Exploring Sensitive Biomarkers Associated With Short-Term Responses and Long-Term Outcomes Using Bayesian Additive Regression Trees.

Identifying good predictive baseline covariates to optimize the target population for a new treatment is an important aim in precision medicine. Early post-baseline biomarker responses may serve as a supportive guide for physicians to decide whether to continue a treatment. We propose an exploratory two-stage subgroup analysis method as a statistical tool to investigate the predictive value of such a short-term response for the benefit of a new treatment in terms of long-term outcomes. We use a flexible probability model, Bayesian additive regression trees (BART), to derive predictive conditional treatment effects (PCTE) based on a short-term post-baseline biomarker response using counterfactual modeling of responses to new and standard treatments for each patient. Constructing patient subgroups according to the PCTE, we analyze an observed long-term outcome to identify a sensitive subpopulation. We carry out extensive simulation studies to examine the operating characteristics of the proposed method. For illustration, we apply the proposed method to data from a randomized clinical trial in patients with locally advanced or metastatic breast cancer. These findings support the predictive value of short-term responses and provide a statistical approach for identifying patients expected to benefit from a new treatment in terms of a long-term outcome.