diseases as well as common diseases (cancer) but also by searching
cooperation with medical specialists (hospital EHRs) and bio-informaticians. We chose to
study disease coding by actually developing a coding system in the field of rare diseases
and participating in the development of a coding tool. Because we discovered a lack
of application of available genetics-knowledge in Primary Care, which is partly due to
limitations in the EHR, we decided to develop a roadmap on this subject.
160
9
Sum
ma
rizin
g D
iScu
SSion
We set ourselves the following aims which in our opinion, we have successfully
fulfilled.
Aims of the thesis1. Assess (aspects of) data quality in (parts of) the Primary Care EHR, focussing on
diagnosis registry as a central item and diagnosis coding as an important tool;
2. Find strategies and solutions to improve quality of (Primary Care) EHR data and to
contribute to the enabling of reuse and sharing of EHR data.
Summary of Results
Part One - Quality of data: literature review & hands on
identification of bottlenecks and areas for improvement
Literature on data quality in primary care is scarce; our review shows that available
studies focus mostly on completeness and some also on correctness of registry of a
small number of data items. Quality of data varies per GP centre and per data category.
In general, coded data such as diagnoses, medication prescriptions and laboratory test
results is registered fairly accurate and complete but there is room for improvement.
Registry of vital parameters, risk factors and allergies & intolerances is often incomplete
and incorrect (chapter 2).
For the studies described in chapters 3, 4 and 5 we used the routine EHR data extracted
from practice centres in the Utrecht area, the Netherlands, that are a member of the Julius
General Practitioners’ Network (JGPN; 120 GPs, 50 practice centres, 290,000 patients).
Coded and free-text primary care data from individual patients enlisted with these centres
is periodically extracted to the central anonymized EHR. From the two studies (chapters 3
and 4) we performed to assess quality (completeness and correctness) of diagnosis registry
in the primary care EHR we learned that the quality of coded data, as demonstrated for
patients with cancer or suspected of having cancer, is suboptimal. GPs do know their
cancer patients but this does not mean that re-users of data can find these cancer cases
using anonymized, coded EHR data easily. In both studies we compared cancer cases found
in the EHR with the Netherlands Cancer Registry (NCR), a reference standard which is
considered reliable. We found that when re-users of data try to select cancer cases using
only coded data on a population level (chapter 3) as well as on an individual level (chapter
4) a large number of cancer cases is due to be missed (up to 40% false-negatives#) and
a large number of cancer cases will be wrongly classified (up to 50%, false-positives#).
We conclude that the quality of coded EHR data improves over the years and that
the type of EHR system used influences data quality. More specifically we found that in
recent years diagnosis registry is more complete but as a drawback also the number of
false-positives increases. In our linkage study (chapter 4) we discovered that for 77%
of the missing (false-negative) cancer cases information about the cancer is available in
161
9
Sum
ma
rizin
g D
iScu
SSion
the EHR elsewhere, merely in un-coded plain text. Also for 38% of seemingly wrong
(false-positive) cases the GP appeared to have correctly registered the cancer diagnosis,
including 31% (of 38%) where the diagnosis is not or not yet retrievable from the NCR.
In our study described in chapter 5 we reused coded as well as free-text EHR data for
a research study to gain experience and by doing so also assessed GP management of
women with breast cancer related concerns. We selected for the period under study all
women from the EHR that presented with physical signs and symptoms of the breast (for
instance pain in the breast or a lump) but also women that presented with fear of breast
cancer or a family history of breast cancer. We found that concerns relating to breast
cancer are presented to an average GP frequently (incidence rate 25.9 per 1,000 women
per year), the larger part consisting of women experiencing physical signs and symptoms
of the breast (85.3% or 23.2 per 1,000 per year). Symptomatic and asymptomatic women
are referred for further investigation equally often (50%), so the GPs diagnostic workup
phase does not seem to be paramount in the decision process. Referral practice for annual
screening and genetic counselling is suboptimal and relevant information concerning
family history of cancer is often missing in the EHR. Identification and management
of women with an increased risk of breast cancer by GPs can be improved as well as
identification and reassurance of women without an increased risk or relevant symptoms.
In this study we presented incidence rates based on extracted EHR data, taking into
account the limitations of routine care data (see recommendations) but without applying
corrections to results because of lack of information on data quality of symptoms and
family history registry in Primary Care. Furthermore, considering the dimensions of data
quality (see table 1 introduction), in all three studies we experienced that data in the
EHR can be incomplete, incorrectly coded and not up to date (current). We also found
examples of lacking concordance and plausibility of data but these two dimensions were
not structurally assessed in our studies.
Part two: Strategies & Solutions for improving data-quality and
enabling reuse and sharing of EHR data
Improving disease coding systems and the development of tools mapping codes between
those systems can help to increase EHR data quality, not just within primary care, but
in health care in general. We have studied the quality of coding systems for EHR use in
the field of rare diseases, in particular of metabolic disorders. Collectively, this group of
diseases is large (>6.000) and growing steadily due to the identification of new diseases or
variants of known diseases and improved clinician awareness. We know from experience
that the annotation of rare diseases by means of adequate coding systems, and thus the
possibility to accurately code patients with these disorders and identify them in EHRs, has
been left behind. This has recently been confirmed by other researchers[3]. Our study,
as presented in chapter 6, demonstrates that there are large gaps in the widely used
existing international coding systems ICD-10 (International Classification of Diseases)
162
9
Sum
ma
rizin
g D
iScu
SSion
(76% missing codes) and SNOMED-CT (Sytematized Nomenclature of Medicine Clinical
Terms) of (54% missing codes) for metabolic disorders. Based on our clinical experience,
we suspect that there may be similar gaps for other types of rare disorders. This has also
been recognized by the SNOMED and Orphanet organisations which have joined efforts
recently to improve coding for rare diseases. Existing gaps are a barrier to database- and
data-sharing efforts especially for rare disorders, where the disease code is often used
as a key to communication. We have shown that with the help of dedicated clinicians
and code development agencies, the problem of coding gaps for rare disorders can be
successfully addressed and that rich and up to date coding systems actually contribute
to the quality of annotation for rare diseases, and thus to healthcare for patients with
these diseases. Although this study was performed in a hospital setting, patients with a
rare disease diagnosis should be recognizable as such also in Primary Care. Furthermore
this study provided insight in the extensive process of developing a high-quality, usable,
up-to-date coding system which will actually be adopted by prospective users.
Another barrier to data sharing for various purposes is the need to standardize
semantics of data values such as diagnoses and other phenotypical codes. Ideally coding
systems are aligned before data entry but often retrospective standardization will be
required. In chapter 7 we describe the development of SORTA, a software tool to ease
data (re-)coding and mapping between coding systems. We participated in this study by
using SORTA for a pilot project to map an existing Dutch coding system for phenotype
coding of physical symptoms to the international Human Phenotype Ontology (HPO) and
demonstrated that existing coding systems can be harmonized with significant speed
and quality improvement compared to earlier manual procedures.
Coding of disease diagnoses and symptoms is pivotal in EHR performance, but not all
types of relevant medical information, e.g. family history, may be captured well in this manner.
The EHR data structure design should allow for storing basically all relevant information and
matching codes should be available to capture that information. In addition, the quality
of the user interface itself is one of the other factors contributing to EHR performance.
These aspects were studied in the context of delivery of genetic services, which despite
readily available genetic knowledge is reported to be inadequate in Primary Care, among
other medical specialties. We have confirmed this problem in chapter 5 where we found
suboptimal referral practices for annual screening/genetic counseling but also information
missing in the EHR concerning family history of women with breast cancer related concerns.
In chapter 8 we identify existing barriers to implementation of genetic services in Primary
Care such as shortcomings in design and interface of EHR systems to register genetic
information. We propose a step-by-step roadmap including adjustments to the EHR and
to existing coding systems to integrate genetics in General Practice and clinical research.
This roadmap can be used as an example for introducing other complicated additions or
adjustments to the EHR or to coding systems driven by needs in daily medical practice.
163
9
Sum
ma
rizin
g D
iScu
SSion
Lessons LearnedSummarizing and interpreting the results of the studies performed we conclude there
are a number of lessons to learn from this thesis:
1. Data-quality in primary care currently is suboptimal, even for a key-item such as
the coded diagnosis and for a serious disease such as cancer; relevant information
regarding important risk factors such as family history is either frequently missing,
incorrectly coded or cannot be found easily;
2. Despite suboptimal quality and subsequent reusability with clear limitations, the
primary care EHR is a rich and voluminous source of (mostly uncoded) medical data,
often comprising many years of follow-up
3. Because of suboptimal quality, primary care data should only be reused by people
that fully understand the context of routine care data-entry and take explicitly into
account the limitations of this data which can be assessed using the checklist in
appendix 1 of this thesis;
4. There is a need to improve data-quality since reuse and sharing are desirable and
expanding, ideally at the source (at data entry) and supported by adequate coding
options. GPs can and should be facilitated and supported to achieve this in a number
of ways (see recommendations below);
5. Adequate and up-to-date coding systems are pivotal for data reuse and sharing,
not only for common but also for rare diseases and can successfully be developed
using not only coding agencies but also dedicated clinicians and can be facilitated
by software tools;
6. These coding systems are only valuable if they are continually be maintained, provide
adequate synonyms and relevant crosslinks to other systems and are equipped with
a guideline for use and an extensive fool-proof thesaurus;
7. Obligatory coding in EHR systems results in more complete registry but also leads to
(over-) registration errors;
8. Linkage of EHR records to other data-sources can be useful to validate diagnoses but
is currently complex and time-consuming;
9. The Primary Care EHR can be complementary to other data sources, even to a known
reliable reference standard such as the Netherlands Cancer Registry;
10. Concerning reuse of Primary Care EHR data, there are many stakeholders involved,
all interested in data reuse but from different perspectives: patients, GPs, the Dutch
Association of General Practitioners (NHG), EHR suppliers, health inspection/insurance
companies, (quality assessors), hospitals and out-of-hours clinics, researchers,
Academic Practice Based Research Networks, departments of Vocational Training
for GPs, Coding agencies/organizations, and owners of External data sources such
as the NCR.
164
9
Sum
ma
rizin
g D
iScu
SSion
RecommendationsThe lessons learned can be translated into recommendations for the various stakeholders
involved.
Patients
Patients should be made aware of anonymous and non-anonymous EHR data reuse which requires their consent. Patients should be made aware that there is a
difference between reuse of data for purposes of care, which cannot be done anonymously,
and re-use for purposes such as research, which can be done with anonymized patient
data. Also, there are ways to let re-users work with anonymized data and still provide
options to contact the patient if necessary through a third party. This means there are
options to avoid privacy issues.
Patients should be stimulated to take responsibility and make sure all important diagnoses and information regarding allergies and intolerances is known and registered in their medical file. The Primary Care EHR is central
for registry of a patient’s medical condition over the years and the GP in particular
has an overview of this data. GPs receive results from laboratory and diagnostic
tests and medical reports whenever their patients visit a medical specialist or a
paramedical professional. Patients have the right to read and check their own medical
files including their EHR record at the GP and should do so at least once to suggest
possible corrections and additions. They should take responsibility and make sure all
important diagnoses are recorded as Episodes in their EHR, as well as information
regarding allergies and intolerances. If patients are convinced their GP is adequately
keeping records and they trust their GP to keep doing this, it is safe and better to let
him/her do the file-keeping. In recent years a number of companies have introduced
software to maintain your own online medical file as a patient, beside or instead of
the doctors’ file, such as www.zorgdoc.nl, www.patient1.nl or www.healthvault.
com. It is however not easy to assemble and interpret all the right information to
keep your own medical file, as is illustrated by the premature exit of Google Health in
2011 (http://www.medischcontact.nl/Nieuws/Laatste-nieuws/Nieuwsbericht/125705/
Google-health-stopt.htm ). However, for patients with chronic or rare diseases the
keeping of personal records with certain measurements and symptoms or complaints
can be very useful, especially when these personal records could be combined with
medical files in the future. When the EHR record is complete and correct, only privacy
issues could be barriers to data reuse and sharing, for instance when “opting-in” to
share data through the LSP[4], and these issues should be individually weighed by
every patient.
165
9
Sum
ma
rizin
g D
iScu
SSion
General Practitioner (GPs)
GPs should improve data-quality by investing in updating and coding key-items in their EHR files and by optimizing working processes at the GP practice. Our research shows that GPs do know their patients but relevant information is
missing or hidden, because it is uncoded, within the EHR text and this is barring and
confounding data reuse and sharing. Data quality in many GP EHRs can be improved.
Despite the fact that we acknowledge the need to lessen the administrative burden
and we therefore support like almost 70% of Dutch GPs the movement “het roer
moet om”[5] we do feel data reuse will only expand and GPs can benefit if they
improve their data quality. They should however be supported and facilitated much
more than they are today (see recommendations below to Dutch College of General
Practitioners and EHR suppliers). There are several ways to improve data quality from
a GP perspective as we will explain.
GPs should invest in updating key-items (such as the important diagnoses, allergies
& intolerances and risk factors) in their EHR files starting by making sure there are coded
Episodes, correctly dated, for every disease the patient suffered or suffers that could
have medical consequences or could be important regarding future medical decisions for
the patient and/or his family. It is important to add the code for a disease only after the
diagnosis has been determined and for suspected disease to code the main symptom,
in line with the Guideline for Adequate EHR registry (ADEPD) Guideline ([6]). Also, the
date of the Episode should actually be the date the final diagnosis was made, not the
date of data entry.
GPs should evaluate and update working processes at the GP practice to integrate
diagnosis registry after a letter from a hospital or diagnostic laboratory is received.
Although most EHR systems do not adequately support registry of a positive family
history for genetic disease such as cancer, it is useful to record this information since
it can have major consequences (suggestion: create a separate Episode and code this
with one of the available ICPC-1 codes for positive family history A29.01 – A29.07).
We expect data reuse and sharing to expand in the coming years since this is the key
to transition of care, such as the follow- up care for cancer patients from hospitals to
Primary Care. We also expect, like the NHS in England (https://www.gov.uk/government/
publications/personalised-health-and-care-2020 ), that it will not be long before patient
empowerment will be taken a step forward and patients will seek and gain the right to
access their full EHR record online and will even be able to add notes to their EHR (but
not edit medical entries made by the GP). From a GP perspective it is undesirable to reuse
and share low quality data for any purpose but certainly not with the patient since this
can have many negative consequences (see discussion chapter 4), besides the risk of the
patient losing confidence in their GP.
166
9
Sum
ma
rizin
g D
iScu
SSion
We recommend GPs to participate in relevant projects which require data reuse and sharing. We recommend GPs, that have improved the quality of their
data, to participate in projects, for instance research, by sharing patient data. Sharing
patient data for other purposes will further stimulate to improve its quality. Privacy
barriers can be addressed and suboptimal coding could potentially be solved by arising
text mining techniques[7]. Participating in a Practice Based Research Network at a local
Academic centre can be feasible for GPs for instance to obtain benchmarking information
(“spiegelinformatie”) which can be valuable for management purposes. Last but not
least, many EHR systems are provided with often unrecognized but useful ICT functions,
for instance to build selection queries which can also be used to assess patient mix and
assemble relevant management information.
Dutch College of General Practitioners (NHG)
the Dutch College of General practitioners (nHG) should on the one hand support and facilitate improvement of data quality, on the other restrict the unlimited reuse and sharing of EHR data. This is especially true for uncoded plain
text in the EHR which is more prone to misinterpretation; reuse should be restricted to
those that fully understand the Primary Care context in which the data was registered.
The focus should be on improving, if possible (re-) coding, and subsequently sharing
and reusing only a limited set of key data items, including the diagnosis. The quality of
these data might be improved by the implementation of tools that can assist in recoding
text (f.i. text mining and optimizing thesauri). This is in line with national developments
such as the Continuity of Care record developed in the project Registry at the Source
fom the NFU (Netherlands Federation of University Medical Centres) (http://www.nfu.nl/
thema/registratie-aan-de-bron/ ). These key-items should be chosen as a subset of the
seventeen data-items that are part of the Continuity of Care record for hospitals, coded
with SNOMED.
the Dutch College of General Practice should study and test techniques such as voice-recognition and text-mining to facilitate recording of high quality data at the source. Supporting the improvement of data quality can be done in a number of
ways, without losing sight of the actual goal of the EHR: supporting the primary process
in every-day General Practice. For the GP this means facilitating easy, fast and user-friendly
recording of consultations. In most EHRs this is suboptimal now and actions taken to
improve data quality should not add to this burden but rather enhance functionality.
This means that existing and upcoming techniques like voice-recognition and text-mining
should be studied and tested.
167
9
Sum
ma
rizin
g D
iScu
SSion
furthermore the implementation of the EHR Reference model by EHR suppliers should be prioritized and stimulated and amendments should be made to the aDEPD and the Reference model. Also, the publication of the EHR reference model
for EHR suppliers provides a firm basis but the implementation of the standard needs
immediate attention. Many EHR suppliers have not yet implemented recent versions of
the model. Furthermore the Reference Model and ADEPD guidelines should be amended
on, among other things, the following items:
1. The Episode date should be the date the diagnosis was made (not the date of the
first attached consultation) and hence the system should propose this date and
furthermore it should be possible to alter this date. For instance for assessing familial
risk of cancer, but also for research purposes it is important to know the age at
diagnosis;
2. Enable attaching consultations to more than one Episode which will hugely simplify
en fasten registration of consultations with multiple symptoms. Investigate the options
and consequences of selecting more than 1 ICPC code per consultation;
3. Enabling the registry of a family history within the context of the EMR is necessary
and should be facilitated for instance following the roadmap we presented in
chapter 8;
4. Develop and add to the Reference Model a list of integrity checks for the EMR
(for instance: it should not be possible to register prostate cancer with a female
patient);
5. Design ways that support easy registration of suspected disease/differential diagnoses,
pathological-proven disease, recurrent disease, etcetera;
6. Enable registry of a suspected or proven rare disease including relevant coding, by
making use of existing and supporting connected coding systems, and develop a way
to make these patients “visible” for the GP.
Investigate the digital integration of guidelines in the EHR and monitor the further development of the ICPC coding system. Furthermore it would be useful
to investigate ways to integrate guidelines in the EHR, for instance to support GPs in risk
assessment of familial cancer and subsequent referral. Last but not least, the monitoring
and further development of the ICPC-1 coding system maintains to be an important issue.
This could be improved by extending the ICPC-1 with useful codes (for instance such as
suggested in chapter 8), but also by providing a fool-proof thesaurus that would suggest
coding options during registration of consultations. This thesaurus should comprise
adequate synonyms and relevant crosslinks (f.i. with SNOMED) to facilitate data-sharing
between sources.
168
9
Sum
ma
rizin
g D
iScu
SSion
EHR suppliers
Investigate how user-interface and system design can be adjusted to support high-quality data entry at the source. We demonstrated that the type of EHR
system used influences data quality in the EHR. EHR suppliers should investigate
(in cooperation with an academic research team) how their user-interfaces and system
design can be adjusted to actually improve data quality. This thesis provides a number
of directions: facilitate user-friendly and accurately coded diagnosis registry without
encouraging false-positives, add options to directly suggest and register the correct
date of diagnosis, suspected, recurring and metastasizing disease, treatment, increased
markers (eg for Prostate Specific Antigen/ or PSA) and a positive family history within
the context of the EHR. Extending integrity checks at data entry would improve data
quality as well.
Integrate referral guidelines into the EmR and facilitate feedback by easy-to-use selection queries. Furthermore EHR suppliers should think about optimizing the
availability of [online] up-to-date and easy to use referral guidelines by integrating them
into the EMR. Also, facilitating feedback on a practice level by providing easy-to use
selection queries for the GP would be worthwhile. Last but not least, the feasibility of
voice-recognition and text-mining to facilitate structured data entry and retrieval should
be investigated.
Quality Assessors (health inspection/insurance companies)
Stop measuring the quality of registration and find ways to adequately measure quality of care in dialogue with GPs. Quality Assessors should be aware
that data quality in Primary Care is suboptimal, even for a key item such as a cancer
diagnosis. The current list of indicators (www.nhg.org/themas/publicaties/download-
indicatoren) in Primary Care that can be calculated by retrieving information from the
EHR relies heavily on adequate disease coding, for instance because the total number
of patients with a certain disorder is used as a denominator. Taking into account the
patient mix of a practice is justly becoming more important, which is another reason
to aim for a reliable denominator. Also, by identifying patients with a certain disorder
such as asthma or diabetes, information registered within those patients records such
as smoking or blood pressure measurements are counted. This means that incorrect or
incomplete diagnosis registry will bias results. Furthermore, we suspect that the quality
of data for items in the EHR such as risk factors will also be suboptimal. This means
that the “paper tiger” that is being created before our eyes, measures, inaccurately,
the quality of registration instead of the quality of care and should be stopped, the
sooner the better.
169
9
Sum
ma
rizin
g D
iScu
SSion
Education
add “adequate health file recording” to the mD curriculum and teach GPs necessary skills. Since digital health file recording is standard procedure in General
Practice and more recently also in hospitals, there is a need for a new subject which should
be added to the MD curriculum at University called “ adequate health file recording”.
Doctors and GPs (to be) should be made more aware of consequences of recording
choices and be taught necessary skills such as correct coding. This subject should also be
integrated more widely into courses for practising GPs and other medical professionals too.
Researchers & Practice Based Research Networks (PBRNs)
When working with routine care data, validate diagnoses. Using EHR data means
understanding and taking into account the limitations of routine care data. If researchers
choose to work with EHR data, the diagnoses should be validated, either by linkage to
external sources or other means. Linkage to other sources could decrease the number
of false-negative records and hence more cases could be traced and included. False-
positive records can only (partly) be identified by studying the full EHR text, which is
time-consuming and may be undesirable considering privacy issues.
PbRns should seize the opportunities to support participating General Practitioners in improving the data quality in their EHRs for instance through providing benchmarking information. In this way they could provide additional
advantages to practices to participate in their Networks. By providing bench marking
and management information GPs could actually assess their patient mix and find EHR
records that need quality improvement.
Future Research crossing boundariesIn this thesis we have been able to successfully assess data quality in Primary Care for
certain data items and have also been able to identify strategies and solutions to improve
data quality to actually enable reuse and sharing. We realize these are pieces of a large
jigsaw that has to be completed in the coming years.
We believe we have only been able to obtain results and devise recommendations
because we have crossed boundaries between academic disciplines: primary care, clinical
genetics, medical informatics, computer science and bio-informatics. Working with
scientists from other disciplines provides new insights and solutions to research questions.
A number of research challenges remain to be studied in the near future, all of them
interdisciplinary. First of all, beside completeness and correctness, other dimensions of data
quality should be evaluated: concordance, plausibility and currency, not only for diagnosis
registry but also for other key data items, for instance risk factors, treatments and allergies.
170
9
Sum
ma
rizin
g D
iScu
SSion
Secondly, user-interface designers should be involved in these studies: what aspects
of user interfaces in the different EHRs lead to differences in data quality and how can
user interfaces be improved to enhance data quality?
Thirdly, the design, implementation and evaluation of actual interventions in the
GP practice to improve data quality could provide the effective interventions needed to
improve daily practice.
In the fourth place, possibilities of natural language recognition to suggest coding
alternatives during data entry should be investigated.
Last but not least, it is necessary to experiment with patient entered data to assess
usability of this data for various purposes, first of all care. This can be done for instance by
developing an app or software tool, that patients can use to enter family history information.
About our patientJanuary 2020: a 48-year-old male visits his General Practitioner (GP) for a persistent mucus
producing cough. A few days ago he made the appointment online and entered the reason
for consultation and his complaints in the text box. Also he answered a few multiple
choice questions presented by the EHR system triggered by the reason for encounter,
about his complaints. Just before the allotted time the GP reads this information and
looks at the patients’ personal health data which includes data from various apps the
patient uses such as exercise-apps. She notices that the patient has lost some weight but
also that the training frequency and duration of this running-enthusiast have decreased
substantially in the last 4 weeks. The GP asks some additional questions and performs
a physical examination that turns out to be normal. She summarizes her findings orally
using speech recognition software and along the way selects relevant codes, prompted
by the system, for symptoms, signs and differential diagnosis. On her screen a pop-up
appears (based on the guideline “Acute cough”) asking her if a request for a chest X-ray
should be send to the nearest hospital selected on diagnostic quality, reimbursement by
the patient’s health care insurance company and shortest waiting list. She clicks “yes”
and schedules an e-consultation for follow-up a week later. The GP has a few minutes
to spare and chats with her patient about his wife, children and his new job.
This thesis hopefully contributes to the improvement of EHR data in general and to the
exposure of the true goldmine these data can become, with the ultimate goal to improve
care for patients with common and rare diseases.
Notes# False-negatives are cancer cases that are present in the NCR but not in the Primary
Care EHR
False-positives are cases that are registered in the Primary Care EHR as having cancer
but are not present in the NCR
171
9
Sum
ma
rizin
g D
iScu
SSion
Checklist before EHR data reuse and/or sharing
nr Item Check
Relevance
1 List the data items you need
2 Critically assess the items you just listed on necessity to answer your (research) question. Delete every item that is not absolutely necessary
Data Quality (for each data item)
3 Gather existing information on the quality of data for each item (dimensions: completeness, correctness, concordance, plausibility, currency)
origin (for each data item)
4 Who entered this data?
5 For what purpose was this data entered?
6 What information is captured with this data?
7 What information is NOT captured with this data?
8 Could entry of this data be biased in any way?
9 Are there other ways to enter the same data in this system?
10 Could another user with the same role, decide to enter this data differently or not at all?
Condition (for each data item)
11 When was the data entered (relative to disease process)?
12 Is there any metadata available?
13 Was the data changed since entry, why and by whom?
14 Was there financial benefit for registering this data at all or in a certain way?
15 List possible errors that could have occurred at data entry
format (for each data item)
16 What format was used entering the data? If coded: what coding system?
17 If a coding system was used: what version, using which instructions? Check out the alternative codes in the system for registry of this item.
18 Were there any restriction rules in the EHR system for entry of this data?
assessment
19 Critically assess every data item using the information gathered and determine usefulness for answering (research) question.
NB: privacy /policy issues are not included in this list
172
9
Sum
ma
rizin
g D
iScu
SSion
REfEREnCES1. Burstein MD, Robinson JO, Hilsenbeck SG, et al. Pediatric data sharing in genomic
research: attitudes and preferences of parents. Pediatrics 2014;133:690–7. doi:10.1542/peds.2013-1592
2. Lateef F. Patient expectations and the paradigm shift of care in emergency medicine. J Emerg Trauma Shock 2011;4:163. doi:10.4103/0974-2700.82199
3. Fung KW, Richesson R, Bodenreider O. Coverage of rare disease names in standard terminologies and implications for patients, providers, and research. AMIA Annu Symp Proc 2014;2014:564–72.http://www.ncbi.nlm.nih.gov/pubmed/25954361 (accessed 16 Jun2016).
4. Zorgcommunicatie) V (Vereniging Z voor. Sharing your medical file and the LSP (Brochure: Uw medische gegevens elektronisch delen?). https://www.vzvz.nl/page/Zorgconsument/Links/Informatie/Informatiemateriaal
5. Het manifest van de bezorgde huisarts. Het roer moet om (free translation: ‘We need a radical change’. www.hetroermoetom.nu; www.hetroergaatom.lhv.nl
6. The Dutch College of General Practitoners. Guideline adequate EHR registry. Revised version 2013. Available at: https://www.nhg.org/themas/publicaties/richtlijn-adequate-dossiervorming-met-het-epd.
7. Hoogendoorn M, Szolovits P, Moons LMG, et al. Utilizing uncoded consultation notes from electronic medical records for predictive modeling of colorectal cancer. Artif Intell Med 2016;69:53–61. doi:10.1016/j.artmed.2016.03.003
CHaPtERNederlandse Samenvatting
Dankwoord
Curriculum Vitae
10
177
10
Ned
erla
Nd
se sam
eNv
at
tiNg
nEDERLanDSE SamEnvattInG
ReflectieWanneer wij naar onze huisarts stappen voor een hoest die maar niet over gaat, of wanneer
we bezorgd zijn over een knobbel in de borst of wat bloed bij de ontlasting, dan gaan
we ervan uit dat onze huisarts dit noteert. We verwachten ook dat hij of zij de resultaten
van het lichamelijk onderzoek registreert in het huisartsgeneeskundig dossier (ook wel
HIS = Huisartsen Informatie Systeem), samen met een hypothese of diagnose die onze
klachten kan verklaren en natuurlijk een plan voor diagnostiek en/of behandeling. We
gaan er dan ook vanuit dat hij of zij onze zorgen kent wanneer we opnieuw langskomen.
We zijn daarentegen verbaasd wanneer we in het weekend na een sportongelukje een
huisartsenpost of Eerste Hulp bezoeken in ons lokale ziekenhuis en de dokters blijken Niet
in ons dossier te kunnen kijken en Niets te weten over onze medische voorgeschiedenis.
We raken geïrriteerd wanneer we belangrijke diagnoses keer op keer moeten herhalen bij
iedere nieuwe dokter die we zien. Diagnoses zoals astma, een doorgemaakt hartinfarct
of kanker zouden inzichtelijk moeten zijn voor de huisarts die dienst heeft op de post
én voor de eerste hulp arts wanneer onze verstuikte enkel toch gebroken blijkt te zijn.
Vanuit patiënten perspectief is hergebruik van gegevens uit het huisartsgeneeskundig
medisch dossier voor zorgdoeleinden logisch en wenselijk, ondanks zorgen die regelmatig
worden geuit, bijvoorbeeld in de media, met betrekking tot privacy aspecten. Vandaag
de dag verwachten patiënten dat medische gegevens worden vastgelegd, dat deze van
goede kwaliteit zijn en dat deze ook worden gedeeld tussen dokters. We weten uit studies
in het veld van de zeldzame ziektes dat de meeste patiënten geen bezwaar hebben en
blij zijn om te kunnen bijdragen wanneer onderzoekers hun medische gegevens willen
hergebruiken. Dit geldt zelfs voor informatie betreffende ons genoom (genetische data).
Veel patiënten staan positief tegenover vroege detectie van (genetisch) risico op ernstige
ziektes, waarvoor dit soort gegevens nodig zijn, blijkt uit diverse studies maar ook uit
de populariteit van e-health applicaties zoals www.yourdiseaserisk.wustl.edu of www.
testuwrisico.nl. We weten echter niet wat patiënten vinden van het feit dat hun medische
gegevens ook worden gebruikt om kwaliteit van zorg te evalueren, maar we weten wel
dat patiënten een hoge kwaliteit van zorg verwachten.
Zoals we hebben laten zien in de introductie (hoofdstuk 1), zijn het hergebruik en
het delen van medische gegevens wenselijk, niet alleen vanuit het perspectief van de
patiënt, maar ook vanuit het perspectief van de onderzoeker, de kwaliteits-beoordelaar
en de huisarts zelf. We weten dat hergebruik en delen van medische gegevens al op grote
schaal gebeurt, ondanks zorgen met betrekking tot de kwaliteit van deze gegevens en
de daaruit volgende herbruikbaarheid. Wij vonden dat er aan de ene kant behoefte was
om dit probleem in kaart te brengen en te kwantificeren voor de huisartsgeneeskunde,
met een speciale focus op de registratie van diagnoses en diagnose codering en aan de
178
10
Ned
erla
Nd
se sam
eNv
at
tiNg
andere kant dat er behoefte was aan het bedenken van nieuwe manieren om medische
data van goede kwaliteit te verkrijgen. Verder wilden we ook uitzoeken of en hoe
medische gegevens uit het huisartsendossier op een verstandige manier kunnen worden
hergebruikt en gedeeld. We besloten onze horizon breed te maken door zowel met
zeldzame als met meer alledaagse ziektes (veel voorkomende kankers) te werken en door
samenwerking te zoeken met medisch specialisten (ziekenhuis Electronische Patiënten
Dossiers of EPD’s) en met (bio-)informatici. We kozen ervoor om diagnose codering te
bestuderen door zelf een code systeem te ontwikkelen binnen het veld van de zeldzame
ziektes en door te participeren in de ontwikkeling van een coderings-applicatie. Omdat
we ontdekten dat er een gebrek is aan toepassing van beschikbare kennis van genetica
in de huisartsenzorg, gedeeltelijk door beperkingen in de huidige HIS-sen, besloten we
om een “roadmap” te ontwikkelen op dit vlak.
Samenvattend stelden we onszelf de volgende doelen:
Doelstellingen van dit promotie-onderzoek1. Beoordeel (aspecten van) data-kwaliteit in (delen van) het huisartsgeneeskundig
medisch dossier, daarbij focussend op diagnose registratie als centraal onderdeel en
diagnose codering als belangrijk middel;
2. Vind strategieën en oplossingen om de kwaliteit van gegevens in het
huisartsgeneeskundig medisch dossier te verbeteren en om het hergebruik en het
delen van deze gegevens mogelijk te maken.
Samenvatting van de Resultaten
Deel 1 – Data kwaliteit: literatuur onderzoek en ‘hands-on’
identificatie van knel- & verbeterpunten
Literatuur op het vlak van data kwaliteit in de huisartsgeneeskunde is schaars; ons
literatuuronderzoek laat zien dat beschikbare studies vooral gericht zijn op compleetheid
en sommige op correctheid (accuraatheid) van een klein aantal data-items. De kwaliteit
van data varieert per huisartsenpraktijk en per data categorie. In het algemeen worden
gegevens die kunnen worden gecodeerd, zoals diagnoses, medicatie voorschriften en
resultaten van laboratorium onderzoek, vrij correct en compleet geregistreerd maar er
is ruimte voor verbetering. Registratie van vitale parameters, risico factoren en allergieën
& intoleranties gebeurt vaak incompleet en incorrect (hoofdstuk 2).
Voor de drie studies die worden beschreven in hoofdstuk 3, 4 en 5 maakten
wij gebruik van een grote database van het Julius Huisartsen Netwerk waarvoor
periodiek geanonimiseerde gegevens worden geëxtraheerd uit de medische dossiers
van 120 huisartsen uit 50 huisartsenpraktijken in de regio Utrecht (290.000 patiënten).
Uit de twee studies die wij uitvoerden om de kwaliteit van diagnoseregistratie in
het huisartsgeneeskundig dossier te meten (hoofdstuk 3 en 4), leerden we dat de
179
10
Ned
erla
Nd
se sam
eNv
at
tiNg
kwaliteit van gecodeerde data, zoals aangetoond voor patiënten met kanker of
een vermoeden op kanker, suboptimaal is. Huisartsen kennen hun kankerpatiënten
maar dit betekent niet dat her-gebruikers van data deze kankerpatiënten eenvoudig
kunnen vinden door gecodeerde, geanonimiseerde huisartsendata te gebruiken. In
beide studies vergeleken we kankergevallen gevonden in het huisartsendossier met
de Nederlandse Kanker Registratie (NKR) (www.cijfersoverkanker.nl), een referentie
standaard die wordt gezien als zeer betrouwbaar. Wanneer her-gebruikers van data
proberen om kankergevallen te vinden door gecodeerde gegevens te gebruiken op
populatieniveau (hoofdstuk 3) maar ook op individueel niveau (hoofdstuk 4), zal een
groot deel van de kankergevallen worden gemist (tot wel 40% fout-negatieven#) en
een groot deel van de gevonden kankergevallen zal fout geclassificeerd zijn (tot 50%
fout-positieven#).
Wij trekken uit deze studies de conclusie dat de kwaliteit van gecodeerde data in
het huisartsgeneeskundig dossier verbetert over de jaren heen, maar ook in recente
jaren suboptimaal is en dat het type HIS systeem de kwaliteit beïnvloedt. Meer specifiek
vonden we dat in recente jaren de diagnose registratie completer is maar als nadeel heeft
dat het aantal fout-positieven stijgt. In onze ‘linkage’-studie beschreven in hoofdstuk 4,
waarvoor we patiënten 1-op-1 koppelden aan de NKR, ontdekten we dat voor 77%
van de missende (dus fout-negatieve) kankergevallen, er wel informatie over de kanker
beschikbaar is, elders in het medisch dossier van de patiënt, meestal in de vorm van
ongecodeerde platte tekst. Ook vonden we dat voor 38% van de ogenschijnlijk foutieve
(fout-positieve) kanker gevallen, de huisarts toch de kankerdiagnose correct heeft
geregistreerd, waarbij voor 31% (van de 38%) de diagnose niet of nog niet beschikbaar
is in de Nederlandse Kanker Registratie.
Om ervaring op te doen hergebruikten wij zelf in onze studie beschreven in hoofdstuk 5
gecodeerde gegevens maar ook vrije tekst uit het huisartsgeneeskundig dossier voor
een onderzoek naar het management van vrouwen die met borst kanker gerelateerde
problemen de huisarts bezoeken. Wij selecteerden voor de onderzoeksperiode alle
vrouwen uit het HIS die zich presenteerden met fysieke klachten en symptomen van de
borst (bijvoorbeeld pijn in de borst of een knobbel in de borst) maar ook alle vrouwen
die naar de huisarts stapten met angst voor borstkanker of met borstkanker in de
familie. We ontdekten dat borstkanker gerelateerde problemen vaak bij de huisarts
worden gepresenteerd (incidentie 25.9 per 1.000 vrouwen per jaar), het grootste deel
bestaat daarbij uit vrouwen met fysieke klachten en symptomen (85.3% of 23.2 per
1.000 per jaar). Ongeveer de helft van de vrouwen wordt doorverwezen voor (meestal
beeldvormend) onderzoek, ongeacht of zij klachten hebben van de borst of niet. Kennelijk
weegt de werkhypothese van de huisarts niet het zwaarst in het besluitvormingsproces.
Verwijzingen voor jaarlijkse screening en genetische counseling blijken suboptimaal en
relevante informatie betreffende de familie anamnese van kanker mist vaak in het HIS.
180
10
Ned
erla
Nd
se sam
eNv
at
tiNg
De identificatie en het management van vrouwen met een verhoogd risico op borstkanker
kan worden verbeterd net zoals de identificatie en geruststelling van vrouwen zonder
verhoogd risico of relevante symptomatologie.
In de laatstgenoemde studie presenteerden we incidentie cijfers gebaseerd routine
zorg gegevens uit huisartsgeneeskundige dossiers, hierbij rekening houdend met de
beperkingen van die data (zie ook aanbevelingen) maar zonder correcties toe te passen
aan de resultaten omdat we onvoldoende informatie hadden over de data kwaliteit. Ook
ondervonden we in alle drie de hierboven beschreven studies dat huisartsgeneeskundige
data incompleet, incorrect gecodeerd en niet tijdig kan zijn. We vonden ook voorbeelden
van onvoldoende concordantie en geloofwaardigheid, maar deze dimensies van data
kwaliteit zijn door ons niet structureel beoordeeld.
Deel 2 – Strategieën en Oplossingen om de kwaliteit van gegevens in
het huisartsgeneeskundig medisch dossier te verbeteren en om het
hergebruik en het delen van deze gegevens mogelijk te maken
Het verbeteren van diagnosecode systemen en de ontwikkeling van tools (digitaal
gereedschap) om codes te ‘mappen’ tussen systemen kan helpen om de kwaliteit van
data in het medisch dossier te verbeteren, niet alleen binnen de huisartsgeneeskunde
maar binnen de gezondheidszorg in het algemeen. Wij hebben de kwaliteit van
coderingssystemen voor het medisch dossier binnen het veld van de zeldzame ziektes,
en meer specifiek de metabole ziektes, bestudeerd. De totale groep zeldzame ziektes is
groot (> 6.000 ziektes) en neemt nog verder toe door de ontrafeling van nieuwe ziektes
of varianten van bekende aandoeningen maar ook door een verbeterde ‘awareness’
bij clinici. We weten uit ervaring dat het annoteren van zeldzame ziektes via adequate
coderingssystemen en dus de mogelijkheid om patiënten accuraat te registreren in
medische dossiers, onvoldoende is en dit is recent bevestigd door andere onderzoekers.
Onze studie, zoals weergegeven in hoofdstuk 6, laat zien dat er grote gaten zitten
in wereldwijd veelgebruikte code-systemen zoals ICD-10 (International Classification
of Diseases) (76% missende codes) en SNOMED-CT (Systematized Nomenclature of
Medicine Clinical Terms) (54% missende codes) voor metabole ziektes. Wij verwachten
dat er vergelijkbare gaten zullen zijn voor andere groepen zeldzame ziektes. Recent is
dit probleem ook herkend door twee organisaties die zich intensief bezighouden met
codering, te weten SNOMED en Orphanet en is er initiatief genomen om gezamenlijk de
codering van zeldzame ziektes te verbeteren. Gaten in codesystemen vormen barrières
voor het delen van data, met name voor zeldzame ziektes, waar de diagnosecode vaak
wordt gebruikt als een sleutel tot communicatie. Wij hebben laten zien dat met de
hulp van ervaren clinici en codeer organisaties het mogelijk is gaten in codesystemen
te dichten en dat het ontwikkelen van rijke en up-to-date codesystemen daadwerkelijk
bij kan dragen aan de kwaliteit (zeldzame) diagnoseregistratie en daarmee indirect
ook aan de zorg voor patiënten met een (zeldzame) aandoening. Ook al werd deze
181
10
Ned
erla
Nd
se sam
eNv
at
tiNg
studie uitgevoerd in een ziekenhuisomgeving, patiënten met een zeldzame ziekte
zouden herkenbaar moeten zijn, ook binnen de huisartsgeneeskunde en dus in
het huisartsgeneeskundig dossier. De studie in dit hoofdstuk gaf ook inzicht in het
uitgebreide proces van het ontwikkelen van een kwalitatief goed, bruikbaar en up-
to-date codesysteem.
Een andere barrière voor het delen van data is de noodzaak om semantiek van
data velden zoals diagnoses maar ook andere fenotypische codes te standaardiseren.
In het ideale geval worden codesystemen eerst op elkaar afgestemd, dus voor data
invoer, maar vaak zal echter retrospectief standaardisatie nodig zijn. In hoofdstuk 7
beschrijven we de ontwikkeling van SORTA, een software tool waarmee data (her)-
codering en ‘mapping’ tussen codesysteem wordt gefaciliteerd. Wij namen deel in
deze studie door SORTA in de praktijk te gebruiken voor een pilot project waarin
een bestaand Nederlands codesysteem voor symptoom (of fenotype) codering werd
‘gemapt’ met een internationaal codesysteem voor fysieke symptomen (HPO = Human
Phenotype Ontology) en wij lieten zien dat bestaande codesystemen vrij snel en met
voldoende verbetering in kwaliteit kunnen worden geharmoniseerd, vergeleken met
eerdere handmatige procedures.
Het coderen van ziektes en symptomen is cruciaal voor een goed medisch dossier,
maar niet alle relevante medische informatie kan goed worden gevangen op deze
manier, bijvoorbeeld de familie anamnese. Het ontwerp van een huisartsgeneeskundig
dossier zou echter registratie van alle relevante medische informatie mogelijk moeten
maken, daarbij waar mogelijk gebruik makend van codesystemen. Verder blijkt dat
de kwaliteit van de user interface ook een belangrijke factor is die bijdraagt aan
data kwaliteit en prestatie van een huisartsinformatie systeem. Deze aspecten
werden door ons bestudeerd in het kader van het leveren van genetisch advies,
wat ondanks beschikbare genetische kennis niet adequaat blijkt te zijn binnen de
huisartsgeneeskunde, maar ook in diverse andere medische specialismen. We hebben
dit probleem bevestigd in hoofdstuk 5 waar we suboptimale verwijs routines vonden
voor jaarlijkse screening en genetische counseling maar ook missende informatie
in het medisch dossier met betrekking tot de familie anamnese van vrouwen met
borstkanker gerelateerde klachten. In hoofdstuk 8 identificeren we obstakels voor
de implementatie van beschikbare genetische kennis binnen de huisartsgeneeskunde
zoals tekortkomingen in het ontwerp en de interface van HIS systemen om genetisch
relevante informatie op te slaan. We introduceren een gefaseerde ‘roadmap’ inclusief
aanpassingen op het HIS en bestaande codesystemen om integratie van genetica
binnen de huisartsgeneeskunde en klinisch onderzoek te verbeteren. Deze roadmap
kan gebruikt worden als een voorbeeld voor het introduceren van andere complexe
toevoegingen en veranderingen aan het HIS of aan codesystemen, altijd vanuit een
behoefte in de dagelijkse medische praktijk.
182
10
Ned
erla
Nd
se sam
eNv
at
tiNg
Lessons LearnedWanneer we de resultaten van de uitgevoerde studies samenvatten en interpreteren
kunnen we de volgende lessen daaruit halen:
1. De kwaliteit van data binnen de huisartsgeneeskunde is suboptimaal, zelfs voor een
sleutel-item zoals de gecodeerde diagnose en voor een ernstige ziekte zoals kanker;
relevante informatie betreffende risicofactoren zoals familie-anamnese mist vaak, is
onvoldoende gecodeerd of kan niet makkelijk worden gevonden;
2. Ondanks suboptimale data-kwaliteit en daaruit volgende herbruikbaarheid met
duidelijke restricties, is het huisartsgeneeskundig medisch dossier een rijke en
volumineuze bron van (voornamelijk ongecodeerde) medische data, die vaak vele
jaren van ‘follow-up’ bestrijkt;
3. Doordat de kwaliteit van data suboptimaal is, zou data uit het huisartsgeneeskundig
medisch dossier alleen moeten worden hergebruikt door mensen die de context van
de huisarts en dus de context rond invoer van deze routine-zorg-gegevens begrijpen
en die expliciet rekening houden met de beperkingen van deze data;
4. Het is noodzakelijk om de kwaliteit van data in het HIS te verbeteren omdat het
hergebruik en het delen van deze data op zich wenselijk is en ook zal toenemen,
idealiter aan de bron (bij data invoer) en ondersteund door adequate codesystemen.
Huisartsen kunnen en moeten hierin worden gefaciliteerd op een aantal manieren
(zie aanbevelingen);
5. Adequate en up-to-date codesystemen zijn cruciaal voor data hergebruik en delen,
niet alleen voor alledaagse maar ook voor zeldzame ziektes en kunnen succesvol
worden ontwikkeld in een samenwerking tussen codeerinstanties en clinici, hierbij
gefaciliteerd door software tools;
6. Deze codesystemen zijn waardevol wanneer ze continue worden onderhouden,
wanneer ze voorzien zijn van adequate synoniemen, relevante crosslinks naar andere
systemen en een duidelijke handleiding voor gebruik;
7. Verplicht coderen van velden in het huisartsgeneeskundig dossier zorgt voor een
meer complete registratie maar ook tot over-registratie en fouten;
8. Het 1-op-1 linken van huisartsgeneeskundige dossiers aan andere data bronnen kan
waardevol zijn om diagnoses te valideren maar is op dit moment nog een complex
en tijdrovend proces;
9. Het huisartsgeneeskundig medisch dossier kan complementair zijn aan andere data
bronnen, zelfs aan een betrouwbare referentiestandaard zoals de Nederlandse Kanker
Registratie (NKR);
10. Er zijn veel ‘stakeholders’ betrokken wanneer het gaat om hergebruik en delen
van medische informatie uit huisartsendossiers: patiënten, huisartsen, het NHG
(Nederlands Huisartsen Genootschap), leveranciers van HISsen (Huisartsen Informatie
Systeem), verzekeringsmaatschappijen/inspectie voor de gezondheidszorg (kwaliteit
183
10
Ned
erla
Nd
se sam
eNv
at
tiNg
beoordelaars), ziekenhuizen, huisartsenposten, onderzoekers, academische
onderzoeksnetwerken, huisartsenopleidingen, codeerinstanties en eigenaars van
externe data bronnen zoals het NKR.
Aanbevelingen De lessons learned kunnen vertaald worden in aanbevelingen voor de diverse stakeholders
(samenvatting, zie voor uitwerking de Engels tekst);
Patiënten
y Patiënten zouden bewust gemaakt moeten worden van het huidige anonieme en
niet-anonieme hergebruik van hun huisartsgeneeskundig medisch dossier.
y Patiënten zouden gestimuleerd moeten worden om verantwoordelijkheid te nemen
door te controleren of al hun belangrijke diagnoses en informatie mbt allergieën en
intoleranties bekend zijn bij hun huisarts en daar ook zijn geregistreerd.
Huisartsen
y Huisartsen zouden de kwaliteit van hun data moeten verbeteren door te investeren
in het updaten en coderen sleutel-items in hun dossiers en door het optimaliseren
van werkprocessen rond registratie.
y Wij bevelen aan dat huisartsen, die hun data kwaliteit hebben verbeterd, actief
participeren in relevante projecten om hun data te delen en te hergebruiken.
Nederlands Huisartsen Genootschap (NHG)
y Het NHG zou aan de ene kant de verbetering van kwaliteit van data in het
huisartsgeneeskundig dossier moeten nastreven en faciliteren, maar aan de andere
kant het ongelimiteerde hergebruik en delen van deze data tegen moeten gaan.
y Het NHG zou digitale technieken die een hoge kwaliteit van data bij invoer kunnen
faciliteren, zoals spraakherkenning, natuurlijke taalherkenning en tekst-mining tools,
actief moeten onderzoeken.
y De implementatie van nieuwere versies van het HIS referentiemodel door leveranciers
zou gestimuleerd moeten worden en diverse aanpassingen zouden moeten worden
gedaan aan het referentiemodel en de ADEPD richtlijn omdat deze bij kunnen dragen
aan het verhogen van de data kwaliteit.
HIS leveranciers
y Onderzoek hoe userinterface en systeem ontwerp aangepast kunnen worden zodat
hoge kwaliteit van (gecodeerde) data bij invoer wordt gestimuleerd.
y Integreer richtlijnen in het HIS zodat digitale suggesties kunnen worden gedaan aan
de huisarts en faciliteer feedback aan huisartsen via eenvoudig te bouwen zoekvragen
en selectie-queries.
184
10
Ned
erla
Nd
se sam
eNv
at
tiNg
Kwaliteitsbeoordelaars
y Stop het meten van de kwaliteit van registratie en vind manieren om de kwaliteit
van zorg adequaat en zonder verstoring van het zorgproces te meten in een dialoog
met huisartsen.
Opleiders
y Voeg een vak “adequate dossiervoering” toe aan het basiscurriculum en leer huisartsen
en huisartsen in opleiding de benodigde vaardigheden.
Onderzoekers en Academische onderzoeksnetwerken
y Als er gewerkt wordt met routine zorg gegevens, valideer dan de diagnose.
y Onderzoeksnetwerken zouden huisartsen actief moeten stimuleren om de kwaliteit van
data in het huisartsgeneeskundig dossier te verbeten en zouden concreet hierbij kunnen
ondersteunen door (spiegel) informatie terug te koppelen aan participerende huisartsen.
Suggesties voor verder onderzoekIn dit proefschrift menen we succesvol te zijn geweest in het beoordelen van de data
kwaliteit van bepaalde items in het huisartsgeneeskundig dossier en zijn we ook in staat
geweest om strategieën en oplossingen te bedenken waarmee data kwaliteit kan worden
verbeterd zodat het hergebruik en het delen van data mogelijk kan worden gemaakt.
We realiseren ons dat dit slechts stukjes van een grote puzzel zijn die de komende jaren
zal moeten worden gemaakt.
Wij zijn er van overtuigd dat we deze resultaten hebben kunnen behalen en
aanbevelingen hebben kunnen formuleren omdat we over de grenzen hebben gekeken
van academische disciplines: huisartsgeneeskunde, klinische genetica, medische
informatica en bio-informatica. Juist het samenwerken met wetenschappers uit andere
disciplines brengt inzicht en oplossingen voor onderzoeksvragen.
Er zijn een aantal (interdisciplinaire) onderzoeks-uitdagingen op dit gebied voor de
toekomst. Allereerst zouden, naast compleetheid en correctheid, ook andere dimensies
van data kwaliteit moeten worden bestudeerd: concordantie, geloofwaardigheid en
tijdigheid, niet alleen voor diagnose registratie maar ook voor andere sleutel-items
zoals risico factoren, doorgemaakte behandelingen en allergieën en intoleranties. In de
tweede plaats zouden interactie-designers moeten worden betrokken bij deze studies:
welke aspecten van de userinterfaces van de verschillende HISsen leiden tot de door
ons geconstateerde verschillen in data kwaliteit en hoe zouden userinterfaces kunnen
worden aangepast zodat de data kwaliteit verbetert?
In de derde plaats zouden het ontwerp, de implementatie en de evaluatie van
diverse interventies, zoals genoemd in dit hoofdstuk (door middel van studies) in de
huisartsenpraktijk zelf kunnen bijdragen aan het ontwikkelen van effectieve interventies
om de data kwaliteit te verbeteren.
185
10
Ned
erla
Nd
se sam
eNv
at
tiNg
Ten vierde zouden de huidige mogelijkheden van natuurlijke taal herkenning en
het realiseren van digitale codeersuggesties tijdens data invoer verder moeten worden
onderzocht.
Tenslotte is het noodzakelijk om ervaring op te doen met door de patiënt ingevoerde
gegevens om de bruikbaarheid van deze data te bepalen voor diverse doeleinden,
natuurlijk in de eerste plaats de zorg zelf. Dit kan bijvoorbeeld gedaan worden door
de ontwikkeling van een App of software tool waarmee patiënten relevante medische
familie informatie kunnen invoeren.
Over onze patiënt uit de inleidingJanuari 2020: een 48-jarige man bezoekt zijn huisarts met een persisterende slijm-
producerende hoest. Enkele dagen geleden maakte hij online een afspraak en voerde
daarbij de reden voor het consult en zijn klachten in. Net voor de afspraak leest de huisarts
deze informatie en kijkt gelijk in de persoonlijke gezondheidsinformatie van de patiënt,
waarin ook gegevens worden verzameld vanuit diverse Apps die de patiënt gebruikt. Het
valt haar op dat de patiënt wat gewicht heeft verloren maar ook dat de trainingsfrequentie
en duur van deze hardloopfanaat behoorlijk zijn afgenomen de afgelopen 4 weken.
De huisarts stelt tijdens het consult een aantal aanvullende vragen en doet lichamelijk
onderzoek, wat zonder afwijkingen blijkt te zijn. Ze vat haar bevindingen hardop samen
voor de patiënt en voor de registratie en gebruikt daarbij haar spraakherkenningssoftware.
Tijdens het inspreken kiest ze relevante codes voor symptomen, klachten en de differentiaal
diagnose, grotendeels gesuggereerd door het systeem. Op haar scherm verschijnt een
pop-up, gebaseerd op de in het systeem geïntegreerde NHG richtlijn “Acuut Hoesten”,
met de vraag of er een aanvraag moet worden verstuurd voor een röntgenonderzoek
(X-Thorax) naar het dichtstbijzijnde ziekenhuis met de kortste wachttijd en vergoeding
door de verzekeraar van de patiënt. Ze klikt op “akkoord” en maakt een afspraak met
de patiënt voor een consult de week daarop om de uitslagen te bespreken.
Dit proefschrift draagt hopelijk bij aan de verbetering van digitale gegevens in medische
dossiers in het algemeen en daarbij aan de ware goudmijn die deze gegevens kunnen
zijn, met als uiteindelijk doel de zorg te verbeteren voor patiënten met alledaagse en
zeldzame ziektes.
Noten# fout-negatieven zijn gevallen van kanker die wel in de NKR staan maar niet in het
huisartsgeneeskundig dossier
fout-positieven zijn gevallen die wel geregistreerd staan in het huisartsgeneeskundig
dossier maar niet in de NKR
data
per
cancerCare
EHRbreast
quality
codi
ng
patients
GP
alsostudy women
used
GPs
12
et
4
al3
5
eg
10
Primary
system
cases
fam
ilyris
kreuse
diagnosis
code
d
using
clinical
code
first
one
gene
tic
two
user
s
age
medical
codes
diag
nose
s
EMR
diseases
registryresearchinform
ation
systemsm
atch
ing
gene
ral
dise
ase
registered
Netherlands
patient
found
Dutc
hyears
practice
history
number
results
NCR
metabolic
heal
th
rare
table
year
registratio
n
avai
labl
e
matches
based
database
genetics
sharing
text
improve
use
however
ICPC
rout
ine
chapter
recordsvalues
listconcerns
disorders
SORTA
University
existing
consultatio
nFigure
example
relevant
value
referral
sym
ptom
s
reco
rd
new
including
Reference
Ontology
related
incidence
SNOM
EDCT
Since
miss
ing
knowledge
purp
oses
incr
ease
d
Episode
consultatio
ns
popu
latio
n
date
step
match
time
studies
national
three
False
posit
ive
CINEAS
2011
case
correct
Furthermorest
anda
rd
types
different
HPO
1000
electronic
know
man
agem
ent
centre
JGPN
Utrecht
physical
Practitioners
Published
inte
rnat
iona
l
analysis
proc
ess
importanton
line
Lucene
according
assess
deve
lope
d
extracted
well
means
referred
complaints
practices
poss
ible
present
Literature
Medicine
guideline
with
out
guidelines
compared
positive
simila
rity
Linkage
large
common
completeness
Groningen
words
Terms
thesis
van
provide
potentially
Many
need
development
shows
improved
matched
screening
Human
Classification
register
determine
instance
EHRs
complete
group
performed
differences
field
hospital
design
oftento
tal
189
10
Da
nk
wo
or
D
DankWooRD
Uiteindelijk typ ik dit dankwoord in het vliegtuig van Marseille naar Amsterdam. Op
weg voor een paar dagen Nederland, waar ik inmiddels dakloos ben omdat ons huis is
verhuurd! Vier weken geleden verhuisden Adri en ik met onze jongste twee kinderen
naar Zuid-Frankrijk om een aantal jaar te wonen en werken in de Provence: een (zeer
fijne) nieuwe omgeving om nieuwe ervaringen op te doen en een andere manier van
leven te ontdekken.
De afgelopen jaren heb ik regelmatig gedacht aan het moment dat ik eindelijk het
dankwoord van mijn proefschrift zou kunnen schrijven: het teken dat het grootste
project wat ik ooit heb aangepakt zou zijn afgerond. Op welke momenten je daar als
promovenda aan denkt? Nou, om wat voorbeelden te noemen: dat zijn de momenten
dat je al uren en uren alleen achter je laptop zit te ploeteren met SPSS of Excel en dat
het lijkt of je in de cijfers verzuipt… Die keren dat je een artikel waar je je uiterste best
op hebt gedaan en waar je heel trots op bent, na twee dagen weer in je inbox vindt
omdat het is afgewezen! De fase dat het lijkt alsof er nooit een eind aan het project
gaat komen, je man en kinderen geïrriteerd raken door je afwezige gedrag en dat je
jezelf afvraagt waarom je hier ooit aan bent begonnen? Maar dan komt er weer een
mail binnen van je promotor met steevast onderaan de tekst “hou vol!”. En dan gaat
je copromotor er eens goed voor zitten, geeft advies, regelt studenten (dank Jessika
Roskam en Rosanne Ader!) die komen helpen met het taaiste werk en wordt het ene na
het andere artikel wél geaccepteerd. Kortom een project met bergen en dalen, net als
onze nieuwe leefomgeving de Provence, en net als het leven zelf.
Dank Mattijs Numans, mijn promotor, voor de kans die je me bood in 2011 om dit
project op te starten en de vrijheid om het zelf vorm te geven. Dank voor je inspiratie
en je “hou vol” mailtjes op de juiste momenten vanuit Starbucksen over de hele
wereld maar bij voorkeur uit de VS. Dank Rolf Sijmons, ook promotor, voor je morele
ondersteuning, je humor, enthousiasme en inzichten. Ik waardeer het enorm dat je
gewoon naast me kwam zitten achter de computer en stukken niet lopende tekst
samen met mij herschreef. Welke promotor doet dat nou? Charles Helsper, merci voor
je inzet en je concrete hulp op lastige momenten. Dank voor het investeren van tijd
en het bellen, ook na werktijd, net voor een deadline met twee kleine jongetjes die
aan je broek hangen!
Leden van de leescommissie, de professoren Damoiseaux, Van der Horst, Cornel
en Brinkkemper en doctor Cornet, dank voor de bestuderen van mijn proefschrift en
vooral voor het goedkeuren daarvan. Beste Sjaak, wat bijzonder dat we elkaar hier
weer treffen, meer dan 20 jaar geleden studeerde ik af bij jou in de Informatica aan de
Universiteit Twente.
190
10
Da
nk
wo
or
D
Paranimfen en vriendinnen Ilya en Floor, de één ken ik al lang, de ander nog maar een
paar jaar, maar jullie zijn mij allebei zeer dierbaar. Jullie weten uit eigen ervaring wat een
promotie-traject inhoudt en hebben mijn gemopper aangehoord en me tips gegeven
en gemotiveerd wanneer dat nodig was. Ik word nu al iets minder gestresst bij het idee
dat jullie naast me staan op 27 januari 2017!
Mijn vriendinnen en vrienden, natuurlijk, allemaal verschillend, vanuit verschillende
activiteiten of rollen die je zo hebt in het leven. Gerreke, wij delen ervaringen, niet allemaal
even leuk, maar begrijpen elkaar daardoor met een enkel woord. Merci voor je steun
en motivatie in allerlei situaties. Anja, o wat mis ik ons trim-zwem- & bijpraatavondje op
maandagavond. Dank voor je vrolijkheid en grenzeloze optimisme waar ik me aan op
kan trekken! Christi, schoonzus en (hoe bijzonder) ook vriendin, dank voor je interesse
en meeleven. Margriet, mijn bijna-collega in Frankrijk (hoe jammer), kom gewoon over
een paar jaar met Bram en Ellen deze kant op, gaan we onderzoek opzetten bij sporters
die de Ventoux op fietsen! Madelon en Isa, congres- en reismaatjes, dank voor jullie
meedenken en discussiëren maar vooral ook jullie gezelligheid op reis. Jullie waren voor
mij de collega’s die ik eigenlijk niet had omdat ik tussen drie universiteiten rondzworf.
Emmy en Hugo, ook al zien we elkaar minder dan toen we nog om de hoek woonden,
onze warme vriendschap blijft! Richard, eindelijk ondernemer, dank voor je vriendschap,
optimisme en meeleven privé en zakelijk, al vanaf onze studietijd in Twente. Ik ga ervan
uit dat we elkaar gewoon regelmatig in Frankrijk zien, of in Morzine, of in Mormoiron.
Steven en Marion, laten we vooral samen blijven skiën, met Floor en ons gezin! Ik
waardeer jullie meeleven en betrokkenheid. Vrienden en vriendinnen die ik hier nog niet
heb genoemd; dank voor jullie vriendschap en meeleven.
Mijn collega’s uit de redactie van Huisarts & Wetenschap (H&W): Just Eekhof, Hans
Hawkeye van der Wouden, Bèr Pleumeekers, Lidewij Broekhuizen, Sjoerd Hobma, Wim
Verstappen, Marianne Dees, Henny Helsloot, Susan Umans, Marissa Scherptong-Engbers,
Nadine Rasenberg en Ivo Smeele; ik wil jullie hier zeker noemen want hoeveel makkelijker
is het voor mij geworden door alle dingen die ik leer in de redactie op het gebied van
wetenschap & onderzoek. Dank voor inzichten, kritische noten maar vooral veel humor.
Ook mijn huisartsopleiders wil ik niet ongenoemd laten, want zij hebben soms
last gehad van mijn promotiewerkzaamheden naast de opleiding maar hebben daar
nooit moeilijk over gedaan. Bert ter Horst, Mieke van Dillen en Wietze Eizenga, dank
voor jullie flexibiliteit, motivatie en het voorbeeld wat jullie voor mij zijn als huisarts.
Wietze vooral, jij begrijpt hoe onderzoek werkt, dank voor onze gesprekken en dank
voor het vertrouwen wat je uitte: ik denk dat jij me net het zetje hebt gegeven om de
Frankrijk plannen door te zetten. Ik verwacht je zeker hier met Brigitte om nog eens de
Ventoux op te fietsen met je retro racefiets maar dan bij wat koeler weer (of gewoon
wat vroeger op de dag).
191
10
Da
nk
wo
or
D
Gepke Visser, metabool kinderarts, wat een geploeter was het he, die metabole
ziektecodes. Het werd mijn eerste geaccepteerde artikel, hoofdstuk 6, dank voor het
samenwerken, soms in het WKZ, soms aan de koffietafel bij jou thuis! Dick Lindhout,
klinisch geneticus en emeritus hoogleraar, door jou begon de liefde voor de genetica,
dank voor de kansen die je me jaren geleden alweer bood! Morris Swertz en Chao Pang,
dank voor de samenwerking die geleid heeft tot hoofdstuk 7.
En dan wil ik hier ook mijn Franse collega’s noemen die mij zo hartelijk hebben
ontvangen en geduldig begeleiden. Ik voel me meer dan welkom. Sébastien Adnot, Lies
de Vos, Francis en Els van der Velden en Philippe Morvan de la Maison de Santé Bel Air
à Carpentras: merci pour votre bienvenue à Carpentras et votre accompagnement dans
les semaines passées. Je suis sûr que nous allons collaborer bien!
Pa en ma, dank voor de veilige, liefdevolle en stevige basis die jullie mij en Henk en Gerard
in het leven hebben gegeven. Ik weet dat ik soms rechts ga waar jullie links hadden
gekozen maar weet dat ik de basis niet vergeet en ook waardeer. Henk en Gerard, mijn
broers, we zien elkaar niet zo vaak maar als dat wel zo is, is het altijd vertrouwd en
gezellig. Nel, schoonma, je voelt net zo close als een “gewone” moeder, dank dat je er
altijd voor mij en mijn gezin bent.
Adri, wij zijn niet zo van veel mooie woorden en dat ga ik hier maar niet veranderen
al heb ik wel iets te zeggen natuurlijk. Geen woorden maar daden kenmerkt jou: dank
voor al je inzet en liefdevolle opvang thuis als ik druk was, dank dat je me altijd vrij laat
in mijn keuzes en motiveert om de dingen te doen die me interesseren. Nog steeds ben
ik gefascineerd door al die dingen die jij kunt bedenken en vervolgens ook kunt maken
(en ik niet). Wat een geluk dat jij met mij het leven wilt delen.
Kristel, het is echt leuk om te zien hoe jij je hebt ontpopt tot (bijna) verloskundige,
een mooie, vrolijke en levenslustige vrouw. Hoe leuk is het dat je me soms belt laat op
de avond om te vertellen over een indrukwekkende bevalling en dat ik na een dienst jou
kan vertellen over wat ik nou weer heb meegemaakt. Je gaat nu zelfs onderzoek doen
op Curaçao en werkt voor een onderzoeksproject in het AMC (yes, research!). Dank
voor je nuchtere kijk op het leven en ook op dit promotietraject en je aanstekelijke lach.
Nathalie, toekomstig collega dokter (huisarts?) en wereldburger. Ik heb het geluk
dat ik ook hier kan schrijven: hoe leuk als je dochter je belt om te vertellen over een
college met een indrukwekkende patiënt! Met jou kan ik altijd discussiëren, jij leest mijn
zelfgeschreven stukken en corrigeert ze. Merci voor je open blik, je brede interesse, je
meedenken en scherpe inzichten. Ik ben benieuwd waar je avonturen je naar toe gaan
leiden de komende jaren, maar ik heb er alle vertrouwen in!
Mike, stoere, slimme en sportieve “gast”, in sommige dingen lijken wij nogal op
elkaar. Zoals jij in een boek of in Duckstad kunt zitten, zo kan ik dat ook. Zoals jij “domme
en verstrooide” dingen kunt doen, kan ik dat ook (slagboompje?). Ik hoop dat je ook
192
10
Da
nk
wo
or
D
nog wat goede genen hebt meegekregen van mij, maar anders zijn er genoeg van papa,
daar ben ik zeker van. Als je niet met Jesse een bedrijf begint later (paintballhal?), is
onderzoeker misschien iets voor je?
Marilyn, jij bent vrolijk, expressief en heel sportief, maar vooral echt een lieverdje. Ik
kan me nog goed herinneren dat ik (weer) een afwijzing kreeg van een wetenschappelijk
blad en dat er even later een briefje op mijn laptop lag: “ lieve Annet, ik vind je artikel
WEL goed!”. Jij vond het vaak niet leuk als ik alweer weg moest of alweer achter de
laptop moest kruipen. Ook al ben je nog maar 11 (bijna 12) je begreep het wel. Dank je
en houd vooral niet op met mij duidelijk zeggen als ik teveel met andere dingen bezig
ben. Ik ben heel benieuwd hoe jij je verder gaat ontwikkelen, wie weet wel tot dierenarts.
Annet, november 2016
195
10
CU
RR
ICU
LUM
VITA
E
CuRRICuLum vItaE
Annet Sollie was born on Friday the 13th of March 1970 in Zwolle, but how lucky can
a person be. She is happily married to her soulmate Adri Wisse and together they are
blessed with four wonderful children: Kristel, Nathalie, Mike and Marilyn. Besides this
she got to be a doctor after all. She enjoys using her past education and experience to
“make ICT solutions actually work for the doctor”.
Currently she is working in Carpentras at the Maison de Santé Bel Air in the south
of France as a general practitioner (médecin généraliste). She also works as an editor
for Huisarts & Wetenschap (H&W) and occasionally as a consultant on ICT & Healthcare
projects for her own company Soll-Vite.
Professional skills and interestsRare diseases in primary care
Phenotype coding & coding systems
Medical Genetics in primary care
Electronic Health Record systems
Working experienceFor more information on finished courses, publications, projects and working experience
in the ICT sector please visit:
Linkedin: nl.linkedin.com/in/annetsollie
Researchgate: researchgate.net/profile/Annet_Sollie
Twitter: twitter.com/annetsollie
About: about.me/asoll
Education1992 - 1988 Secondary education (VWO) prof. dr. Greijdanusschool in Zwolle
1988 - 1994 Computer Science, Technical University of Twente, Enschede
specialisation in Business Administration
1999 - 2000 Propaedeutic Psychology, Open University
2001 - 2009 Medical School, University of Utrecht
2012 – 2016 Residency General Practice