M‐Learning for Individual Treatment Rule With Survival Outcomes

Individualized treatment rules (ITRs) tailor treatments to individuals based on their unique characteristics to optimize clinical outcomes and resource allocation. Current approaches use outcome modeling or propensity score weighting to control confounding in complex medical data. To avoid model misspecification and the impact of extreme weights, matched-learning (M-learning) was recently proposed for continuous outcomes. In this paper, we expand the existing M-learning methodology to estimate optimal ITRs under right-censored data, as time-to-event outcomes are common in medical research. We construct matched sets for individuals by comparing observed times and incorporate an inverse probability censoring weight into the value function to handle censored observations. Additionally, we consider a full matching design as a possible alternative to the matching with replacement in M-learning. We demonstrate that the proposed value function is unbiased for the true value function without censoring. To gain insight into the empirical performance, we conduct an extensive simulation study that compares M-learning with two matching designs and a weighed learning approach. Results are evaluated based on winning probabilities and estimated values. The simulation reveals that all methods are generally fine in the absence of unmeasured confounders, and different methods show somewhat different performances under various scenarios. But their performance drops substantially in the presence of unmeasured confounders. Finally, we apply these methods to estimate optimal ITRs for patients with atrial fibrillation (AF) complications from an electronic medical record database, where full matching design shows slightly better performance.