Feature Creation and Model Uncertainty in Observational Medical Data

• Activity Number: 509
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Rebecca Ferrell*+ and Tyler H. McCormick
• Companies: University of Washington and University of Washington
• Keywords: variable selection ; time series ; medical records ; model uncertainty ; time discretization

Abstract

Large-scale observational health databases (such as electronic medical records or administrative claims data) capture continuous-time, unsolicited recordings of patient experiences. As with many emerging data sources without a formal sampling design, these data require substantial pre-processing before using standard statistical tools. For observational health databases, pre-processing often involves coding for characteristics present at a designated baseline period through discretization of the temporal element of the records, e.g. coarsening the health event timelines over a specified "lookback period" into a binary or count feature to capture prior disease history. Though there is rich literature examining model selection, very little work examines these pre-processing, "feature creation" choices. We explore this problem through the framework of model uncertainty in the context of medical event prediction. Through simulations and an application to health claims data, we demonstrate the effect of decisions to encode time-varying information as static baseline covariates on predictive performance and discuss approaches to account for uncertainty in defining lookback periods.