The design of complex, adaptive clinical trials often involves the selection of a design from a large number of potential designs. For example, designs with different interim analysis timings or randomization schemes may be considered. To choose an optimal design, candidate designs are simulated and compared using operating characteristics.
However, the optimization process can be complicated. When the number of candidates is large, the optimization may be limited to searching over a small number of designs. Second, individual components of the design may interact in unexpected ways, making stepwise selection of parameters difficult. Finally, individual simulations may be computationally demanding, limiting the accuracy of design comparisons.
This paper applies the principles of Bayesian optimization to the design of clinical trials in order to efficiently automate the selection of a design. Several examples highlight the advantage of this approach over a “hand tuning” design process in terms of its ability to explore the design space in a structured manner while minimizing computational burden. Practical considerations for implementing this procedure are discussed.
|