Identifying Potential Patient Population Based on Electronic Medical Record Data

• Activity Number: 466
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Jin Zhou*+ and Haoda Fu
• Companies: University of Arizona and Eli Lilly and Company
• Keywords: EMR ; RCT ; Subgroup Identification ; Variable selection ; observational study ; personalized medicine

Abstract

To compare a treatment with a control via a randomized clinical trial (RCT), the assessment of the treatment efficacy is often based on an overall treatment effect over a specific study population. To increase the probability of study success, it is important to choose an appropriate and relevant study population where the treatment is expected to show overall benefit over the control. In this paper, we explore several possible methods to identify such a population using Electronic medical records (EMRs) data. In the literature, there are efforts towards using RCT data for patient subgrouping. Compared to RCT, EMR are referred as one type of non-randomized data or "observational" data. The key challenges behind borrowing information from EMR are how to tease out the signal from the confounding factors and establish the causal relationship. We first estimate propensity score using comprehensive EMR data, which is used for sample bias adjustment. Several variable selection methods were adopted for high dimensional EMR data. Our method is an individualized treatment recommendation system, which is based on a new optimization scheme adjusting for baseline/subject level effects.