Equity through efficiency: Automating patient to trial matching using AI and OncoLLM.

144 Background: Despite high patient interest in clinical trials, fewer than 10% of adult oncology patients enroll, often due to missed matching opportunities and labor-intensive screening processes. At the Medical College of Wisconsin (MCW) Cancer Center, we deployed an AI platform (Triomics PRISM) powered by a domain-specific large language model (OncoLLM) to prospectively screen every upcoming oncology visit against our active trial portfolio, with a focus on improving equitable access through scalable pre-screening. Methods: We conducted a prospective quality improvement initiative guided by the RE-AIM framework. In July 2024, PRISM was deployed across five disease-oriented teams (GI, GU, Breast, Thoracic, GYN Onc) to screen 100% of upcoming visits. The platform used structured and unstructured EHR data to evaluate eligibility against > 100 recruiting trials and generated match summaries for coordinators and physicians. Coordinators completed a 2-hour onboarding session and integrated the tool into routine workflows. We assessed RE-AIM domains: Reach (proportion of visits screened), Effectiveness (accrual changes, match accuracy), Adoption (daily use, coordinator engagement), Implementation (integration fidelity, time spent per review), and Maintenance (sustained coverage over time). Concordance between PRISM matches and final trial enrollment was used to assess accuracy. Results: Between July 1, 2024 and December 31, 2024, the platform automatically screened > 19,000 patient visits (~6,000/month), and reduced coordinator review time from 20–25 minutes to 3–12 minutes per patient. Among those enrolled during the study period, 72% were identified by PRISM prior to their clinic visit. Clinical trial accruals increased by 39% in Q3 and 27% in Q4 compared to the same periods in 2023. Importantly, trial matches showed > 95% concordance with final enrollment choices, affirming accuracy. Coordinators demonstrated enthusiastic adoption, with several completing hundreds of deep trial matches, suggesting sustained and meaningful engagement. Monthly platform coverage has remained consistently high, with > 97% of all upcoming patients reviewed every month since deployment. The AI-supported workflow also surfaced trial access opportunities in community patients and newly referred individuals, enabling more equitable screening. Conclusions: A scalable OncoLLM-based solution enabled comprehensive trial screening across multiple oncology indications, improving patient-trial matching speed, increasing accrual, and ensuring access for patients previously missed by manual approaches. Our deployment offers a reproducible model for increasing equity in clinical trial access by reducing reliance on variable human screening capacity and transforming the trial-matching process into an inclusive, data-driven system.