Efficient site selection for clinical trials using simulated annealing.

person Ian Ruchlin ConcertAI, Cambridge, MA info_outline Ian Ruchlin, Vibhor Govil, Deepak Samar, Jitesh Chawla, Pyeush Gurha

Authors person Ian Ruchlin ConcertAI, Cambridge, MA info_outline Ian Ruchlin, Vibhor Govil, Deepak Samar, Jitesh Chawla, Pyeush Gurha Organizations ConcertAI, Cambridge, MA Abstract Disclosures Research Funding ConcertAI Background: Identifying sites with eligible patients for multi-site oncology clinical trials is challenging given increasingly complex inclusion/exclusion criteria, limited site resources, and the need to ensure diversity in trials. This is a complex systems problem with many factors that present nonlinear behaviors. Given manifold combinations of sites, research capabilities, and study characteristics, artificial intelligence solutions may help research sites and sponsors optimize site selection to achieve enrollment, speed, and diversity of goals. We report on one of the first deployments of a multi-AI model solution to optimize site selection for an ongoing clinical trial. Methods: The clinical trials were translated into a standardized digital form expressed as rules fully encompassing the inclusion and exclusion criteria. A patient cohort was created from the criteria using the ConcertAI oncology research database, including site affiliations. The sites had numerical features specific to the protocol indication, such as physician count and projected patient count. Each site was scored by its features with the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) to create higher alignment to the best matched sites. Feature weights allowed TOPSIS to emphasize desirable attributes in an interpretable way. A simulated annealing algorithm, adept with nonlinear objectives, explored site combinations, yielding a high-scoring, constraint satisfying solution. Independent simulations measured convergence and provided a distribution of solutions. Shapley analysis clarified site contributions to score and constraints. The implementation was compiled and parallelized. Results: Utilizing a Multiple Myeloma clinical trial protocol with 7 criteria, we formed a cohort of 15,881 patients across 1,246 US clinical trial sites. Each site had 12 features. The weights optimized patient diversity while enrolling at least 10 patients per month across all sites to achieve the targeted study accrual timeline. The result was a set of 3 sites enrolling 12 patients per month with 51% diversity, in agreement with the optimal solution obtained through detailed, multi-week, manual analysis. As constraints increased, and exact verification became impractical, convergence was gauged by decreasing variance in independent solutions. Conclusions: Our study introduced a robust, efficient, and interpretable approach to clinical trial site selection. It addressed the challenges posed by increasingly complicated protocols, requirements for study participant diversity, and heterogeneous site features. The multi-AI model scalability underscored its practical utility in streamlining the site selection process for clinical research networks, academic centers, industry sponsored trials, cooperative group studies, and investigator-initiated studies.