Adaptive design strategies for ethical and efficient clinical trials using simulated annealing

We introduce a flexible Adaptive Design Strategy for treatment allocation in clinical trials based on a randomized simulated annealing algorithm. The proposed approach provides a unified and modular framework for implementing a broad range of adaptive randomization objectives, including covariate-adaptive (CA), response-adaptive (RA), covariate-adjusted response-adaptive (CARA), and hybrid RA+CA and CARA+CA designs. The procedure accommodates multiple treatments, different outcome types, and any baseline covariate information available. Its implementation is compatible with both parametric and nonparametric predictive models; in this work, we use Bayesian Additive Regression Trees to capture complex covariate-response relationships without committing to a low-dimensional parametric specification. The allocation rule combines a simulated-annealing-based recommendation with a fallback randomization step used to ensure positive assignment probabilities. In a finite stratified setting, we establish strong consistency of the resulting stratified sample mean estimators. Extensive simulation studies with homogeneous and heterogeneous treatment effects show that our proposal achieves a favorable trade-off between ethical allocation and statistical efficiency. A redesign of the ACTG 175 clinical trial further illustrates the practical advantages and flexibility of the proposed framework.