Optimal dose escalation methods using deep reinforcement learning in phase I oncology trials.

Area of Science:

• Clinical Pharmacology
• Biostatistics
• Artificial Intelligence

Background:

• Identifying the maximum tolerated dose (MTD) is critical in phase I clinical trials for novel anticancer drugs.
• Existing MTD-finding methods often show insufficient percentages of correct selection (PCS), hindering dose determination for later trial phases.

Purpose of the Study:

• To develop an advanced action rule for dose escalation/de-escalation and trial continuation/stopping.
• The goal is to maximize the percentage of correct MTD selection (PCS) in phase I trials.

Main Methods:

• Utilized deep reinforcement learning (DRL) to construct an adaptive action selection rule.
• Defined specific states, actions, and reward functions tailored for MTD identification.

Main Results:

• The DRL-based method demonstrated improved PCS compared to traditional designs.
• Simulations showed superior performance against 3+3, CRM, BLRM, BOIN, mTPI, and i3+3 designs.

Conclusions:

• Deep reinforcement learning offers a powerful framework for optimizing MTD selection in early-phase oncology trials.
• The proposed DRL strategy enhances the reliability of MTD identification, potentially improving subsequent trial phases.