A Cross-Fitted Graph-Fused Borrowing Analysis of Basket Trials

Basket trials in oncology evaluate a single treatment across multiple cancer types or subtypes that share a common genetic or molecular aberration. Statistical methods that borrow information across baskets often yield more reliable inference than simple pooling or independent analyses. However, many existing analysis methods rely on prior distributions or hyperparameters whose specification may be subjective or difficult to interpret in some clinical contexts. The resulting inferences may depend on these modeling choices, which can complicate interpretation or communication in certain settings. In this paper, we propose a cross-fitted graph-fused borrowing approach that produces basket-specific p-values for analyzing basket trial data. Our method combines (i) sample splitting and cross-fitting to separate model selection from inference, (ii) data-driven discovery of basket similarity using graph-fused penalties, and (iii) selective and resampling-based safeguards to ensure valid type I error control after adaptive borrowing. Simulation studies demonstrate that the proposed method exhibits satisfactory operating characteristics across a wide range of scenarios. We further illustrate the practical utility of our method through applications to two oncology basket trials involving imatinib and vemurafenib, respectively.