An Adaptive Biomarker-based Umbrella Trial Design Using Bayesian Latent Class Model

A Bayesian adaptive umbrella trial design is proposed to identify effective treatments for biomarker-defined subgroups. This design employs a mixture regression model to jointly model short-term binary and long-term time-to-event outcomes and introduces a latent class model to adaptively cluster subgroups with similar treatment-outcome distributions. The model aims to dynamically borrow information between subgroups in the same class, enhancing the efficiency and robustness in testing subgroup-specific treatment effects. To strengthen the individual ethics of the trial, the design incorporates response-adaptive randomization with early stopping rules for futility and superiority. The simulation results confirm that the proposed design demonstrates desired performance.