Upload
karin-verspoor
View
64
Download
0
Tags:
Embed Size (px)
DESCRIPTION
Background Pain is a feature of approximately 70% of all Emergency Department (ED) presentations. It has been demonstrated that mandating recording of a patient’s feeling of pain can improve service delivery for ED patients. However, there is a substantial group of patients (approximately 21% of ED visits in our 12-month sample) for which there exists an inconsistency between pain score and the Australian Triage Scale (ATS) score assigned by the nurse; where a patient reports high levels of pain but they are assigned a lower-urgency triage category. It has been unclear until now whether this “inconsistent” group of patients has been receiving optimal care. Methods To better understand the characteristics in this inconsistent group, we performed topic modeling of the clinical notes collected during ED triage assessments. We divided the notes into two subgroups, according to whether or not the patient’s self-reported level of pain was consistent with the triage urgency recorded in the ATS score. We performed topic modeling of these two subgroups separately, using the implementation of Latent Dirichlet Allocation (LDA) in the Mallet toolkit. We have experimented with several representations of the notes, including unigrams (tokens), bigrams, and the medical concepts contained in each note, as determined with the MetaMap medical concept recognition tool. An ED nurse reviewed the topics generated in each case and assigned a descriptor to them. Results When considering the token-based presentation of the notes, the labels in the consistent group are related to road trauma, cardiac pain, change of consciousness, ongoing chest pain, limb injury, renal illness and pain due to illness. In the inconsistent group, we find topics related to either conditions related to ongoing conditions (including postoperative complications or worsening abdominal pain), urinary and respiratory problems, infections and injury related complications. When considering the concept-based representation of the notes, the labels in the consistent set denote gastrointestinal diseases, neurological illness, dizziness, chest pain, testicular pain, shortness of breath and trauma. The labels in the inconsistent set denote different issues caused by trauma and distress due to pain, infection and urinary condition. This includes injuries in several body parts like in the limbs and back. The latter topic containing body parts appears to have been enabled by the abstraction of individual terms into concepts. Conclusions Topic modeling of Emergency Department data shows substantial promise for helping to characterise particular subpopulations of interest, and incorporating pre-processing of clinical notes to capture variation in clinical terminology appears to have value. While this initial work has focused on the pain-related chief complaints, we have also recently begun to explore temporal characteristics of the data through analysis of how derived topics change ove
Citation preview
Topic modeling of Emergency Department Triage notes for characterising pain-related
chief complaintsKarin Verspoor, The University of MelbourneAntonio Jimeno Yepes, The University of MelbourneSimon Kocbek, RMIT UniversityWray Buntine, Monash UniversityTheresa Vassiliou, Royal Melbourne HospitalMarie Gerdtz, Royal Melbourne Hospital
World health organisation (WHO) guidelines on pain management (2007)
Emergency care acute pain management manual (2011)
Australasian College of Emergency Medicine (ACEM) Policy (58) (2009)
Background
Aims of the study• Apply topic modeling to explore the triage
notes for pain-related descriptors• Apply temporal (dynamic) topic modeling
to identify temporal variations in symptoms
Safety
Risk
Acuity & Severity
Presenting Complaint
Time to treatment
Pain Score
ComplexityNeed for admission
Assigned ATS 1 - 5
2 - 5 minute Assessment
Allocated to treatment stream
Triage process
Numerical Rating Score (NRS)
Pain Score at triage
Sample57,984 Patients
ATS 4
(10,655)
ATS 3
(9,200)
ATS 2
(2,572)
Exclusions
Presenting complaints for psychiatric distress
ATS 1 – 5 , “Unable” 0/10
(35,612)
Pain related distress
and ATS 2 – 4
(22,372)
Subgroups of interest
ATS 2 ATS 4
ATS 3ConsistentATS High urgency &
High Pain score
InconsistentATS Low urgency &
High Pain score
Moderate pain (4-6)
Mild pain (1-3)Severe pain (7-10)
Topic Modeling• Unsupervised machine learning task that uncovers the hidden topical patterns in text collections.• Based on a probabilistic model that allows documents to have mixtures of topics.• Topic:
– distribution over terms in a vocabulary. – represented with a list of top most probable words/tokens.
• Our static model: – grouping of the most related tokens in each patient subgroup of interest, using topic modeling– ED records text is split into tokens or mapped to concepts in the UMLS
Topic Modeling
Blei, MLSS 2012
Topics by consistency subgroup
• First token based topic for both sets showing the top 10 most relevant words
• We find that the terms related to the inconsistent groups denote painful but not life threatening conditions
Inconsistent (Topic 0) Consistent (Topic 0)
swelling hr
painful chest
knee hx
swollen bp
arm sob
yesterday reg
shoulder spo
forte central
injury ht
present increased
Dynamic topic models• Capable of analysing
the time evolution of topics.
• Data can be split into epochs (e.g. months, weekdays-weekend)
• First order Markov model: current epoch depends on the previous epoch. Blei, MLSS 2012
Results – Dynamic Model
Topic
Problem
Top representative words
T1 Flu aches, runny, chills, flu-like, fever
T2 Asthma sentences, speaking, ventolin, talk
T3 Angina gtn, patch, anginine, spray, aspirin
T4 Arm foosh, rotation, shortening, rotated
Topic Top representative words
Car car, loc, driver, hit, speed, head
Finger finger, cut, vasc, intact, rom, hand
Abdomen abdo, flank, chronic, lower
Issues and future work
• Preprocessing of the data to address problems with clinical language used in triage notes.• Explore different numbers of topics.• Pain-related topics analysis: Assessment of topic coherence based on nurses’ feedback
• Dynamic topic models: – More data is needed (statistical significance) –Adaptation of topic model to deal with periodic effects
Acknowledgements
Our collaborators at the Royal Melbourne Hospital ED• A/Prof Jonathan Knott• Theresa Vassiliou• Marie Gerdtz• Rochelle Wynne