ENTROPY LEARNING FOR DYNAMIC TREATMENT REGIMES

摘要

Estimating optimal individualized treatment rules (ITRs) in single- or multi-stage clinical trials is a key element of personalized medicine and, as a result, is receiving increasing attention within the statistical community. Recent works have suggested that machine learning approaches can provide significantly better estimations than those of model-based methods. However, a proper inference for estimated ITRs has not been well established for machine learning-based approaches. In this paper, we propose an entropy learning approach for estimating optimal ITRs. We obtain the asymptotic distributions for the estimated rules in order to provide a valid inference. The proposed approach is demonstrated to perform well through extensive simulation studies. Finally, we analyze data from a multi-stage clinical trial for depression patients. Our results offer novel findings not revealed by existing approaches.