Optimal adaptive allocation using deep reinforcement learning in a dose-response study.

Area of Science:

• Clinical Trials
• Pharmacometrics
• Machine Learning

Background:

• Accurate dose-response curve estimation and dose selection in Phase II trials are critical for drug development.
• Traditional methods often use equal subject allocation, which may not optimize performance metrics.
• Existing optimal allocation methods can be limited by sample size and asymptotic assumptions.

Purpose of the Study:

• To develop an adaptive allocation rule that directly optimizes performance metrics for dose-response studies.
• To enhance the efficiency and accuracy of dose selection in clinical trials.
• To explore the application of deep reinforcement learning in adaptive clinical trial design.

Main Methods:

• Constructed an adaptive allocation rule using deep reinforcement learning (DRL).
• Defined appropriate states and rewards within the DRL framework to guide allocation.
• Evaluated the DRL-based method against equal allocation, D-optimal, and TD-optimal methods via simulation.

Main Results:

• The proposed DRL-based adaptive allocation rule successfully improved the targeted performance metrics.
• Demonstrated superior performance compared to equal allocation, D-optimal, and TD-optimal methods.
• Achieved notable improvements, particularly when optimizing for mean absolute error, showing superiority across multiple metrics.

Conclusions:

• Deep reinforcement learning provides a powerful framework for creating effective adaptive allocation rules in clinical trials.
• The developed method offers a significant advancement over existing techniques for optimizing dose-response studies.
• This approach holds promise for more efficient and accurate drug development processes.