ezDI: A Supervised NLP System for Clinical Narrative Analysis

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 412–416, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics ezDI: A Supervised NLP System for Clinical Narrative Analysis Parth Pathak, Pinal Patel, Vishal Panchal, Sagar Soni, Kinjal Dani, Narayan Choudhary, Amrish Patel ezDI, LLC. {parth.p, pinal.p, vishal.p, sagar.s, kinjal.d, narayan.c, amrish.p} @ezdi.us Abstract This paper describes the approach used by ezDI at the SemEval 2015 Task-14: ”Anal- ysis of Clinical Text”. The task was di- vided into two embedded tasks. Task-1 re- quired determining disorder boundaries (in- cluding the discontiguous ones) from a given set of clinical notes and normalizing the dis- orders by assigning a unique CUI from the UMLS/SNOMEDCT1. Task-2 was about find- ing different type of modifiers for given disor- der mention. Task-2 was divided further into two subtasks. In subtask-2a, gold set of disor- der was already provided and system needed to just fill modifier types into the pre-specified slots. Subtask 2b did not provide any gold set of disorders and both the disorders and its re- lated modifiers are to be identified by the sys- tem itself. In Task-1 our system was ranked first with F-score of 0.757 for strict evalua- tion and 0.788 for relaxed evaluation. In both Task-2a and 2b our system was placed second with weighted F-score of 0.88 and 0.795 re- spectively. 1 Introduction Extracting medical information from clinical natural text has gained a lot of attraction over the past few years. Approximately 80% of patient related infor- mation resides under unstructured transcribed text. Amount of this unstructured text is increasing con- stantly and automated methods of extracting crucial information is of paramount interest to health care informatics industry. Task-14 of SemEval 2015 on 1http://www.nlm.nih.gov/research/umls/ ”analysis of clinical text” addresses the same con- cern. Task-14 of SemEval-2015 was in continuation of the 2013 ShaRe/CLEF Task-1 (Suominen, H. et al., 2013) and task-7 of SemEval 2014. The task was divided into two parts. In continuation of last year, task-1 was about finding disorder mentions from the clinical text and associating them with their related CUIs (concept unique identifiers) as given in the UMLS (Unified Medical Language System). This year one additional task (Task-2) of disorder modi- fier slot filing was added. Task-2 was further sub- divided into two parts. In subtask-2a, a gold set of disorder mentions was provided and the participants had to only fill the pre-specified slots with the nor- malized modifiers. In task 2b, no gold set of disorder mentions was provided. Figure1 provides detailed overview about task 1 and 2. Clinical NLP has evolved a lot in the tasks re- lated to medical entity detection. NLP systems like cTAKES (Savova, Guergana K., et al., 2010), MetaMap (A. Aronson, 2001) and MedLEE (C. Friedman et al., 1994) have focused on rule based and dictionary look-up approaches for thid task. Re- cently a few attempts have been made to use su- pervised and semi-supervised learning models. In 2009, Yefang Wang (Wang et al., 2009) used cas- …

Content truncated to 500 words