A Bayesian hierarchal modeling approach to shortening phase I/II trials of anticancer drug combinations.

Abstract

In phase I/II anticancer drug-combination trials, trial design to evaluate toxicity and efficacy has been studied by dividing the trial into 2 stages, followed by seamless execution of the 2 stages. In the first stage, admissible dose combinations in toxicity are identified, followed by patient assignment among the identified admissible dose combinations using adaptive randomization in the second stage. When patients are assigned using adaptive randomization, it is desirable to determine a more appropriate dose combination by taking into consideration both drug efficacy and toxicity; however, during the course of this determination and evaluation of toxicity and efficacy, there remains a concern that the trial duration might be prolonged. Therefore, we proposed a trial design to assign patients adaptively to more appropriate dose combinations in both toxicity and efficacy and to shorten trial duration without compromising trial performance. When selecting the dose combination for subsequent cohorts, unobserved data are treated as missing data, which are imputed using a data augmentation algorithm involving a gamma process. Probabilities associated with toxicity and efficacy are estimated applying a Bayesian hierarchical model to the imputed data, thereby allowing more patients to be assigned more appropriate dose combinations in both toxicity and efficacy through adaptive randomization. Results of simulation studies suggested that the proposed approach shortened trial duration without significantly compromising the performance of the trial as compared with existing approaches. We believe that the proposed approach will expedite drug development time and reduce costs associated with clinical development.