Bayesian Hierarchical Mixture Models for Adaptive Basket Trial Design

The purpose of an oncology basket trial is to test the effectiveness of a single drug for different cancer types with the same somatic mutations. The goal is to determine whether the drug under investigation is effective for all cancer types or whether it specifically targets certain types. Studies have revealed variation in the response rate across different baskets, but some baskets with the same genetic variant may have similar response rates. Therefore, studies have mainly focused on measuring heterogeneity among baskets to control the strength of information-sharing. However, this is not the only approach to designing a basket trial. In phase II trial design, determining which baskets respond effectively to treatment is more vital than focusing solely on similarity in treatment efficacy among baskets. Therefore, in this article, we propose a data-driven Bayesian hierarchical mixture model in which data are incorporated to classify the prior in the model. The classification rule separates baskets into classes, thus, facilitating the sharing of information among baskets in the same class. This study provides an alternative perspective for developing information-sharing tools for basket trials that do not solely depend on interbasket similarity when binary outcomes and response rates are interesting. This study also demonstrates the advantages of the developed model in comparison with other popular model designs.