Tests and classification methods in adaptive designs with applications.

Area of Science:

• Biostatistics
• Genomic Medicine
• Clinical Trial Design

Background:

• Adaptive clinical trials are crucial for genomic studies, requiring robust methods for biomarker identification and patient stratification.
• Efficiently identifying biomarkers and grouping patients improves trial power and reduces sample size.

Purpose of the Study:

• To evaluate and compare statistical tests for biomarker identification in the first stage of adaptive designs.
• To assess various classification methods for patient grouping in the second stage of adaptive designs.
• To apply the best-performing methods to the Adaptive Signature Design (ASD) using breast cancer data.

Main Methods:

• Biomarker identification methods evaluated: model-based, t-test, Wilcoxon Rank-Sum test, and Regularized Generalized Linear Models.
• Classification methods evaluated: Random Forest, Elastic-net, Support Vector Machine (SVM), Gradient Boosting Machine (GBM), and Extreme Gradient Boosting (XGBoost).
• Performance assessment via simulation studies, using the $$ score for identification and Area Under the ROC Curve (AUC) for classification.

Main Results:

• Identified optimal statistical tests for biomarker discovery in early-stage adaptive trials.
• Determined the most effective classification algorithms for patient stratification in later stages.
• Demonstrated the practical application and performance of selected methods within the Adaptive Signature Design framework using real-world breast cancer data.

Conclusions:

• The study provides a framework for selecting superior statistical methods in adaptive genomic clinical trials.
• The findings contribute to optimizing trial design for biomarker-driven research and personalized medicine.
• The evaluation offers practical guidance for researchers utilizing adaptive designs in oncology and other fields.