Bayesian adaptive randomization in the I-SPY2.2 sequential multiple assignment randomized trial

Abstract: The I-SPY2 phase 2 clinical trial is a long-running platform trial evaluating neoadjuvant treatments for locally advanced breast cancer, assigning subjects to tumor subtype-specific experimental agents via a response-adaptive randomization algorithm that updates randomization probabilities based on accruing evidence of efficacy. Recently, I-SPY2 has been reconfigured as a sequential multiple assignment randomized trial (SMART), known as I-SPY2.2, in which subjects who are predicted to not achieve a satisfactory response to an initial assigned therapy are re-randomized to a second subtype-specific treatment followed by standard rescue therapy if a satisfactory response is not predicted. The I-SPY2.2 SMART thus supports evaluation of entire treatment regimes that dictate the choice of treatments at each stage on the basis of the outcome pathological complete response (pCR). The transition of I-SPY2 to a SMART required development of a trial-specific response-adaptive randomization scheme in which randomization probabilities at each stage are updated based on evolving evidence on the efficacy of full regimes, so as to skew probabilities toward treatments involved in regimes that the current evidence suggests are optimal in the sense of maximizing the probability of pCR. The approach uses Thompson sampling, which updates randomization probabilities based on the posterior probability that treatments are implicated in optimal regimes. We present the proposed algorithm and empirical studies that demonstrate it improves within-trial regime-specific pCR rates and recommends optimal regimes at similar rates relative to uniform, nonadaptive randomization.