Learning from Medical Summaries: The University of Michigan at TREC 2015 Clinical Decision Support Track.

The goal of TREC 2015 Clinical Decision Support Track was to retrieve biomedical articles relevant for answering three kinds of generic clinical questions, namely diagnosis, test, and treatment. In order to achieve this purpose, we investigated three approaches to improve the retrieval of relevant articles: modifying queries, improving indexes, and ranking with ensembles. Our final submissions were a combination of several different configurations of these approaches. Our system mainly focused on the summary fields of medical reports. We built two different kinds of indexes – an inverted index on the free text and a second kind of indexes on the Unified Medical Language System (UMLS) concepts within the entire articles that were recognized by MetaMap. We studied the variations of including UMLS concepts at paragraph and sentence level and experimented with different thresholds of MetaMap matching scores to filter UMLS concepts. The query modification process in our system involved automatic query construction, pseudo relevance feedback, and manual inputs from domain experts. Furthermore, we trained a re-ranking sub-system based on the results of TREC 2014 Clinical Decision Support track using Indri’s Learning to Rank package, RankLib. Our experiments showed that the ensemble approach could improve the overall results by boosting the ranking of articles that are near the top of several single ranked lists.