Effective Feature Representation for Clinical Text Concept Extraction Yifeng Tao 1,2 , Bruno Godefroy 1 , Guillaume Genthial 1 , Christopher Potts 1,3,* 1 Roam Analytics 2 Carnegie Mellon University 3 Stanford University Yifeng Tao et al. NAACL Clinical NLP 2019 1
Background: Healthcare Text Datasets o Crucial information of healthcare recorded only in free-form text Clinical Diagnosis Detection Social Scientific Prescription Reasons Penn Adverse Drug Reactions Chemical-Disease Relations Init Maximization Commercial Drug-Disease Relations Expectation Crowdsourcing Expert annotation FDA Drug Labels convergence End Drug-Disease Relations Dataset [Figures from: 1. Lamjed Ben Jabeur et al. Uprising microblogs: A Bayesian network retrieval model for tweet search. 2012, 2. https://www.sjm.com.br/utilidades/pubmed-busca, 3. http://anakin.uta.cloud/uncategorized/the-need-for-drug-donations, 4. https://www.autismawareness.com.au/news-events/aupdate/is-there-an-over-diagnosis-of-autism] Yifeng Tao et al. NAACL Clinical NLP 2019 2
Background: Healthcare Text Datasets o Clinical text datasets are scarce and expensive o Privacy considerations o Domain specialists 10000 8000 # texts 6000 4000 2000 0 Diagnosis Prescription Penn Adverse Chemical-Disease Drug-Disease Detection Reasons Drug Reactions Relations Relations Yifeng Tao et al. NAACL Clinical NLP 2019 3
Task: Clinical Text Annotation P OSITIVE C ONCERN Diagnosis Detection Asymptomatic bacteriuria , could be neurogenic bladder disorder . P RESCRIBED R EASON Prescription Reasons I will go ahead and place him on Clarinex for his seasonal allergic rhinitis . ADR Penn Adverse Drug Reactions (ADR) #TwoThingsThatDontMixWell venlafaxine and alcohol - you’ll cry and ADR throw chairs at your mom’s BBQ . D ISEASE D RUG Chemical–Disease Relations (CDR) Ocular and auditory toxicity in hemodialyzed patients receiving desferrioxamine . T REATS Drug–Disease Relations Indicated for the management of active rheumatoid arthritis and should not be C ONTRA used for rheumatoid arthritis in pregnant women . Yifeng Tao et al. NAACL Clinical NLP 2019 4
Previous Models OTHER DISCONTINUED REASON REASON OTHER DISCONTINUED REASON REASON CRF CRF CRF CRF CRF CRF CRF CRF LSTM LSTM LSTM LSTM sparse features sparse features sparse features sparse features word word word word embedding embedding embedding embedding Soma Soma Stop Stop cost cost for for o LSTM-CRF o HB-CRF o General text o Clinical text o Distributed word embeddings o Sparse hand-built features Yifeng Tao et al. NAACL Clinical NLP 2019 5
Model: ELMo-LSTM-CRF-HB o Dense ELMo word embeddings + Sparse hand-built features OTHER DISCONTINUED REASON REASON CRF CRF CRF CRF dense features dense features dense features dense features LSTM LSTM LSTM LSTM sparse features sparse features sparse features sparse features ELMo ELMo ELMo ELMo Soma Stop cost for Yifeng Tao et al. NAACL Clinical NLP 2019 6
Performance: Per-token Macro-F1 Scores o Hyperparameters tuned through cross-validation o Each experiment repeated for five times rand-LSTM-CRF HB-CRF ELMo-LSTM-CRF ELMo-LSTM-CRF-HB *** *** 90 *** 80 * F1 Score 70 60 ** 50 40 Diagnosis Prescription Penn Adverse Chemical-Disease Drug-Disease Detection Reasons Drug Reactions Relations Relations *: p <0.05, **: p <0.01, ***: p <0.001 Yifeng Tao et al. NAACL Clinical NLP 2019 7
The Role of Text Length o LSTM: handles short texts well o HB-CRF: robust on long texts Yifeng Tao et al. NAACL Clinical NLP 2019 8
CRF Potential Scores o LSTM features always more OTHER DISCONTINUED REASON REASON important o HB features make substantial dense features dense features dense features dense features contribution LSTM LSTM LSTM LSTM sparse features sparse features sparse features sparse features ELMo ELMo ELMo ELMo Soma Stop cost for Yifeng Tao et al. NAACL Clinical NLP 2019 9
Major Improvements in Minor Categories 100 10 100 10 Prescription Reasons Diagnosis Detection 90 90 8 8 Improvement (%) Imrpovement (%) F1 score (%) F1 score (%) 80 80 6 6 70 70 4 4 60 60 2 2 50 50 40 0 40 0 OTHER POSITIVE RULED-OUT CONCERN OTHER REASON PRESCRIBED DISCONTINUED (74888) (24489) (2797) (2780) (83618) (9114) (5967) (2754) 100 10 100 120 Chemical-Disease Relations Drug-Disease Relations 9 90 90 100 8 Improvement (%) Improvement (%) 7 F1 score (%) F1 score (%) 80 80 80 6 70 5 70 60 4 60 40 60 3 2 50 20 50 1 40 0 40 0 OTHER TREATS UNRELATED PREVENTS OTHER DISEASE CHEMICAL (10634) (3671) (1145) (320) (104530) (6887) (6270) Label/Category (Support) Yifeng Tao et al. NAACL Clinical NLP 2019 10
Conclusion o A unified feature representation for clinical text sequence labeling o Sparse, ontology-driven features o Dense LSTM features o Best performance on five distinct healthcare datasets o Takes advantages of both feature types o Makes maximal use of small, expensive, domain-specific healthcare texts o A new labeled clinical dataset o Identifies the treatment relations between drugs and diseases o Extensive analysis to identify what information our model makes use of, and why its performance is consistently improved Yifeng Tao et al. NAACL Clinical NLP 2019 11
Acknowledgement o Roam Analytics o Christopher Potts o Bruno Godefroy o Guillaume Genthial o Kevin Reschke o NLP Group Yifeng Tao et al. NAACL Clinical NLP 2019 12
Penn Adverse Drug Reactions (ADR) Results Penn Adverse Drug Reactions 100 200 90 Improvement (%) 150 F1 score (%) 80 70 100 60 50 50 40 30 0 OTHER ADR INDICATION (5023) (283) (29) F1 score (%) Improvement (%) o The Role of Text Length o Major Improvements in Minor Categories Yifeng Tao et al. NAACL Clinical NLP 2019 13
Example of Hand-built Features Yifeng Tao et al. NAACL Clinical NLP 2019 14
Procedure for Building Drug-Disease Relations Dataset Init Maximization Expectation Crowdsourcing Expert annotation FDA Drug Labels convergence End Drug-Disease Relations Dataset Yifeng Tao et al. NAACL Clinical NLP 2019 15
Statistics of Datasets Yifeng Tao et al. NAACL Clinical NLP 2019 16
Hyperparameters of Experiments Yifeng Tao et al. NAACL Clinical NLP 2019 17
Recommend
More recommend