Simulation-Guided Selection of a Bayesian Adaptive Phase II Design for a Nine-Arm Cilostazol-Albumin Trial in Aneurysmal Subarachnoid Hemorrhage

Background: The Cilostazol Albumin Treatment in Subarachnoid Hemorrhage (CATS) trial evaluates eight active cilostazol-human albumin regimens plus control in patients with aneurysmal subarachnoid hemorrhage. We summarized the rationale for the primary statistical design, compared alternative Phase II methodologies, and evaluated reduced-arm sensitivity scenarios. Methods: The binary primary endpoint is Common Data Elements-defined delayed cerebral ischemia within 14 days after randomization. The selected design is Bayesian adaptive, with a burn-in phase, response-adaptive randomization among active arms while maintaining fixed control allocation, four interim analyses, early stopping for expected success or futility, and a two-dimensional normal dynamic linear model. Primary operating characteristics were obtained from 1,000 virtual trials per scenario using Fixed and Adaptive Clinical Trial Simulator version 7.0.0. Exploratory simulations evaluated six-, four-, and two-active-arm configurations and simplified alternative designs. Results: Compared with fixed equal allocation, the Bayesian adaptive design preserved an approximately 10% false-success probability under the global null while improving probability of success and efficiency in clinically relevant scenarios. Under the Realistic scenario, probability of success increased from 0.61 to 0.86, expected sample size decreased from 400 to 308, and expected duration decreased from 235 to 187 weeks. Under common thresholds, null probability of success was 0.098 for the full anchor and 0.073 for Reduced-6; Reduced-6 probabilities of success were 0.774 and 0.765 in the Realistic and Realistic2 scenarios. However, Reduced-6 omitted two monotherapy anchors and was less robust in Backwards2. In the comparator simulation, the selected design had probability of success of 0.858 and expected sample size of 308.3 under the Realistic scenario, compared with 0.624 to 0.845 and approximately 352 to 400 for simplified comparators. Conclusions: For identifying the most promising cilostazol-human albumin regimen for Phase III rather than confirming efficacy, the Bayesian response-adaptive design with two-dimensional normal dynamic linear model borrowing is more efficient and better aligned than simplified comparators. The full nine-arm design remains preferable because it preserves the complete therapeutic discovery space and is more robust to misspecified or non-smooth response surfaces.