Automated Activity Recognition in Clinical Documents

International Joint Conference on Natural Language Processing, pages 1129–1133, Nagoya, Japan, 14-18 October 2013. Automated Activity Recognition in Clinical Documents C. Thorne M. Montali D. Calvanese Free University of Bozen-Bolzano Bolzano, Italy surname@inf.unibz.it E. Cardillo C. Eccher Fondazione Bruno Kessler Trento, Italy surname@fbk.edu Abstract We describe a first experiment on the identification and extraction of computer- interpretable guideline (CIG) components (activities, actors and consumed artifacts) from clinical documents, based on clinical entity recognition techniques. We rely on MetaMap and the UMLS Metathesaurus to provide lexical information, and study the impact of clinical document syntax and semantics on activity recognition. Keywords. Clinical entity recognition, computer interpretable guideline, UMLS Metathesaurus. Introduction. Clinical practice guidelines are sys- tematically developed documents specifying the activities, resources and personnel required to cure or treat an specific illness or medical condition (Field and Lohr (1990)). The need to instan- tiate them into clinical protocols and workflows has given rise to computer-interpretable guide- lines (CIGs) (De Clercq et al. (2008)), i.e., formal representations of the care process or plan, and to several natural language processing (NLP) tech- niques aimed at automating the costly manual CIG generation process (Kaiser et al. (2007), Serban et al. (2007)). All NLP approaches leverage on an- notated biomedical resources (e.g., the CLEF cor- pus from Roberts et al. (2007) and Mykowiecka and Marciniak (2011)), or on frameworks such as cTAKES (Savova et al. (2010)). The key lexical-semantic resource in this domain is the US National Library of Medicine’s Unified Medical Language System (UMLS) Metathesaurus (Bo- denreider (2004)), complemented by its front-end MetaMap (Aronson and Lang (2010)). In this paper we conduct a first experiment on how to apply entity recognition techniques in- spired by Abacha and Zweigenbaum (2011), to recognize CIG components in medical documents. The process dimension of CIGs consists of four pillars: (1) activities to be executed; (2) the re- sources they use or consume; (3) the actors that execute them; (4) control flows and gates that tem- porally constrain activities. We focus in this pa- per on activities, the main building block of CIGs, and to a lesser extent on resources and actors. All these components are denoted by content words and can be used to build CIG fragments. We rely on MetaMap annotations and evaluate our tech- niques over an UMLS-annotated clinical corpus. CIGs and Activities. Activities are entities diffi- cult to identify with current resources: within clin- ical documents, in fact, not only verbs (VBs) but also proper nouns (PNs), common nouns (NNs) and, more in general, noun phrases (NPs)1 can refer to them. Figure 1 shows an example from the type-2 diabetes guideline of the National In- stitute for Health and Clinical Excellence (NICE) (NICE - NHS (2009)) expressing a conditional CIG/process fragment, annotated automatically with MetaMap. To correctly extract the “deep” intended representations it is necessary to recog- nize that the two entities “blood glucose control” and “oral glucose-lowering medication” are activ- ity tokens. MetaMap annotations provide a clue, but we still need to “filter out” …

Content truncated to 500 words