Characterizing Subpopulations with Better Response to Treatment Using Observational Data - an Epilepsy Case Study

Electronic health records and health insurance claims, providing observational data on millions of patients, offer great opportunities, and challenges, for population health studies. The objective of this study is to utilize observational data for identifying subpopulations that are likely to benefit from a given treatment, compared to an alternative. We refer to these subpopulations as "better responders", and focus on characterizing them using linear scores with a limited number of variables. Building upon well-established causal inference techniques for analyzing observational data, we propose two algorithms that generate sparse linear scores for identifying better responders, as well as methods for evaluating and comparing such scores. We applied our methodology to a large dataset of ~135,000 epileptic patients derived from claims data. Out of this sample, 85,000 were used to characterize subpopulations with better response to next-generation ("Newer") anti-epileptic drugs (AEDs), compared to an alternative treatment by first-generation ("Older") AEDs. The remaining 50,000 epileptic patients were then used to validate and compare the ability of our scores to identify large subpopulations of epileptic patients with significantly better response to newer AEDs.