Cohort Sherpherd II: Verifying Cohort Constraints from Hospital Visits

This paper describes the updated system created by the University of Texas at Dallas for content-based medical record retrieval submitted to the TREC 2012 Medical Records Track. Our system updates our work from the previous year by building a structured query for each cohort that captures the patient’s age, gender, hospital status, and medical assertion information. Further, all keywords that encode any medical phenomena from the query are recursively decomposed before being expanded using knowledge from UMLS, SNOMED, Wikipedia, and PubMed co-occurrences. An initial ranking of hospital visits is then obtained using BM25 relevance on an interpolation of these decomposed keywords. Finally, hospital visits are re-ranked according to the constraints extracted in the structured query. Four runs were submitted, comparing pair-wise combinations of complete vs. shallow keyword decomposition and full vs. negation-only assertion processing. Our highest scoring submission achieved an infNDCG score of 0.426.