ezDI: A Hybrid CRF and SVM based Model for Detecting and Encoding Disorder Mentions in Clinical Notes

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 278–283, Dublin, Ireland, August 23-24, 2014. ezDI: A Hybrid CRF and SVM based Model for Detecting and Encoding Disorder Mentions in Clinical Notes Parth Pathak, Pinal Patel, Vishal Panchal, Narayan Choudhary, Amrish Patel, Gautam Joshi ezDI, LLC. {parth.p,pinal.p,vishal.p,narayan.c,amrish.p,gautam.j}@ezdi.us Abstract This paper describes the system used in Task-7 (Analysis of Clinical Text) of SemEval-2014 for detecting disorder men- tions and associating them with their re- lated CUI of UMLS1. For Task-A, a CRF based sequencing algorithm was used to find different medical entities and a binary SVM classifier was used to find relation- ship between entities. For Task-B, a dic- tionary look-up algorithm on a customized UMLS-2012 dictionary was used to find relative CUI for a given disorder mention. The system achieved F-score of 0.714 for Task A & accuracy of 0.599 for Task B when trained only on training data set, and it achieved F-score of 0.755 for Task A & accuracy of 0.646 for Task B when trained on both training as well as development data set. Our system was placed 3rd for both task A and B. 1 Introduction A clinical document contains plethora of informa- tion regarding patient’s medical condition in un- structured format. So a sophisticated NLP sys- tem built specifically for clinical domain can be very useful in many different clinical applications. In recent years, clinical NLP has gained a lot of significance in research community because it contains challenging tasks such as medical entity recognition, abbreviation disambiguation, inter- conceptual relationship detection, anaphora res- olution, and text summarization. Clinical NLP has also gained a significant attraction among the 1http://www.nlm.nih.gov/research/umls/ This work is licenced under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are added by the organizers. License details: http: //creativecommons.org/licenses/by/4.0/ health care industry because it promises to de- liver applications like computer assisted coding, automated data abstraction, core/quality measure monitoring, fraud detection, revenue loss preven- tion system, clinical document improvement sys- tem and so on. Task-7 of SemEval-2014 was in continuation of the 2013 ShaRe/CLEF Task-1 (Sameer Prad- han, et al., 2013). This task was about finding disorder mentions from the clinical text and as- sociating them with their related CUIs (concept unique identifiers) as given in the UMLS (Unified Medical Language System). UMLS is the largest available medical knowledge resource. It contains 2,885,877 different CUIs having 6,497,937 differ- ent medical terms from over 100 different medi- cal vocabularies. Finding accurate CUIs from free clinical text can be very helpful in many healthcare applications. Our aim for participating in this task was to explore new techniques of finding CUIs from clinical document. Over the last few years many different Clin- ical NLP systems like cTAKES (Savova, Guer- gana K., et al., 2010), MetaMap (A. Aronson, 2001), MedLEE (C. Friedman et al., 1994) have been developed to extract medical concepts from a clinical document. Most of these systems focus on rule based, medical knowledge driven dictio- nary look-up approaches. …

Content truncated to 500 words