Outcome Polarity Identification of Medical Papers

Outcome Polarity Identification of Medical Papers Abeed Sarker , Diego Moll´a-Aliod Centre for Language Technology Macquarie University Sydney, NSW 2109 {abeed.sarker,diego.molla-aliod}@mq.edu.au C´ecile Paris CSIRO – ICT Centre Sydney, NSW 2122 cecile.paris@csiro.au Abstract A medical publication may or may not present an outcome. When an outcome is present, its polarity may be positive, negative or neu- tral. Information about the polarity of an out- come is a vital one, particularly for practi- tioners who use the outcome information for decision making. We model the problem of automatic outcome polarity identification as a three-way document classification problem and attempt to solve it via supervised machine learning. We combine domain knowledge and linguistic features of medical text, and apply natural language processing to extract features for the chosen classifiers. We introduce two novel features — Relative Average Negation Count and Sentence Signature — and show that they are effective in improving classifica- tion accuracy. We also include features, such as n-grams and semantic orientation of terms, that have been used for similar text classifica- tion problems in other domains. Using these features, we obtain a maximum accuracy of 74.9% for the classification problem. Our ex- periments suggest that through careful feature selection, machine learning can be used to solve this problem. 1 Introduction The phenomenal growth of biomedical literature has presented medical practitioners, particularly those practicing Evidence Based Medicine (EBM), with the problem of information overload. The popu- lar practice of EBM requires practitioners to review medical literature before making clinical decisions (Sackett et al., 1996; Greenhalgh, 2006). When reviewing medical publications, EBM practitioners are mostly interested in identifying the outcomes presented and their polarities. The polarity of an outcome can be positive (e.g. the study shows that drug X is useful for patients suffering from condition Y), negative (e.g. the study suggests that drug X is not recommended for patients suffering from condi- tion Y) or the publication may present a neutral out- come or may not present an outcome at all (e.g. the study does not produce conclusive results regarding the efficacy of drug X for condition Y). Manually as- sessing the outcomes presented by multiple medi- cal papers on a given topic is a time-consuming task and often cannot be efficiently performed at point of care (Ely et al., 1999). Hence, there is a strong need for automatic outcome polarity identification tech- niques to aid the decision making process of practi- tioners. 1.1 Motivation In order to appease the problem of information over- load faced by medical domain experts, research has focused on information retrieval, automatic sum- marisation and question answering of medical doc- uments (Lin and Demner-Fushman, 2007; Fiszman et al., 2009). Intelligent text processing systems that perform automatic summarisation and question answering for this domain can benefit significantly from techniques that can automatically detect the polarity of outcomes presented in documents. Such techniques will be particularly useful for multidocu- ment summarisation, where the detection of contra- dictory or consistent outcomes presented in separate documents is vital. Furthermore, recent …

Content truncated to 500 words