1206 - Clinical Implementation of an AI-Assisted Workflow for Systematic Clinical Trial Eligibility Prescreening
Presenter(s)
J. T. Rosenthal1, E. Hahesy2, S. Chalise3, D. Pazgan-Lorenzo4, D. Rose4, B. A. Mueller5, S. N. Powell5, Z. Zhang6, M. Zhu3, M. R. Sabuncu7, A. Li4, and L. Z. Braunstein3; 1Weill Cornell Medical College, New York, NY, 2Memorial Sloan Kettering Cancer Center, New York, NY, United States, 3Memorial Sloan Kettering Cancer Center, New York, NY, 4Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, 5Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, 6Department of Epidemiology and Biostatistics, Memorial Sloan Kettering Cancer Center, New York, NY, 7Cornell Tech, New York, NY
Purpose/Objective(s): Determining patient eligibility for clinical trials is a complex task that requires retrieving and synthesizing information from heterogeneous longitudinal clinical documents. This labor-intensive, manual process precludes systematic screening, leading to missed opportunities to offer eligible patients the option to participate in trials and to low overall trial accrual rates. We hypothesized that an AI-assisted workflow could improve eligibility determination efficiency, enabling the implementation of a broad prescreening program in a high-volume breast radiation oncology clinical service.
Materials/Methods: We previously developed and retrospectively validated a retrieval-augmented large language model (LLM) system for determining clinical trial eligibility. To implement this AI-integrated workflow into clinical practice, we engineered 3 components: 1) smart job scheduling to programmatically fetch appointment schedules, identify upcoming new patient visits, and trigger workflows; 2) a user-friendly application to facilitate human-in-the-loop review by a clinical research coordinator (CRC) for verification, adjudication, and oversight of AI outputs; and 3) delivery of results to clinicians via a personalized report in advance of upcoming clinics. All data processing, including LLMs, was conducted on secure, scalable, institutionally approved HIPAA-compliant infrastructure and services.
Results: A 4-week implementation period was conducted in January 2026. 382 patients across 25 clinicians were screened for 7 actively accruing therapeutic trials, for a total of 2674 patient-trial pairs (“cases”). 869 cases (32.5%) flagged as potential matches underwent secondary review by a CRC, with a median review duration of 12.4 seconds using our application (interquartile range: 8.7–21.7). A total of 72 protocol-eligible patients were identified and reported to the attending radiation oncologists in advance of their scheduled visits. Usage costs for LLMs averaged $1.10 per case.
Conclusion: These results demonstrate the feasibility of integrating AI into radiation oncology clinical workflows to enable broad prescreening of clinical trial eligibility. A human-in-the-loop AI workflow increases the efficiency of eligibility determination, enabling a single CRC to run systematic trial screening for an entire clinical service. Future work will prospectively evaluate whether this workflow intervention translates to increased trial accrual rates.