Optimizing Graphical Procedures for Multiplicity Control in a Confirmatory Clinical Trial via Deep Learning.

Area of Science:

• Biostatistics
• Clinical Trial Design
• Statistical Inference

Background:

• Confirmatory clinical trials often use graphical approaches for intersection hypotheses testing.
• Controlling Type I errors in the strong sense is crucial, with weighted Bonferroni-type procedures being a common method.
• Optimizing the graphical testing procedure based on prior knowledge (e.g., Phase II results) is key for Phase III studies.

Purpose of the Study:

• To evaluate existing derivative-free constrained optimization methods for graphical testing procedures.
• To propose and assess a novel deep learning enhanced optimization framework using feedforward neural networks (FNNs).
• To improve the balance between robustness and time efficiency in optimizing multiple testing procedures.

Main Methods:

• Numerical approximation of the objective function using feedforward neural networks (FNNs).
• Optimization using available gradient information with constraints on testing procedure features.
• Evaluation through simulation studies comparing the FNN-based approach with existing derivative-free and stochastic search methods.

Main Results:

• The FNN-based optimization framework demonstrates a superior balance of robustness and time efficiency compared to existing derivative-free methods.
• The proposed optimizer achieves moderate multiplicity adjusted power gains, particularly when dealing with a large number of hypotheses.
• The method effectively optimizes multiple testing procedures for specific study objectives, as shown in a case study.

Conclusions:

• Deep learning, specifically FNNs, provides an effective framework for optimizing complex multiple testing procedures in clinical trials.
• The proposed method enhances efficiency and statistical power in confirmatory trials.
• This approach offers a valuable tool for designing more effective and robust clinical trial strategies.