Vector Space Models for Encoding and Retrieving Longitudinal Medical Record Data

Abstract

Vector space models (VSMs) are widely used as information retrieval methods and have been adapted to many applications. In this paper, we propose a novel use of VSMs for classification and retrieval of longitudinal electronic medical record data. These data contain sequences of clinical events that are based on treatment decisions, but the treatment plan is not recorded with the events. The goals of our VSM methods are (1) to identify which plan a specific patient treatment sequence best matches and (2) to find patients whose treatment histories most closely follow a specific plan. We first build a traditional VSM that uses standard terms corresponding to the events found in clinical plans and treatment histories. We also consider temporal terms that represent binary relationships of precedence between or co-occurrence of these events. We create four alternative VSMs that use different combinations of standard and temporal terms as dimensions, and we evaluate their performance using manually annotated data on chemotherapy plans and treatment histories for breast cancer patients. In classifying treatment histories, the best approach used temporal terms, which had 87 % accuracy in identifying the correct clinical plan. For information retrieval, our results showed that the traditional VSM performed best. Our results indicate that VSMs have good performance for classification and retrieval of longitudinal electronic medical records, but the results depend on how the model is constructed.