madhumita sushil, simon Šuster, kim luyckx, walter daelemans · 2020. 7. 16. · ucto. clin28...
TRANSCRIPT
Unsupervised patient representationswith interpretable classification decisions
Madhumita Sushil, Simon Šuster, Kim Luyckx, Walter Daelemans
CLIN28
Patient Representations Task-independent generalized semantic representations of patients, s.t.
Similar patients - similar representations
Patient similarity encompasses holistic patient condition
2
CLIN28
Patient Similarity
3
CLIN28
Pipeline - Representation Learning and Evaluation
4
CLIN28
Data
5
icons by pictohaven, LAFS, ProSymbols, Lucas Almeida, Juraj Sedlák; thenounproject.com
MIMIC-III adult patient; min. 1 non-discharge note
25k-3k-3k(train-val-test)
1-879 notes, 14 categories
PATIENT DOCUMENT
PATIENTTOKENS
Avg. 13k tokens
UCTO
CLIN28
Unsupervised Representation Learning
6
PATIENTTOKENS
BAG-OF-WORDS (BoW)
MEDICAL CONCEPTS
+ ASSERTION
BAG-OF-CONCEPTS (BoCUI)
SDAE-BoW / SDAE-BoCUI
Vectors
CLAMP
Stacked Denoising Autoencoder
(SDAE)
SDAE
CLIN28
Unsupervised Representation Learning
7
PATIENTTOKENS
DOC2VEC Vectors
DOC2VEC
CLIN28
Unsupervised Representation Learning
8
DOC2VEC Vectors SDAE-BoW
Vectors
SDAE-BoW +
DOC2VECVectors
CLIN28
Supervised Representation Evaluation
9
13% PATIENT DEATH
4% PATIENTDEATH
12% PATIENT DEATH
FEEDFORWARD NEURAL NETWORKS
IN-HOSPITALMORTALITY
30 DAYSMORTALITY
1 YEARMORTALITY
PRIMARYDIAGNOSTICCATEGORY
PRIMARYPROCEDURAL
CATEGORYGENDER
57% MALE40% MAJORITY 39% MAJORITY
Patient Representations
CLIN28
Results - Representation Evaluation
10
All generalized representation models significantly outperform sparse models when no. of positive instances is low (30 days mortality).
CLIN28
Results - Representation Evaluation
11
For all tasks except distant patient mortality, BoW model is a strong baseline (strong lexical features present).
CLIN28
Results - Representation Evaluation
12
Recommended to combine SDAE and doc2vec vectors for unknown tasks.
CLIN28
Model InterpretabilityCritical for
● Error analysis● Exploratory analysis
Goal: Quantitative explanation of a trained model without retraining
13
CLIN28
Feature Extraction - Pretraining AutoencodersRank features according to mean squared feature reconstruction error across all instances
14
CLIN28
Feature Extraction - Pretraining AutoencodersRank features according to mean squared feature reconstruction error across all instances
14
CLIN28
Feature Extraction - Pretraining AutoencodersRank features according to mean squared feature reconstruction error across all instances
0.87-0.88 Spearman correlation coefficient between feature reconstruction error and frequency
Feature entropy too high?
14
CLIN28
Finding Influential features - Classification PhaseGradient-based feature sensitivity calculation across two neural networks
For arbitrary number of instances and output classes
Transferable to different representation learning architectures
15
CLIN28
Feature Extraction - Classification Phase
16
i4
i1
i2r1
i3
r2
r3
r1
r2
r3
Collecting all encoder layers of SDAE
o1
o2
Supervised classifier
CLIN28
Feature Extraction - Classification Phase
16
i4
i1
i2r1
i3
r2
r3
r1
r2
r3
Collecting all encoder layers of SDAE
o1
o2
Supervised classifier
*
So1, i1(1) =
Sensitivity
CLIN28
Feature Extraction - Classification Phase
Feature influence = Maximum root mean squared sensitivity across all instances
16
i4
i1
i2r1
i3
r2
r3
r1
r2
r3
Collecting all encoder layers of SDAE
o1
o2
Supervised classifier
*
So1, i1(1) =
Sensitivity
CLIN28
Most influential features for one instance each
17
CLIN28
ConclusionsGeneralized patient representations help for low number of positive instances
Recommended to combine autoencoder and doc2vec representations
During representation learning, autoencoder has high reconstruction error for frequent terms, perhaps due to their high entropy
The most influential features for classification however are frequent terms, and context of features plays a role.
18
CLIN28
Unsupervised patient representations from clinical notes with interpretable classification decisions
Madhumita Sushil, Simon Šuster, Kim Luyckx, Walter Daelemans
Workshop on Machine Learning for Health, NIPS 2017.
THANK YOU!
19
CLIN28
Results
20