Adaptive Seamless Dose-Finding Trials
37 Pages Posted: 22 Aug 2020
Date Written: June 26, 2020
We propose a non-parametric online-learning framework to conduct early stage dose-finding clinical trials with a simultaneous consideration of efficacy and toxicity. It has two major benefits: efficient use of patient responses and immunity to model mis-specifications. First, unlike most Phase I trials which only keep track of the toxicity, our framework makes efficient use of patient responses and infers the efficacy of each dose at the same time. Second, our framework utilizes application-specific structures of the dose-efficacy and dose-toxicity curves without imposing any parametric forms and is thus immune to model mis-specifications. Because of the discontinuity arising from the binary response (the dose is safe or not), the standard approaches in continuum-armed bandits do not apply. We then propose two algorithms, which are easy to understand and implement, and analyze their regret. The first one follows dose escalation principles, which makes it appealing for adoption for practice, and analyzes the efficacy and toxicity simultaneously. The second one, which is asymptotically optimal up to a logarithmic factor, uses bisection search to identify a safe dose range and then applies upper confidence bound algorithms within the range to identify efficacious doses. We test our proposed algorithms with a benchmark commonly used in practice on synthetic and real data sets and the results show that they significantly outperform the benchmark.
Keywords: Dose-Finding, Joint Efficacy and Toxicity, Continuum-Armed Bandit, Dose Escalation, Upper Confidence Bound
Suggested Citation: Suggested Citation