Exploring Semi-Supervised Coreference Resolution of Medical Concepts using Semantic and Temporal Features

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 731–741, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Exploring Semi-Supervised Coreference Resolution of Medical Concepts using Semantic and Temporal Features Preethi Raghavan∗, Eric Fosler-Lussier∗, and Albert M. Lai† ∗Department of Computer Science and Engineering †Department of Biomedical Informatics The Ohio State University, Columbus, Ohio, USA {raghavap, fosler}@cse.ohio-state.edu, albert.lai@osumc.edu Abstract We investigate the task of medical concept coreference resolution in clinical text using two semi-supervised methods, co-training and multi-view learning with posterior regulariza- tion. By extracting semantic and temporal features of medical concepts found in clinical text, we create conditionally independent data views; co-training MaxEnt classifiers on this data works almost as well as supervised learn- ing for the task of pairwise coreference resolu- tion of medical concepts. We also train Max- Ent models with expectation constraints, using posterior regularization, and find that poste- rior regularization performs comparably to or slightly better than co-training. We describe the process of semantic and temporal feature extraction and demonstrate our methods on a corpus of case reports from the New England Journal of Medicine and a corpus of patient narratives obtained from The Ohio State Uni- versity Wexner Medical Center. 1 Introduction The clinical community creates and uses a variety of semi-structured and unstructured electronic tex- tual documents that include medical reports such as admission notes, progress notes, pathology re- ports, radiology reports and hospital discharge sum- maries. The documents, collectively termed clini- cal narratives, account for various medical condi- tions, procedures, diagnoses and assessments in a patient’s medical history. Researchers have inves- tigated ways in which clinical text can be automati- cally processed for enabling access to relevant infor- mation for physicians and health researchers (Embi and Payne, 2009). One application is to support pa- tient recruitment into clinical trials (research studies that try to answer scientific questions to find bet- ter ways to prevent, diagnose, or treat a disease) by matching patient characteristics against eligibil- ity criteria (Raghavan and Lai, 2010). While there has been significant efforts to move to structured data collection, clinical narratives remain a critical data source for these tasks. Extracting structured information from unstruc- tured clinical text using natural language processing (NLP) is complicated by the distinct clinical report- ing sub-language characterized by incomplete sen- tences and domain specific abbreviations (Friedman et al., 2002). The large number of clinical narra- tives generated per patient, over the years, along with redundant information within and across narra- tives, further adds to the complexity of using infor- mation structured using NLP. There is a tendency to copy and edit parts of an old clinical narrative when- ever a new one is created, thus leading to redundant information in clinical narratives of a patient. Fur- thermore, since different types of clinical narratives are created for different purposes, certain narratives may summarize information from various other, at times older, clinical narratives. All of this makes the task of automatically processing unstructured clin- ical …

Content truncated to 500 words