A simulation study evaluating clinical trial designs for the Cilostazol Albumin Treatment in Subarachnoid Hemorrhage trial found that a Bayesian adaptive Phase II design increased the probability of identifying successful treatments while reducing expected patient enrollment compared to fixed equal allocation. Published in BMC Medical Research Methodology, the investigation assessed an adaptive regimen-selection framework intended to evaluate eight active combinations of cilostazol and human albumin alongside a concurrent control in patients diagnosed with aneurysmal subarachnoid hemorrhage.
The simulated clinical investigation focused on a primary binary endpoint defined as delayed cerebral ischemia occurring within fourteen days, measured in accordance with established Common Data Elements standards. The selected trial framework integrated multiple adaptive components, including protected control allocation to ensure adequate concurrent comparator data, response-adaptive allocation to assign more patients to promising regimens, predetermined interim stopping rules, and information borrowing across treatment combinations using a two-dimensional normal dynamic linear model.
To establish baseline operating characteristics, the investigators generated one thousand simulated clinical trials per evaluation scenario using the Fixed and Adaptive Clinical Trial Simulator version 7.0.0. In addition, reproducible independent-arm simulations utilizing five thousand trials per scenario—and expanding to fifty thousand trials per scenario for recalibrated allocation assessments—were conducted to explore specific comparison strategies and evaluate sensitivity to missing outcome data, varying control risk levels, burn-in periods, and omitted candidate regimens. Throughout these analyses, the researchers formally distinguished the overall probability of reaching any positive Go decision from the statistical power required to identify a truly beneficial regimen.
In primary operating evaluations, the selected adaptive design demonstrated controlled error rates alongside substantial gains in decision performance. Under global-null conditions where no intervention was effective, the estimated family-wise type I error rate was 0.104 for the adaptive design, closely matching the 0.100 benchmark observed under a standard fixed equal-allocation design. When active treatments were beneficial, the adaptive design achieved notable advantages in treatment identification over fixed designs.
Across specific treatment gradient scenarios where all active combinations provided therapeutic benefit, the probability of reaching a Go decision—which simultaneously represented statistical power to detect an effective intervention—reached 0.858 under an albumin-favored gradient and 0.838 under a cilostazol-favored gradient. By comparison, fixed equal-allocation strategies yielded Go probabilities of only 0.610 and 0.600 under identical respective gradients. Furthermore, the adaptive protocol lowered expected patient enrollment to 308.3 individuals in the albumin-favored scenario and 307.9 individuals in the cilostazol-favored scenario, compared to a required total of 400.0 individuals under the fixed-allocation benchmark.
In parallel independent-arm simulations that omitted dynamic model borrowing, the trial’s Statistical Analysis Plan allocation rule increased statistical power from 0.697 to 0.777 in the albumin-favored gradient and from 0.703 to 0.780 in the cilostazol-favored gradient relative to fixed allocation under identical decision rules calibrated to a ten percent global-null target. However, sensitivity analyses highlighted vulnerabilities associated with incomplete follow-up, as introducing a ten percent rate of missing outcome data reduced statistical power from 0.750 to 0.707 within the albumin-favored gradient.
The simulations also revealed critical practical risks regarding candidate regimen selection, demonstrating that reduced-arm designs entirely failed to detect a beneficial monotherapy if that single agent was omitted from the starting candidate set. The authors concluded that while the selected adaptive architecture improved Go probability and reduced participant exposure without inflating family-wise type I error, removing monotherapy arms could exclude the only effective intervention. Crucially, the researchers cautioned that these simulation comparisons do not prove superiority over broader classes of standard adaptive designs, nor do they isolate the independent statistical contribution of two-dimensional model borrowing.
Subject of Research: Bayesian adaptive Phase II clinical trial design simulations for aneurysmal subarachnoid hemorrhage
Article Title: Simulation-guided selection of a Bayesian adaptive Phase II design for a nine-arm cilostazol-albumin trial in aneurysmal subarachnoid hemorrhage
Article References: Qureshi, A. I., Raza, H., Alam, N., Beall, J., Gajewski, B. J., Martin, R. H., & Suarez, J. I. (2026). Simulation-guided selection of a Bayesian adaptive Phase II design for a nine-arm cilostazol-albumin trial in aneurysmal subarachnoid hemorrhage. BMC Medical Research Methodology. https://doi.org/10.1186/s12874-026-03080-z
Image Credits: AI Generated
DOI: 10.1186/s12874-026-03080-z
Keywords: aneurysmal subarachnoid hemorrhage, cilostazol, human albumin, Bayesian adaptive design, Phase II clinical trial, CATS trial, delayed cerebral ischemia, Common Data Elements, response-adaptive allocation, normal dynamic linear model, model borrowing, FACTS simulator, type I error rate, interim stopping, expected enrollment, independent-arm simulations, monotherapy arms, regimen selection
News Source: Ophelia Keating. (November 18, 2026). Adaptive Design Improves Regimen Selection in Simulated Brain Hemorrhage Trial. Scienmag.



