A Knowledge-Based Approach to Medical Records Retrieval.

Abstract

The NLM LHC team approached the cohort selection task of the 2011 Medical Records Track as a question answering problem. We developed 60 training topics and then manually converted those topics to question frames. We started with the evidence-based medicine well-formed question frame and expanded it to explicitly capture temporal and causal relations. We then implemented a syntactic-semantic method for extracting the question frames from the free text topics. Based on the clinical documentation standards and knowledge of the clinical documentation structure, we split each report type into sections corresponding to different categories of clinical content, with the result that each section contained a specific class of data. We then ranked each document section according to its likelihood of containing answers to specific question frame slots. For example, if a question concerns medications prior to admission, the answers should be found in the Medications on Admission and the Medical History sections. In addition, we split each section into Positive (containing asserted findings, problems, and interventions), Negative (in which findings are negated) and Possible (that includes all uncertain statements). After structuring the questions and the documents, we developed algorithms for expressing question frames in the languages of the two search engines used for retrieval: Essie and Lucene. In addition to the UMLS synonymybased