Key Concept Identification for Medical Information Retrieval

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 579–584, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Key Concept Identification for Medical Information Retrieval Jiaping Zheng College of Information and Computer Sciences University of Massachusetts Amherst, MA jzheng@cs.umass.edu Hong Yu Bedford VA Medical Center Bedford, MA Department of Quantitative Health Sciences University of Massachusetts Worcester, MA hong.yu@umassmed.edu Abstract The difficult language in Electronic Health Records (EHRs) presents a challenge to patients’ understanding of their own con- ditions. One approach to lowering the barrier is to provide tailored patient ed- ucation based on their own EHR notes. We are developing a system to retrieve EHR note-tailored online consumer ori- ented health education materials. We ex- plored topic model and key concept identi- fication methods to construct queries from the EHR notes. Our experiments show that queries using identified key concepts with pseudo-relevance feedback signifi- cantly outperform (over 10-fold improve- ment) the baseline system of using the full text note. 1 Introduction Allowing patients access to their own electronic health records (EHRs) can enhance medical un- derstanding and provide clinical relevant bene- fits (Wiljer et al., 2006), including increased med- ication adherence (Delbanco et al., 2012). How- ever, EHR notes present unique challenges to the average patients. Since these notes are not usually targeted at the patients (Mossanen et al., 2014), languages that may be difficult for non-medical professionals to comprehend are prevalent, in- cluding medical terms, abbreviations, and medi- cal domain-specific language patterns. The valu- able and authoritative information contained in the EHR is thus less accessible to the patients, who ul- timately stand to benefit the most from the infor- mation. To address the challenges, we are developing a system to link medical notes to targeted education materials from trusted resources. The textual nar- ratives in the EHR notes are not conducive to ef- fective and efficient query. We therefore explored topic model and key concept identification meth- ods to construct short queries from the EHR notes. 2 Related Work Patent retrieval (Fujii et al., 2007) is similar to this work, as the queries are usually long and complex patent documents. Several methods have been proposed to construct shorter queries from the documents. For example, words in the summary section of a patent document can be ranked by TFIDF scores and extracted to form a query (Xue and Croft, 2009). Sentences that are similar to pseudo-relevant documents according to a language model are also used to reduce query length (Ganguly et al., 2011). Other similarity measures such as Kullback-Leibler divergence are used to extract terms, which are expanded to gen- erate queries in the patent retrieval domain (Mahd- abi et al., 2012). However, the patent retrieval do- main is recall-driven, while in our scenario, pa- tients are generally not expected to read relevant education documents exhaustively. Various methods have also been proposed to re- trieve documents relevant to passages of text or web documents. A model extended from CRF is proposed to identify …

Content truncated to 500 words