Design of a prospective implementation study to evaluate the efficacy of an AI-assisted workflow intervention to increase breast cancer clinical trial participation

We previously developed and retrospectively validated an AI-integrated workflow for conducting rapid, accurate, and cost-effective automated clinical trial eligibility prescreening, but it remains unknown whether this tool can increase trial accrual rates when implemented in real-world clinical workflows. To evaluate its impact on practice, this prospective, controlled implementation study will use a periodic on-off design in which participating clinicians alternate monthly between the AI-assisted workflow and usual care. During intervention periods, the Memorial Sloan Kettering Multi-Agent Trial Coordination Hub (MSK-MATCH) system will conduct automated eligibility prescreening for all upcoming new patient visits. For cases triaged by the AI system for secondary human review, a clinical research coordinator (CRC) will use a secure web interface to verify predictions and resolve ambiguities. The results of this human-in-the-loop eligibility prescreening will be compiled into a single summary report delivered directly to each attending physician in advance of their scheduled clinic each week. The primary endpoint is the rate of accrual to any breast radiation oncology therapeutic clinical trial within 60 days of initial visit. Secondary endpoints include (1) the number of patient-trial pairs evaluated for trial eligibility, and (2) the mean cost of all screening activities per screened pair. Statistical analysis will use a two-sided paired t-test comparing intervention and control periods within each clinician. At the conclusion of the study, semi-structured interviews will be conducted to collect feedback and assess perceptions of this AI workflow intervention among participating clinicians and CRCs. This study is planned to begin in January 2026 in the breast radiation oncology service at Memorial Sloan Kettering Cancer Center. With an enrollment target of 2,500 new patient visits, the design provides 80% power to detect a minimum absolute difference in accrual rate of 1.57% with a type I error rate alpha = 0.05. An interim analysis for efficacy and futility is planned after 1,250 visits, anticipated in June 2026. This quality-improvement study is IRB-exempt, and waivers have been added to all trial protocols to cover screening activities. Findings from this study will provide evidence to guide the responsible and effective adoption of AI-based workflow interventions in clinical oncology.