UNT Medical Information Retrieval at TREC 2016.

This paper provides a description of a project to design and evaluate an information retrieval system for clinical decision support track. The target document collection for retrieval consisted of 1.25 million biomedical related documents taken from the Open Access Subset of PubMed Central. The topics provided by TREC for query construction consisted of 30 patient narrative cases, each of which includes a Note section, a Description section, and a Summary section. The PMCID, title, abstract, keywords, subheadings of article body, introduction and conclusion paragraphs were extracted from the documents. Terrier was used as the platform for indexing and retrieval. Several models, including the LemurTF_IDF weighting model with pseudo relevancy feedback, were applied to retrieve and rank relevant documents. Out of the five runs submitted, two runs were performed by merging the retrieval results of top five individual weighting models, and the remaining three runs were obtained through passing three types of queries constructed, manually and automatically, using the Note and the Summary sections. The automatic runs are observed to receive a better performance than the manual runs. The automatic run using the Note section for query construction performed better than other runs. Overall performance of the system is around the median when compared to all TREC 2016 CDS Track submissions .