From Bayesian dose-finding to AI-enabled decision support in oncology clinical trials.

e15145 Background: Bayesian methods are increasingly used in oncology clinical development to support dose finding, adaptive trial designs, and decision-making under uncertainty. Recent FDA draft guidance on Bayesian methodology clarifies regulatory expectations and supports broader use of Bayesian approaches across development phases. However, utilization remains limited in small and mid-sized biotechnology and pharmaceutical organizations, due to limited specialized expertise and operational complexity. Advances in automation and clinical trial data integration create opportunities to extend Bayesian frameworks toward AI-enabled decision support. Methods: We review Bayesian and model-informed methods used across Phase I-III oncology trials, focusing on trial design, interim monitoring, and development decision-making. Approaches include model-assisted and model-based dose-escalation designs, Bayesian predictive probability methods, and adaptive designs such as response-adaptive randomization and platform trials conducted under master protocols. Methods are evaluated based on modeling assumptions, simulation-based operating characteristics, regulatory considerations, and feasibility for integration into automated, AI-enabled decision support workflows using standard clinical trial data. Results: Examples demonstrate the application of Bayesian methods across clinical development phases. In Phase I oncology studies, the Bayesian Optimal Interval design supports efficient dose identification while controlling toxicity risk through prespecified decision rules. In Phase II dose-expansion, Bayesian continuous monitoring supports early assessment of futility or efficacy and informs development decisions. In later phase development, Bayesian adaptive designs, including response-adaptive randomization and platform trials implemented under master protocols, support efficient evaluation of multiple treatments and adaptive trial modification. These designs have been implemented in oncology programs (NCT06726161, NCT06590857, NCT05595460, NCT07165132). For illustrative purposes, simulation results are presented to characterize operating characteristics under clinically relevant scenarios rather than actual trial parameters. Conclusions: Bayesian and predictive modeling approaches provide a consistent foundation for model-informed decision-making in oncology clinical trials and a pathway toward AI-enabled decision support. The presentation will highlight key opportunities for AI to enhance Bayesian decision frameworks, including automation and integration of real-time trial data. By standardizing and automating established Bayesian analyses, future AI-driven systems may support broader and more efficient application while maintaining statistical rigor and regulatory acceptability, consistent with recent FDA guidance.