Abstract
AbstractThe calibration-free odds type (CFO-type) of designs, as data-driven decision-making Bayesian approaches, leverage historical cumulative data across various dose levels, primarily aiming at identifying the maximum tolerated dose (MTD). Inheriting the ideas from game theory or “tug-of-war”, CFO mimics the games of force: one pushes the dose down while the other pushes it up. Extensive simulations validate that CFO-type designs maintain an optimal balance between efficiency and safety in MTD identification, with performance metrics that are comparable to, or occasionally surpass the state-of-the-art methods. This article primarily introduces the R packageCFOfor implementing and assessing CFO-type designs in phase I clinical trials. Besides, we propose integrating the mechanism of exploration and exploitation from reinforcement learning into the CFO design, leading to a novel approach: the randomized CFO (rCFO) design. TheCFOpackage encompasses various variants tailored to accommodate different scenarios. Beyond the fundamental CFO design, these include the two-dimensional CFO (2dCFO) designed for drug-combination trials, accumulative CFO (aCFO) for accruing all dose information, time-to-event CFO (TITE-CFO), and fractional CFO (fCFO) which are developed to specifically address late-onset toxicity. Moreover, hybrid designs such as TITE-aCFO and f-aCFO, which integrate both late-onset toxicity and all dose information for decision making, are also included.CFOprovides a robust set of functions used for determining subsequent cohort doses, selecting the MTD, and conducting simulations to evaluate design operating characteristics. The properties and results are presented to trial investigators through simple textual and graphical outputs. The user-friendly interface, adaptability to various design considerations, and the comprehensive implementation of CFO-type designs positionCFOas a noteworthy tool for phase I clinical trials.
Publisher
Cold Spring Harbor Laboratory