Using Bayesian proof-of-concept criteria to design randomized phase II oncology trials stratified by biomarker status.

Abstract

e21150 Background: In the early stage development of a targeted therapy, assumed to be more effective in a specific subpopulation, it is very challenging to decide whether clinical trials should be conducted in all patients or only in the specific subpopulation. We propose to use randomized Phase II study results to help make the decision about which population to use in a subsequent Phase III study. Traditional frequentist designs often use Type I/II error rates only to determine sample sizes and make decisions based on p-values. However, critical values needed for statistical significance are often implicit and a significant p-value does not necessarily imply a clinically meaningful effect size. Therefore, we propose some Bayesian proof-of-concept (PoC) criteria with clearly defined effect size criteria. Methods: Bayesian PoC criteria are proposed for all patients, biomarker positive patients, and biomarker negative patients, respectively. These criteria require 1) a reasonably high probability that treatment is better than control and 2) the estimated hazard ratio is in favor of the treatment and reaches a clinically meaningful size. We then establish a decision rule based on the PoC criteria for the three study populations. Frequentist probabilities of making the right decision for various scenarios are investigated. Additionally, when the prevalence rate for positive biomarker is low, it may be desirable to enrich the Phase II study in biomarker positive patients, which requires adjusted estimates for the hazard ratios in the original population. Results: Simulations show that the proposed design has good frequentist operating characteristics. The proposed design has been applied in a new randomized Phase II trial in breast cancer. Conclusions: A new Phase II oncology design using Bayesian PoC criteria is proposed. This design is easy to interpret, has good operating characteristics, and can avoid some potential shortcomings in traditional designs.