Adaptive Design of Clinical Trials: A Sequential Learning Approach

The inefficiency of the conventional clinical trial design is often attributed to two main reasons: invariant type I and II errors as well as fixed sample sizes. To the best of our knowledge, we represent the first paper to present a novel and tractable Bayesian decision-theoretical framework on multi-period adaptive clinical design that accommodate different forms of experiments (or controls). Our framework builds on the sequential hypothesis testing paradigm in which the clinical trial designer can either choose among different experiments with different information, or to terminate the trial. We show that the log-likelihood ratio (LLR) converges to a diffusion process via a limiting approximation, where the designer runs a series of increasingly less informative experiments whose distributions under the null and the alternative converge in symmetric KL-divergence. The optimal solution to the resulting stochastic control problem is analytically solved. The insight is to use more informative controls when the belief is more certain, and to use less informative ones otherwise. We demonstrate using real world clinical trial data that our model performs considerably better than existing policies in terms of overall expected economic benefits. We believe this paper provides invaluable managerial insight and practical guidance for decision makers in adaptive clinical trials to better design experiment and make informed decisions.