Evaluation of Machine Learning Approaches for Estimating Individualized Treatment Regimens for Time-to-Event Outcomes in Observational Studies

In this paper we provide an overview and evaluation of machine learning methods for estimating individualized treatment regimens (ITR) for time-to-event outcomes through maximizing restricted mean survival time (RMST) in observational studies with non-randomized treatment assignment. We present extensive simulation studies that closely mimicked real-world data under a set of scenarios representing different degrees of alignment between the observed regimen, which reflects the actual prescribing practice, and the optimal ITR. The simulation results include performance characteristics of the candidate methods in terms of their ability to recover the optimal ITR and various empirical measures of RMST gain based on the comparison between the estimated ITR and the actual prescribing practice. Direct methods for estimating ITR for survival outcomes did not show advantages over indirect methods based on predicted potential outcomes under our implementation and simulation settings. Among indirect methods, gradient boosting for estimating potential survival outcomes has an advantage over random survival forests and parametric methods. Estimating the value of estimated ITRs and associated gains over the actual prescribing practice remains a challenging problem, especially under complex confounding scenarios when either most patients do not receive optimal treatment or when the actual treatment assignments are already close to optimal.