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

Area of Science:

• Clinical Trials Methodology
• Biostatistics
• Pharmacoeconomics

Background:

• Accurate prediction of milestone dates in event-driven clinical trials is critical for decision-making and resource allocation.
• Current methods for predicting event times in blinded randomized clinical trials (RCTs) often assume no treatment effect, leading to biased predictions when a treatment effect exists.

Purpose of the Study:

• To introduce a novel Bayesian Prediction of Event Times (BayesPET) method for predicting event timings in blinded RCTs.
• To address the limitation of existing methods by allowing for different time-to-event distributions between treatment and control arms.

Main Methods:

• Developed the BayesPET method using a mixture Weibull model for interim event times.
• Addressed the label-switching challenge in mixture models using truncated priors.
• Validated the method through extensive simulations and real-world phase 3 clinical trial data.

Main Results:

• The BayesPET method demonstrated superior predictive performance compared to existing methods.
• The model showed effectiveness in both blinded and unblinded trial settings.
• Accurate predictions were achieved even when treatment arms had different time-to-event distributions.

Conclusions:

• The BayesPET method offers a more accurate approach to predicting event times in clinical trials, especially when treatment effects are present.
• This improved prediction supports more effective trial execution and can accelerate the development of new therapies.
• The method enhances strategic planning and resource optimization in clinical trial management.