Adaptive Bayesian Design for Phase I Dose-Finding Trials Using a Joint Model of Response and Toxicity

Abstract

We present a new adaptive Bayesian method for dose-finding in phase I clinical trials based on both response and toxicity. Although clinical responses are usually rare in phase I cancer trials, molecularly targeted therapy may make clinical responses more likely. In addition, biological responses may be common. Thus responses may be frequent enough to help decide how aggressive a phase I escalation should be. The model assumes that response and toxicity events happen depending on respective dose thresholds for the individual, assuming that the thresholds jointly follow a bivariate log-normal distribution or a mixture. The design utilizes prior information about the population threshold distribution as well as accumulated data. The next dose is assigned to maximize a patient-oriented expected utility integrated over the current posterior distribution. The design is evaluated through simulation with population parameters equaling estimates from early Gleevec trials. This exercise provides evidence for the value of the use of the proposed design for future clinical trials.

Key Words:

• Adaptive Bayesian design
• Log-bivariate normal distribution
• Phase I trials
• Threshold

Notes

Note. The six scenarios are combinations of correlation ρ and proportion p. The estimate of OPT has good accuracy (mean close to true OPT and small SD).

Note. The proportions of outcomes from the simulated standard MTD design differed from those of the original paper because the design used in the original paper was somewhat modified from the standard MTD design. The percentages of rt and RT are higher under the assumption of higher risk of toxicity (line B) compared to the simulations based on the toxicity model estimated from data (line A).