gms | German Medical Science

Introduction: In the wake of advancements in personalized medicine, custom-tailored interventions are on the rise. By incorporating patients’ unique characteristics, they enable superior treatment options matching patients’ needs - a trend especially prevalent in oncology. These advancements in personalized medicine pose new challenges for clinical trials and the statistical methods employed.

Basket trials address some of those challenges: In these trials patients from different indications are recruited, with shared biomarkers as a common denominator. The resulting patient subgroups are called „baskets” [1]. Information is then borrowed between baskets if they are sufficiently similar, lessening the impact of low patient numbers on power.

Among numerous frequentist and Bayesian options, the basket trial design proposed by Fujikawa et al. [2] offers a variety of advantages: It foregoes complex hierarchical modelling used in early basket-designs [3] and instead uses pairwise weights to determine the borrowing of information between baskets. It presents a flexible and intuitive Bayesian framework and, due to its conjugated beta-binomial setting, is computationally inexpensive.

In Fujikawa’s design, the pairwise weights used for the pivotal borrowing of information between the baskets are, in addition to a measure of similarity, determined by two prespecified tuning parameters. However, the obscure influence of the tuning parameters on the weights and the reliance on similarity measures could lead to suboptimal choice and according ramifications for the analysis. The aim of the project is therefore to explore a more flexible and transparent method to determine these weights.

Methods: We propose determining these pairwise weights by prespecifying knots and using monotonic splines to interpolate between these knots, hereby replacing Fujikawa’s weights based on the Kullback-Leibler divergence and avoiding uninterpretable hyperparameters. This enables practitioners to directly fine-tune a weighting function which is not reliant on similarity measures.

In order to explore the spline-based method, a simulation study is carried out. Using two knots for maximal and minimal heterogeneity in the response as a foundation, more knots are added to further customize the function which assigns the weights while controlling for the number of baskets and the sample size per basket.

A concluding, comparative simulation study evaluates the performance of the spline-based method for determining the weights-function against the original method using hyperparameters.

Results: The spline-based method is expected to be at least on par with Fujikawa’s original method, while offering substantial adaptability and transparency regarding the function determining the weights.

Discussion/conclusion: We show that choosing weights using monotonic splines leads to flexible borrowing and easily interpretable tuning parameters. This expands Fujikawa’s design by facilitating researchers to make informed decisions about the weighting function used.

The authors declare that they have no competing interests.

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