Tumor-Specific Decisions Using Tumor-Agnostic Evidence from Basket Trials: A Bayesian Hierarchical Approach

PurposeTreatment effect heterogeneity across tumor types remains a challenge to evidence interpretation and implementation of tumor-agnostic drugs (TADs), which are typically approved based on basket trial evidence. We sought to use Bayesian hierarchical models (BHM) to assess heterogeneity and improve estimates of tumor-specific treatment outcomes, which are crucial for healthcare decision-making. MethodsWe fitted BHMs and Bayesian fixed-effect models to evaluate the objective response rate (ORR), the median progression-free survival (mPFS), and the overall survival (mOS). We estimated the posterior distribution of outcomes for each tumor type, the pooled effects, and intra-class correlations (ICC). Using published basket trial evidence for pembrolizumab (KEYNOTE-158/KEYNOTE-164), we obtained the predictive outcomes in a new cancer type drawn from the same population. In the base case, we assumed non-informative priors with uniform distributions for between-tumor standard deviation. We performed sensitivity analyses with various priors to account for uncertainty in the prior specification. ResultsThe BHMs shrunk the original tumor-specific estimates toward a pooled treatment effect. The borrowing of information across tumor types resulted in less variability in the posterior tumor-specific estimates compared to the original trial estimates, reflected in narrower 95% credible intervals (CrLs). We found low heterogeneity for ORR but high heterogeneity for mPFS and mOS across cancers (ICC: 0.22, 0.87, 0.7). The predicted posterior means and 95%CrLs were 0.37 (0.15-0.64) for ORR, 3.75 months (0.24-50.45) for mPFS, and 13.76 months (0.42-276.49) for mOS, respectively. ConclusionsBorrowing information through BHM can improve the precision of tumor-specific estimates, thereby facilitating more robust policy decisions regarding TADs. Our analysis revealed high heterogeneity and uncertainty in survival endpoints. Both pooled and tumor-specific estimates are informative for clinical and coverage decision making. HighlightsO_LIBayesian hierarchical models could enhance precision and reduce uncertainty of estimates derived from basket trial evidence, potentially improving confidence in tumor-agnostic decision making, despite small sample sizes in some tumor types. C_LIO_LIOur study highlights high variability in treatment effects of pembrolizumab across tumor types with respect to survival endpoints, although treatment effects appear more consistent when judged by objective response rate at approval. Understanding heterogeneity in treatment effects following accelerated approvals based on surrogate endpoint is crucial for clinical and coverage decision making. C_LIO_LIThis article demonstrates the use of Bayesian methods to estimate posterior distributions of tumor-specific and aggregated treatment effects (ORR, median PFS, and median OS) from basket trials. Choosing between fixed-effect or random-effects model to evaluate pooled treatment effects depends on the level of heterogeneity in effect sizes across tumor types. C_LI