Building a framework for handling clinical abbreviations a long journey of understanding shortened words "

Transcription

1 Building a framework for handling clinical abbreviations a long journey of understanding shortened words " Yonghui Wu 1 PhD, Joshua C. Denny 2 MD MS, S. Trent Rosenbloom 2 MD MPH, Randolph A. Miller 2 MD, Dario A. Giuse 2 Dr.Ing, Hua Xu 1,2 PhD 1 School of Biomedical Informatics, The University of Texas 2 Department of Biomedical Informatics, Vanderbilt University

2 Abbreviations in Clinical Notes" 53 yo woman with HTN, DM2, obesity, depression/anxiety, H. Pylori gastri7s, GERD, h/o pancrea77s in 2002 presumed 2/2 passed gallstone (lipase 431/amylase 181). Pervasive 17.1% of total word tokens Important 33% abbrevia7ons are diseases/ symptoms Ambiguous % of abbrevia7ons found in the UMLS 2001 were ambiguous (e.g., pt - pa7ent, pt - physical therapy) Dynamic different ins7tu7ons, different types of notes, over the 7me 1

3 A Corpus-based Framework for Handling Clinical Abbreviations" 1 Detect abbreviations from a corpus e.g. if ca is an abbreviation Clinical Corpus 2 Build corpus- specific sense inventories e.g. ca - Carcinoma - Calcium 3 Disambiguate ambiguous abbrevia;ons e.g. Diagnosed with Colon CA a1er ü Carcinoma û Calcium

4 Detect abbreviations using machine learning classifiers" Data set: 70 discharge summaries with abbrevia7ons annotated Features including word forma7on, dic7onary, corpus frequency Different machine learning algorithms including DT, RF, SVM, and ensembles Best performance F- measure 95.7% Wu et al. AMIA Annu Symp Proc Detec7ng abbrevia7ons in discharge summaries using machine learning methods. 3

5 Build sense inventory using clustering" Clinical Notes mg Instance 1 Instance 2 Instance 3 milligram magnesium milligram Senses of mg : - milligram - magnesium Instance Collec;on Feature Vector TCRSD Sense Clustering Sense cluster 1 Sense cluster 2 milligram magnesium Xu et al. J Biomed Inform Dec;45(6): A new clustering method for detec7ng rare senses of abbrevia7ons in clinical notes.

6 Evaluation of clustering-based sense inventory " 13 abbrevia7ons, 52 senses Annotation Cost Average Sense completeness Random EM TCRSD 10 66% 68% 82% 20 69% 75% 85% Xu et al. J Biomed Inform Dec;45(6): A new clustering method for detec7ng rare senses of abbrevia7ons in clinical notes.

7 Profile-based disambiguation using vector space model" Discharge Summaries Sense- tagged Discharge Summaries Sense Profiles Sense Profile Vector 1 Sense Profile Vector 2 Similarity 1 An Instance from Admission Notes Instance Context vector Similarity 2 automatically sense-annotated instances of discharge summaries for each sense of each abb She was ambulating on room air without shortness of breath room air She was ambulating on ra without shortness of breath 6 6

8 Combine profile score and estimated sense frequency from clustering analysis" Sense frequency based on clustering ra right atrium room air rheumatoid arthri7s right atrium room air rheumatoid arthri7s N rheumatoid arthri7s Profile- based similarity scores Evalua7on on 13 abbrevia7ons (200 instances each) showed following accuracy Profile only Sense frequency only Profile + sense frequency Xu et al. Combining Corpus- derived Sense Profiles with Es7mated Frequency Informa7on to Disambiguate Clinical Abbrevia7ons. AMIA Annu Symp Proc. 2012; 2012:

9 Building comprehensive abbreviation sense inventories from Vanderbilt corpus" Abbrevia7on detec7on Sense clustering and annota7on Vanderbilt clinical notes Abbrevia;on List, Pick top 1000 Sense Inventory Notes Abbrevia;on detected Exis7ng database # of clusters for 1000 abbrevia;ons Final # of abbrevia;ons Final # of senses Discharge Summary (123K) 22,546 15, (221) 1,299 Clinic Visit (261K) 107,303 20, (283) 1,499

10 Deliverables " The open source CARD (Clinical Abbrevia7on Recogni7on and Disambigua7on) Framework Abbrevia7on Detec7on List of abbrevia7ons Sense Clustering Sense clusters Sense disambigua7on Sense Inventory Two comprehensive sense inventories: discharge summaries and clinic visit notes Abbrevia7on wrappers for exis7ng NLP systems: MetaMap and ctakes