Bidirectional RNN for Medical Event Detection in Electronic Health Records

Proceedings of NAACL-HLT 2016, pages 473–482, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Bidirectional RNN for Medical Event Detection in Electronic Health Records Abhyuday N Jagannatha1, Hong Yu1,2 1 University of Massachusetts, MA, USA 2 Bedford VAMC and CHOIR, MA, USA abhyuday@cs.umass.edu , hong.yu@umassmed.edu Abstract Sequence labeling for extraction of medical events and their attributes from unstructured text in Electronic Health Record (EHR) notes is a key step towards semantic understand- ing of EHRs. It has important applications in health informatics including pharmacovig- ilance and drug surveillance. The state of the art supervised machine learning models in this domain are based on Conditional Random Fields (CRFs) with features calculated from fixed context windows. In this application, we explored recurrent neural network frameworks and show that they significantly outperformed the CRF models. 1 Introduction EHRs report patient’s health, medical history and treatments compiled by medical staff at hospitals. It is well known that EHR notes contain information about medical events including medication, diagno- sis (or Indication), and adverse drug events (ADEs) etc. A medical event in this context can be described as a change in patient’s medical status. Identifying these events in a structured manner has many im- portant clinical applications such as discovery of ab- normally high rate of adverse reaction events to a particular drug, surveillance of drug efficacy, etc. In this paper we treat EHR clinical event detection as a task of sequence labeling. Sequence labeling in the context of machine learning refers to the task of learning to predict a la- bel for each data-point in a sequence of data-points. This learning framework has wide applications in many disciplines such as genomics, intrusion detec- tion, natural language processing, speech recogni- tion etc. However, sequence labeling in EHRs is a challenging task. Unlike text in the open domain, EHR notes are frequently noisy, containing incom- plete sentences, phrases and irregular use of lan- guage. In addition, EHR notes incorporate abundant abbreviations, rich medical jargons, and their varia- tions, which make recognizing semantically similar patterns in EHR notes difficult. Additionally, dif- ferent events exhibit different patterns and possess different prevalences. For example, while a medi- cation comprises of at most a few words of a noun, an ADE (e.g., “has not felt back to his normal self”) may vary to comprise of a significant part of a sen- tence. While medication information is frequently described in EHRs, ADEs are typically rare events. Rule-based and learning-based approaches have been developed to identify and extract informa- tion from EHR notes (Haerian et al., 2012), (Xu et al., 2010), (Friedman et al., 1994), (Aronson, 2001), (Polepalli Ramesh et al., 2014). Learning- based approaches use sequence labeling algorithms like Conditional Random Fields (Lafferty et al., 2001), Hidden Markov Models (Collier et al., 2000), and Max-entropy Markov Models (McCal- lum et al., 2000). One major drawback of these graphical models is that the label prediction at any time point only depends on its data instance and the immediate neighboring labels. …

Content truncated to 500 words