Interpretable Feature Creation and Model Uncertainty in Observational Medical Data

• Activity Number: 637
• Type: Topic Contributed
• Date/Time: Thursday, August 13, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Rebecca Ferrell* and Tyler McCormick
• Companies: University of Washington and University of Washington
• Keywords: electronic health records ; variable selection

Abstract

With longitudinal observational medical data, the question arises of how to construct the regression design matrix of covariates when a natural form not is suggested by the source data. We examine "feature creation" decisions in a longitudinal and hierarchical context in which we have granular medical events by date with subject identifiers. For observational health databases, feature creation often involves coding for characteristics present at a designated baseline period through discretization of the temporal element of the records, e.g. coarsening patient timelines over a specified period into a feature to capture prior disease history. Our proposed approach to create meaningful covariates in these applications is to develop a set of generic concepts and compile a set of potential definitions for each, which may vary in hierarchical or temporal resolution. These correlated "competing definitions" for the same concept are treated as mutually exclusive to retain interpretability in regression models using structured prior distributions and stochastic search variable selection (SSVS).