Bayesian borrowing approaches in the context of clinical trials for rare diseases: The example of slow-flow vascular malformations

Details

Supervisors

Legrand, Catherine

Faculty

Faculté des sciences

Degree label

Master [120] en statistique, orientation biostatistiques, à finalité spécialisée

Abstract

enBackground: Clinical trials of rare diseases are often limited by small sample sizes, which restrict the precision and reliability of treatment effect estimates. Slow-flow vascular malformations, as rare and clinically heterogeneous disorders, pose significant challenges for clinical trials because of limited patient populations, variable disease presentation, and diverse treatment responses. In this context, Bayesian borrowing approaches provide a principled framework for incorporating external or historical data with current study data. These methods include both prior-based approaches, which explicitly model prior information, and likelihood-based hierarchical models, which borrow information through partial pooling across studies. Methods: An exploratory simulation study was conducted based on two realizations of simulated datasets, representing a large (n = 100) and a small (n = 20) current studies, with 10 historical studies exhibiting varying degrees of similarity. These scenarios were designed to assess borrowing behavior under different levels of similarity. In addition, a real-world analysis was performed using 19 historical studies extracted from the literature to reinforce the estimation of treatment effects from a study evaluating the effect of sirolimus in patients with slow-flow vascular malformations (our current study), focusing on two endpoints: pain response after 12 months and D-dimer decrease after 12 months. A two-stage framework was applied, combining pre-borrowing weights, based on clinical, demographic, and methodological similarity with the current study, with four Bayesian borrowing methods: robust meta-analytic-predictive prior (rMAP), Self-adaptive mixture prior (SAM), elastic prior, and normal-normal hierarchical model (NNHM). Results: Simulation results suggest that all methods adapt the level of borrowing according to the degree of similarity between historical and current data, with increased effective sample size (ESS) and narrower intervals under high similarity levels, and reduced borrowing when the similarity decreases. In the real-world analysis, minimal borrowing was observed for the pain response endpoint, with posterior estimates remaining close to the current study estimate of 76.2%. For the D-dimer endpoint, all borrowing methods show an upward shift in the posterior mean relative to the current-only estimate. Moderate borrowing methods (rMAP and SAM) yield posterior means around 1.19–1.20, with credible intervals approximately 0.28–0.29 to 2.09–2.10, corresponding to moderate precision gains (relative ESS ≈ 14–16%). The Elastic models remain closer to the current-only estimate, with posterior means around 1.15 with credible intervals from 0.23-2.07 and smaller borrowing gains (relative ESS ≈ 10–11%). In contrast, NNHM produces the largest shift, with posterior means around 1.55 and credible intervals approximately 0.60 to 2.44, reflecting the strongest borrowing from historical data and the highest precision gains (relative ESS ≈ 19–24%). Conclusion: Bayesian dynamic borrowing methods provide a flexible framework for integrating historical data while adapting to the level of agreement between current and historical studies. The proposed two-stage approach provides a structured way to account for clinical relevance and methodological quality before integrating external data. Although borrowing led to modest gains in precision and small shifts in estimates, it did not substantially alter clinical conclusions. The recommendations for future work include performing large scale simulations with varying current study sample sizes and different levels of agreement between historical and current data.