gms | German Medical Science

Background: Phase II clinical trials in oncology are conducted to assess whether a treatment shows promising effect in patients, so that further investigation may be recommended. Typically, such trials are designed as two-stage single-arm designs with a binary endpoint and interim decision at a pre-defined time-point during the trial whether to stop for futility. A simple and widespread method to plan and conduct such trials is Simon's two-stage design [1]. In this design, a pre-specified number of patients is treated in the first stage, and the trial only continues to the second stage if a pre-specified number of successful treatments was observed in first stage. However, the timespan until the primary outcomes for all patients in the first stage are observed and the stopping decision is made can be undesirably long. E.g. if 1-year survival is used as endpoint, it may take up to one year from last-patient-in in first stage to first-patient-in in second stage. Kunz et al. [2] proposed a design that uses a short-term endpoint to consider stopping for futility and analyses the original endpoint to evaluate efficacy at the end of the trial.

Methods: Since a critical requirement of this method is to assume the correlation between the short-term and the long-term endpoint, and since in clinical practice this assumption will usually be associated with high uncertainty, it seems natural to embed the method in the Bayesian framework. Further, if recruitment is rather slow, it is possible to estimate the correlation based on a few patients at the time point of interim analysis. Therefore, we implement a Bayesian model that estimates the correlation at time-point of interim analysis using all available information, including the information on the long-term endpoint. The interim decision whether to stop for futility is then guided by evaluating the predictive probability of success (PoS) of the trial. In this study, we evaluate from both frequentist and Bayesian perspective the impact of estimating the correlation and evaluating the PoS on the accuracy of interim decisions in phase II single-arm clinical trials.

Results: Our simulations show that it is possible to implement a Bayesian model such that it controls the frequentist type I error rate and improves the interim decisions under wrongly specified assumptions compared to standard approaches. However, this comes with a loss of power if the initial assumptions are correct. The performance of the method is better, if patient recruitment is slower since more long-term endpoint data is available at interim analysis.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.