Unsupervised dense random survival forests identify interpretable patient profiles with heterogeneous treatment benefit
Precision oncology seeks to match patients with treatments that maximize clinical benefit. Identifying clinically interpretable subgroups with differential treatment effects from randomized clinical trials, however, remains challenging. We propose a subgroup-discovery framework based on dense random survival forests, with a splitting criterion that combines prognostic discrimination and treatment-effect heterogeneity. The framework aggregates patient proximity matrices across multiple forest configurations, identifies subgroups using the resulting treatment-informed similarity representation, and summarizes them through concise covariate rules. An empirical calibration procedure assesses evidence for treatment-effect heterogeneity. Extensive simulations demonstrate the ability of the framework to distinguish heterogeneous from homogeneous treatment-effect settings and recover treatment-responsive subgroups, with empirical Type I error control under the homogeneous scenarios considered. An application to a Phase III oncology trial identifies a treatment-responsive patient profile defined by baseline characteristics, illustrating the potential of the framework to inform subgroup-focused clinical research and future trial design.