UWM-TRIADS: Classifying Drug-Drug Interactions with Two-Stage SVM and Post-Processing

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 667–674, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UWM-TRIADS: Classifying Drug-Drug Interactions with Two-Stage SVM and Post-Processing Majid Rastegar-Mojarad Richard D. Boyce Rashmi Prasad University of Wisconsin-Milwaukee University of Pittsburgh University of Wisconsin-Milwaukee Milwaukee, WI, USA Pittsburgh, PA, USA Milwaukee, WI, USA Rastega3@uwm.edu rdb20@pitt.edu prasadr@uwm.edu Abstract We describe our system for the DDIExtraction-2013 shared task of classifying Drug-Drug interactions (DDIs) given labeled drug mentions. The challenge called for a five-way classification of all drug pairs in each sentence: a drug pair is either non-interacting, or interacting as one of four types. Our approach begins with the use of a two-stage weighted SVM classifier to handle the highly unbalanced class distribution: the first stage for a binary classification of drug pairs as interacting or non-interacting, and the second stage for further classification of interacting pairs from the first stage into one of the four interacting types. Our SVM features exploit stemmed words, lemmas, bigrams, part of speech tags, verb lists, and similarity measures, among others. For each stage, we also developed a set of post-processing rules based on observations in the training data. Our best system achieved 0.472 F- measure. 1 Introduction Potential drug-drug interactions (DDIs), defined as the co-prescription of two drugs that are known to interact, are a significant source of pre- ventable drug-related harm (i.e., adverse drug events, or ADEs) (Nebeker et al., 2004). Gurwitz et al, in their cohort study of ADEs among older Americans receiving ambulatory care, found that 13.3% of preventable errors leading to an ADE involved the co-prescription of drugs for which a “...well established, clinically important interac- tion” was known (Gurwitz et al., 2003). Nearly 7% (23/338) of the ADEs experienced by resi- dents of two academic nursing homes over a nine-month period were attributable to DDIs (Gurwitz et al., 2005). Sixteen cohort and case- control studies reported an elevated risk of hospi- talization in patients who were exposed to DDIs (Hines et al., 2011). Failure to properly manage a DDI is a medical error, and the Institute of Medicine has noted that a lack of drug knowledge is one of the most fre- quent proximal causes of such errors (Committee on Identifying and Preventing Medication Errors, 2007). Indeed, health care providers often have inadequate knowledge of what drug interactions can occur, of patient specific factors that can in- crease the risk of harm from an interaction, and how to properly manage an interaction when pa- tient exposure cannot be avoided (Chen et al., 2005; Hines et al., 2012). Unfortunately, there is no single complete and authoritative source of DDI knowledge (Hines et al., 2012). Rather, there are multiple sources, each tasked with extracting, evaluating, and stay- ing up-to-date with pertinent DDIs reported in the literature, and drug product labeling (Boyce et al., 2012). The dynamic nature of drug knowledge, combined with the enormity of the biomedical literature, makes this …

Content truncated to 500 words