Bayesian Prediction of Event Times Using Mixture Model for Blinded Randomized Controlled Trials.

Accurate prediction of key milestone dates, such as the timing of interim and final analyses, is crucial in event-driven clinical trials with time-to-event endpoints. These predictions facilitate timely decision-making, enhance strategic planning and optimize resource allocation while minimizing patient exposure to potentially ineffective or harmful therapies. Existing methods for predicting event timing in blinded randomized clinical trials (RCTs) typically assume identical time-to-event distributions for the treatment and control arms, implying no treatment effect. This assumption fails when the treatment is more effective than the control, which is often the very outcome the trial seeks to detect, leading to biased predictions. To address this issue, we propose a novel Bayesian Prediction of Event Times (BayesPET) method that allows for different time-to-event distributions between arms in blinded RCTs. We employ a mixture Weibull model for the observed interim event times, while addressing the critical challenge of label-switching in mixture models through truncated priors. Through extensive simulations and real-world applications to phase 3 clinical trials, we demonstrate the BayesPET produces superior predictive performance in both blinded and unblinded settings, supporting effective trial execution and accelerating the development of new therapies.