Optimal dose selection in phase I/II dose finding trial with contextual bandits: a case study and practical...

Area of Science:

• Clinical Trials
• Machine Learning in Medicine
• Pharmacometrics

Background:

• Classical dose-finding trials in early drug development are often inefficient.
• Response-adaptive designs improve efficiency but are not yet optimal.
• Machine learning (ML) methods, like contextual bandits (CB), show promise for optimizing dose selection.

Purpose of the Study:

• To present a case study on using ML for oncology phase I/II dose-finding trial designs.
• To evaluate practical aspects of ML-based designs, including interim analyses and modeling approaches.
• To provide recommendations for implementing ML in early-phase dose selection.

Main Methods:

• Utilized Thompson sampling and Bayesian bootstrap for contextual bandits (CB).
• Compared direct clinical utility modeling with joint efficacy-safety modeling and model-independent approaches (multi-armed bandits).
• Incorporated weak informative prior information and conducted extensive simulations across various dose-response scenarios.

Main Results:

• Simulation results compared different combinations of design settings and modeling methods.
• Evaluated the performance of ML approaches under feasible dose-response relationship scenarios.
• Identified optimal strategies for ML-based dose-finding in phase I/II trials.

Conclusions:

• The proposed ML approach, particularly contextual bandits, offers a more efficient and potentially optimal method for phase I/II dose-finding.
• Practical recommendations are provided for the implementation of these ML methods in clinical trial design.
• The study guides the selection of appropriate modeling strategies and design parameters for ML-based dose selection.