Adaptive randomized designs for cancer clinical trials by using integer algorithms and exact Monte Carlo methods

Project Summary/Abstract Current phase II clinical trial designs for cancer studies are generally not flexible and effective enough to reduce sample size and costs. Unlike traditional single-arm two-stage designs, adaptive designs allow a study to be modified with the information observed from previous stages. Recently, a few adaptive designs were developed

for phase II cancer clinical trials with binary endpoints, and the majority of them cannot be directly applied in practice because of a counter-intuitive feature of the relationship between sample size and the number of responses from previous stages. We developed a new single-arm two-stage design that corrects that counter-

intuitive feature of the study design. These adaptive designs were all developed for single-arm studies. In Aim 1, we will use efficient integer algorithms along with exact Monte Carlo simulation methods to develop adaptive randomized two-arm designs for cancer clinical trials. The proposed adaptive randomized designs are expected

to save between 10% to 35% sample sizes as compared to the conventional group sequential designs. Unlike the existing adaptive randomized designs minimizing expected treatment failures, we will develop the first adaptive randomized designs with the objective to minimize expected sample size. For the existing adaptive

single-arm design using integer algorithms without importance sampling, it could take a few months by using a stand-alone computer, and a few days using a supercomputer. With multiple arms in a study, it would be very computationally intensive. The goal of Aim 3 is to reduce the computation time to no more than 30 minutes

by utilizing importance sampling and integer algorithms on a stand-alone computer. The traditionally used importance sampling does not guarantee the type I error rate and power. For this reason, we will utilize the recently developed exact importance sampling method to guarantee type I error rate and power. A combination

of integer algorithms and importance sampling will be able to reduce the computation time to no more than 30 minutes for the proposed adaptive designs. In addition to new adaptive design development, we will also develop proper statistical inference for adaptive two-stage clinical trials in Aim 2. The existing exact approaches

from commercial software for statistical inference are often based on the conditional framework, by assuming both marginal totals fixed. Such exact conditional approaches are not aligned with the study design for a clinical trial which often only assumes the sample size of each arm fixed, not the total responses. The proposed exact

statistical inferences are proper by considering the nature of adaptive designs with multiple stages and sample size change. Ultimately, we will develop adaptive randomized designs for phase II cancer studies with binary endpoints with the smallest expected sample size. The proposed designs will be available for public use through

a new R package and a new website that will use a powerful supercomputer. Upon completion of this project, our school will take over the cost of maintenance of the software developed from this proposal.