Enhancing Adaptive Trial Integrity and Efficiency: From Bayesian Optimization to AI Implementation

Adaptive clinical trials increasingly leverage Bayesian optimization and AI-enabled decision tools to improve efficiency, enhance patient safety, and preserve statistical integrity. This session brings together five experts to highlight emerging methods for dose-finding, response-adaptive randomization (RAR), and real-time trial monitoring. The session addresses the crucial contemporary challenge in clinical development: integrating complex, efficient adaptive methodologies with the necessary statistical validity and integrity, as well as operational scalability. The talks cover practical advancements, moving from enhanced Phase I design and systematic RAR regularization to navigating regulatory requirements and leveraging cutting-edge AI for trial planning. The content has strong timeliness, as it provides concrete solutions for designs that are increasingly required for efficiency but often face implementation hurdles (e.g., controlling bias and Type I error, dealing with small sample sizes, and adhering to new safety monitoring guidelines). Its appeal lies in combining methodological innovation with practical software and regulatory solutions, relevant to biostatisticians, clinical trial methodologists, and regulatory scientists. The session will also explore how machine learning and modern computational platforms can support reproducible, regulator-ready adaptive designs. Attendees will gain insights into best practices for integrating Bayesian and AI methodologies while maintaining rigor, interpretability, and trustworthiness in clinical development. Speaker 1 (Yaron Racah, PhaseV Trials) will begin by proposing an extension of the popular BOIN (Bayesian optimal interval) design for Phase 1 trials that retains BOIN's operational simplicity but improves its flexibility and statistical performance. The method improves MTD selection accuracy while preserving BOIN's ability to summarize operating rules in simple lookup tables readily communicated to clinical teams. Speaker 2 (Lukas Pin, University of Cambridge, UK) will then present novel methods for informed burn-in decisions in RAR. Specifically, the approach harmonizes adaptivity and inferential precision based on the particular setting of the clinical study. Speaker 3 (Brad Carlin, PhaseV Trials) will then discuss adaptive strategies for navigating new FDA requirements for monitoring overall survival (OS) as a safety endpoint in oncology clinical trials. After a brief review of the new guidance document, he will discuss frequentist and Bayesian stopping rules for OS-based harm, illustrating in the case of a seamless Phase 2-3 trial for hepatocellular cancer. A key challenge here is to balance the need to limit exposure to a potentially harmful drug with the need to avoid prematurely discontinuing the trial. Speaker 4 (Subhajit Sengupta, Cytel) will describe a retrieval-augmented conversational assistant for clinical trial design and simulation. The system combines Azure OpenAI's GPT-4.1 model with a vector database of methodological and product documentation. Benchmarking against expert responses demonstrates the assistant's high correctness and completeness. Finally, the discussant (Sofia Dias, University of Cambridge, UK) will summarize the talks and offer her views on the routine incorporation of the tools presented into modern clinical trials.