Genome Directed Cancer Treatment Use Case for a Learning ...•Genome directed cancer treatment is a driving use case for learning cancer systems ... • Jeff Sosman • Leora Horn

Vanderbilt-Ingram Cancer Center

Genome Directed Cancer Treatment

Use Case for

a Learning Cancer System

March 23, 2012

Mia Levy, MD, PhD Director Cancer Clinical Informatics, Vanderbilt Ingram Cancer Center

Assistant Professor of Biomedical Informatics and Medicine


Biomarkers in the Clinical Continuum

Diagnosis Treatment Selection

Treatment Plan

Management

Treatment Response

Assessment

Risk Biomarker

Diagnostic Biomarker

Prognostic Biomarker

Predictive Biomarker

Response Biomarker


Personalized Cancer Medicine Initiative


Treatment Plan

Management

Treatment Response

Assessment

Predictive Biomarker

Genome directed cancer treatment selection


Traditional View of Cancer

Adeno-

carcinoma

Squamous

Large

Small

Lung Cancer

Arising from Skin

Without Chronic

Sun Damage

Arising from Skin

With Chronic

Sun Damage

Arising from

Mucosal

Surfaces

Arising from

Acral

Surfaces

Melanoma


Lung Panel: 451 patients 46% Patients with Actionable Mutation

54% No Mutation Identified

7/1/10-12/31/11


Personalized Cancer Medicine Initiative

12 ALK fusions

Melanoma Panel: 538 patients 67% Patients with Actionable Mutation

33% No Mutation Identified


Old Method for Reporting Mutation Results

in the Electronic Medical Record

Old Method:

• Report Template

• Scanned into Electronic Health Record as image file (not computable)

Challenges:

• How to report > 40 mutations in 8 genes?

• Whose role to curate knowledge regarding clinical significance?

• Lack clinical trial information


R = Outside Specimen Requested

Order Status (letter)

O = Order Received

A = Outside Specimen Arrived

v = Specimen Accessioned

Yellow = Gene Mutation Detected

Grey = Gene Mutation Not Detected

Red = No Result – Insufficient Specimen

Result Status (colored box)












7 Cancers Lung

Melanoma Breast Colon

Thymic GIST

Thyroid

22 Genes

203 Disease-Gene-Variant Relationships


NEW clinical trial search - 135 Cancer Diagnoses

- 443 Cancer Genes

Vanderbilt-Ingram Cancer Center 53,515

- Mar 15, 2012

120 countries and territories

>1500 site visits per week

Vanderbilt-Ingram Cancer Center 30,308

- Mar 15, 2012


Worldwide Collaboration

• 30 Contributors

• 13 Institutions

• 6 Countries


Scale, Maintain & Sustain

My Cancer Genome

Content

Generation

Content

Dissemination


Decision Support as a Service

Treatment Plan

Selection

Laboratory Testing Facility

Oncology Vendor EHR

Vanderbilt EHR

Public Access

Academic Medical

Center EHR


Learning Cancer System

Select patient

treatment

Assess clinical

outcomes

Compare treatment

effectiveness Implement new evidence for

treatment prioritization

patient

Scalability: Data Driven Approach


Learning Cancer System


Treatment Plan

Management

Treatment Response

Assessment

Aggregate & Analyze

- Primary Site

- Histology

- Stage

- Biomarkers

- Treatment History - Tumor Response

- Host Response

- Quality of Life

- Toxicity

- Survival


Case Reports: Manual Annotation

DIRECT

• Collection of EGFR mutations in NSCLC

• 1596 patient level case reports

• 1876 gene, drug, response instances

• 146 publications

• 150 unique primary EGFR mutations

• 47 unique secondary EGFR mutations

L Horn, H Chen, CM Lovly, J Andrews, P Yeh, MA Levy, W Pao

DIRECT: DNA-mutation Inventory to Refine and Enhance Cancer Treatment. A catalogue of clinically relevant

somatic mutations in lung cancer. J Clinical Oncology 29:2011 (Suppl; abstr 7575)


Case Reports: Automated Extraction

Synthetic Derivative

Vanderbilt EHR

De-identification

Labs, notes,

medications

1.8M pt

>1000 pt with tumor gene mutation analysis

And growing Tumor

Registry

63K pt

Site, Stage, histology,

vital status

Continuous extraction

and integration with

knowledge resources


Chemotherapy Plan Abstraction Method

Medication Event

Extraction

Chemotherapy Plan

Abstraction

Cohort Plan

Analysis


EHR

Data Sources

Physician Notes (100s)

Physician Orders (100s- 1000s)

Pharmacy

Dispensing

Records (100s – 1000s)

Nurse

Administration

Records (100s – 1000s)

30

Data Quality: Extracting Medication Events T

ruth

fuln

ess –

Even

t A

ccura

cy

Free Text Structured

Data Accuracy


Data Completeness

Friedman 2010


Triangulation of Data Sources

Notes Orders Pharmacy Nursing Data Sources

Data Extraction

Methods

Medication Event X(t)

NLP Machine Learning




Transaction System Data Processing

Terminology

Mapping

Probability of Medication Event X(t)

Event

Triangulation


100%

Structured and free-text Format Chemotherapy Medication Events at VUMC

33

% a

t V

ander

bil

t

Clinical Notes

95%

5%

Orders

100%

5%

95%

Pharmacy

Dispensing

Records

Nurse Admin

Records

Free-text

Structured

Data source


Pharmacy Dispensing Records

Jan. 2006 – May 2011

63K patients

Tumor Registry

Sep. 1937 – Dec. 2009

Pharmacy and Tumor Registry Data

5,394

34

` Patients

Medication

Events

487K 17M

Chemo 8,278 213K


Chemotherapy Plan Abstraction Method

Medication Event

Extraction

Chemotherapy Plan

Abstraction

Cohort Plan

Analysis

Recall

Precision

Accuracy

89.9

75.5

69.6

Bhatia, Levy, AMIA 2011


Cohort analysis

36

Cyclophosphamide,

Doxorubicin

Paclitaxel

Trastuzumab

Fulvestrant

Cyclophosphamide,

Docetaxel

Patient

Count

198

138

61

55

48

Breast Cancer Patient

Cohort

n=554

Unique Plans = 107


Cohort analysis

37

Breast Cancer Patient

Cohort

n=554

Unique Plans = 107

Patient

Count

Cyclophosphamide,

Doxorubicin 198

Paclitaxel 138

Periodicity (days) Cycles

Standard of Care Observed Average

15.8

3.8

10.6 8

14

7 12

4

14 4


Therapy Sequence

38

Mamounas, E.P. et al. Paclitaxel After Doxorubicin Plus Cyclophosphamide As Adjuvant Chemotherapy for

Node-Positive Breast Cancer: Results From NSABP B-28. JCO (2005).


Duration of Treatment

as Surrogate of Time to Progression

Fulvestrant (55 Patients)

6.4

39

Robertson, J.F.R. et al. Fulvestrant versus anastrozole for the treatment of advanced breast carcinoma in

postmenopausal women. Cancer 98, 229-238 (2003).


Summary

• Genome directed cancer treatment is a

driving use case for learning cancer systems

• EMR data may be used for such a system

• EMR data quality may be mitigated through

triangulation of multiple sources


Acknowledgements

• William Pao

• Jennifer Pietenpol

• Ashley Lamb

• Christine Micheel

• Cindy Vnencak-Jones

• Christine Lovly

• Jeff Sosman

• Leora Horn

• Paul Yeh

• Mary Beth Bauer

• Adeola Davis

• Haresh Bhatia

• Scott Sobecki

• Riyad Naser

• Mik Cantrell

• Stacy Cooreman

• Jim Tibbetts

• Daniel Carbone

• Ross Oreto

• Anna Belle Leiserson

• Carlos Hooper

• Dario Giuse

• Jonathan Grande

• RuAnn Schleicher

• Jay Cowan

• Michael Assink

Grant Support

• Kleberg Foundation

• Anonymous Foundation

• BMS educational grant

• ONC/NCI Health 2.0 Developer Challenge


Thank you


Method Performance

Methodology Patient

Level Plans

43

Pattern

Recognition

Refined

Plans

KB

Manually annotated

plans used as

Gold Standard

Training Test1

Patients

Med. events

163 341

2,298 5,713

Recall 88.8 91.3

Precision 75.2 82.9

% %

Accuracy 68.7 76.8

Training Test1

Test2

168

3,214

89.9

75.5

%

69.6

Test2

Patients - 7,805

Med. events – 139,659

Documents

Genome Directed Cancer Treatment Use Case for a Learning ...•Genome directed cancer treatment is a driving use case for learning cancer systems ... • Jeff Sosman • Leora Horn