Stochastic approximation with virtual observations for dose-finding on discrete levels.

Area of Science:

• Biostatistics
• Clinical Pharmacology
• Drug Development

Background:

• Phase I clinical studies determine maximum tolerated drug doses.
• Dose-finding is often a quantile estimation problem.
• Continuous biomarkers are often dichotomized, losing information.

Purpose of the Study:

• Propose a novel Robbins-Monro stochastic approximation variant for quantile estimation.
• Utilize continuous biomarker measurements for more accurate dose-finding.
• Address limitations of traditional methods with binary data and discrete doses.

Main Methods:

• Formulate dose-finding as root-finding for the mean of a continuous variable.
• Introduce 'virtual observation' for discrete dose application.
• Adapt Robbins-Monro stochastic approximation for clinical dose-finding.

Main Results:

• Proposed method improves accuracy over the continual re-assessment method.
• Demonstrates robustness to model misspecification in simulations.
• Inherits convergence properties and computational simplicity of stochastic approximation.

Conclusions:

• Novel method enhances Phase I dose-finding accuracy using continuous data.
• Virtual observation enables application with discrete doses.
• Offers a computationally simple and robust alternative for clinical trials.