CAREER: Innovative Methods for Designing Adaptive Clinical Trials

这个教师早期职业发展（Career）项目的目标是为临床试验适应性设计中的最佳学习开发一个灵活的优化框架。自2004年美国联邦药物管理局（FDA）批准这一概念并启动试点研究以来，适应性设计一直备受关注。与预先确定样本量并仅在试验结束时得出结论的固定临床试验不同，适应性临床试验允许在将患者分配到治疗组、调整剂量水平或根据获得的成功或失败证据终止试验方面进行临时修改。适应性试验有望提高新疗法的安全性，缩短有效疗法的上市时间，并限制对劣质疗法的暴露。此外，适应性临床试验可以更容易地纳入患者异质性（例如，基于生物标志物或个体暴露），这可能导致针对特定患者亚群的更有效、个性化的治疗。该教育计划将在本科和研究生课程中纳入适应性设计概念，作为引入最佳学习技术的新手段。此外，该项目还将为通过两个合作医疗系统进行临床试验的医生、护士和其他医疗从业人员提供适应性方法的教育研讨会。该CAREER项目将通过建立两类最优学习问题的统一解决框架来推进知识，即(i)具有任意（可能相关）信念分布的排序和选择问题，目标是学习具有期望属性的群体，以及（ii）具有相关奖励的多武装强盗问题，目标是学习最佳群体和最大化总奖励。在每段时间内，手臂的任何子集都可以选择任意次数，但要根据总拉数的预算。结合贝叶斯统计，采用新颖的近似动态规划方法来研究解空间，并建立基于对偶理论的界，以评估解的质量。研究结果将揭示患者异质性在适应性临床试验设计中的作用，并应扩展到最佳学习的其他应用，如动态定价、收入管理和分类计划。

The goal of this Faculty Early Career Development (CAREER) project is to develop a flexible optimization framework for optimal learning in adaptive design of clinical trials. Adaptive designs have been of great interest since the Federal Drug Administration (FDA) approved the concept and initiated pilot studies in 2004. Unlike fixed clinical trials, where sample sizes are determined in advance and conclusions made only at the end of the trial, adaptive clinical trials allow for interim modifications in assigning patients to treatment groups, adjusting of dosage levels, or terminating trials based on developing evidence of success or failure. Adaptive trials hold the promise of improving safety of new treatments, reducing time to market of efficacious treatments, and limiting exposure to inferior treatments. Moreover, adaptive clinical trials can more easily incorporate patient heterogeneity (e.g., based on biomarkers or individual exposure), which may lead to more effective, personalized treatments for particular patient subsets. The educational plan will incorporate adaptive design concepts in undergraduate and graduate coursework as a novel means of introducing optimal learning techniques. In addition, the project will provide educational seminars on adaptive methods to physicians, nurses, and other healthcare practitioners who are conducting clinical trials through two collaborating healthcare systems. This CAREER project will advance knowledge by establishing unifying solution frameworks to two classes of optimal learning problems, namely (i) ranking and selection problems with arbitrary (possibly correlated) belief distribution and the objective of learning a population with a desired property, and (ii) multiarmed bandit problems with correlated rewards and the dual objective of learning the best population and maximizing the total reward, where at each period any subset of arms can be chosen any number of times subject to a budget on the total number of pulls. Novel approximate dynamic programming methods integrated with Bayesian statistics are employed to study the solution space and establish bounds based on duality theory that will assess the quality of solutions. The results will shed light on the role that patient heterogeneity plays in adaptive clinical trial design and should extend to other applications in optimal learning, such as dynamic pricing, revenue management, and assortment planning.