Adaptive Bayesian Clinical Trials: The Past, Present, and Future of Clinical Research

Abstract

Background/Objectives: Decision-analytic Bayesian approaches are ideally suited for designing clinical trials. They have been used increasingly over the last 30 years in developing medical devices and drugs. A prototype trial is a bandit problem in which treating participants is as important as treating patients in clinical practice after the trial. Methods: This article chronicles the use of the Bayesian approach in clinical trials motivated by bandit problems. It provides a comprehensive historical and practical review of Bayesian adaptive trials, with a focus on bandit-inspired designs. Results: The 20th century saw advances in Bayesian methodology involving computer simulation. In the 21st century, methods motivated by bandit problems have been applied in designing scores of actual clinical trials. Fifteen such trials are described. By far the most important Bayesian contributions in clinical trials are the abilities to observe the accumulating results and to modify the future course of the trial on the basis of these observations. In the spirit of artificial intelligence, algorithms are programmed to learn the optimal treatment assignments over the remainder of the trial. Conclusions: Bayesian trials are still nascent and represent a small minority of clinical trials, but their existence is changing the way investigators, regulators, and government and industry sponsors view innovation in clinical trials.