Optimal rule-based general A+B designs in phase I clinical trials.

Abstract

Lin et al. studied the statistical properties of rule-based general A+B designs using true toxicity probabilities, without and with dose de-escalation, and derived exact formulae for key statistical quantities essential for phase I clinical trials. We systematically evaluate the impact of the parameters of A, B, C, D and E on the key statistical quantities across five dose levels within the A+B=6 framework. Simulation studies are conducted with true toxicity probabilities generated using a dose-toxicity model on known prior toxicity probabilities. General A+B designs with dose de-escalation are more conservative than those without dose de-escalation. For either presence or absence of dose de-escalation, 1) under an under-toxic environment, the 33,222 design is the best design and 24,113 design is the second one, but 33,222 design requires two more patients expected to be treated than 24,113 design. 2) under a moderate-toxic and over-toxic environment, 24113 and 42,111 design demonstrates an absolute and evident superiority, respectively. 3) under an unknown-toxic environment, 33122 design is the optimal choice but requires a larger total number of patients expected to be treated. 4) under an over-/unknown-toxic environment, the traditional "3+3" is one of the top two smallest total number of patients expected to be treated and patients experiencing a dose-limiting toxicity. Some general A+B designs have superior performance than traditional "3+3" design, having the target toxicity level closest to target toxicity rate (TTR) and higher number of patients treated at the dose levels with toxicity probabilities close to TTR.