A Query Taxonomy Describes Performance of Patient-Level Retrieval from Electronic Health Record Data

Performance of systems used for patient cohort identification with electronic health record (EHR) data is not well-characterized. The objective of this research was to evaluate factors that might affect information retrieval (IR) methods and to investigate the interplay between commonly used IR approaches and the characteristics of the cohort definition structure. We used an IR test collection containing 56 test patient cohort definitions, 100,000 patient records originating from an academic medical institution EHR data warehouse, and automated word-base query tasks, varying four parameters. Performance was measured using B-Pref. We then designed 59 taxonomy characteristics to classify the structure of the 56 topics. In addition, six topic complexity measures were derived from these characteristics for further evaluation using a beta regression simulation. We did not find a strong association between the 59 taxonomy characteristics and patient retrieval performance, but we did find strong performance associations with the six topic complexity measures created from these characteristics, and interactions between these measures and the automated query parameter settings. Some of the characteristics derived from a query taxonomy could lead to improved selection of approaches based on the structure of the topic of interest. Insights gained here will help guide future work to develop new methods for patient-level cohort discovery with EHR data.