AI-enabled clinical trial framework for end-to-end site recommendation and enrollment projection.

11022 Background: Reliable accrual forecasting is essential in oncology trials because delayed enrollment reduces statistical power, increases costs, and prolongs timelines. We developed an AI-enabled framework integrating site recommendations with probabilistic enrollment modeling to support data-driven trial planning. Methods: Site-level attributes (Site360) and curated real-world patient datasets (Patient360) were integrated as inputs. Trial objectives including enrollment targets, timelines, sponsor and diversity requirements, and screen failure rates (SFR) were user specified. Candidate sites were ranked using a multi- criteria TOPSIS engine optimized with simulated annealing to identify portfolios meeting protocol requirements. Recommended site attributes informed Monte Carlo enrollment simulations incorporating site activation, patient arrival, and capacity constraints. Each underwent 1000 simulations, generating accrual distributions. Fifty-four CRC, NSCLC, and Breast trials from ClinicalTrials.gov were evaluated. Planned site counts and enrollment intensity (patients per month per planned site; ppmpps) from the same trials served as benchmarks. Enrollment intensity was calculated as enrolled patients divided by (trial duration in months × planned sites), with duration estimated from trial start to primary completion. Scenarios tested SFR 20%, 40%, and 60%. Model outputs were compared with planned benchmarks using Spearman correlation and summarized with medians and IQR. Results: Trials showed heterogeneous planned site counts: CRC median 76 (IQR: 43-119), Breast 204 (114-279), NSCLC 178 (101-227). Model recommended site counts showed moderate to strong rank agreement with planned sites across indications. CRC correlations increased from 0.67 to 0.73 as SFR rose from 20% to 60% (all p < 0.05). Breast correlations ranged 0.49-0.54 and NSCLC 0.61-0.63 across SFRs all with p < 0.05. Higher SFR required more recommended sites, compensating for screen failures. Predicted ppmpps aligned strongly with benchmarks. For Breast, predicted ppmpps decreased from 0.15 (0.10-0.31) at 20% SFR to 0.07 (0.04-0.16) at 60% SFR versus planned 0.08 (0.03-0.17), with ppmpps correlations 0.82, 0.77, 0.66 respectively (all p < 0.01). CRC predicted rates were 0.13 (0.08-0.29), 0.10 (0.07-0.24), 0.07 (0.05-0.18) across SFRs versus planned 0.09 (0.06-0.13), correlations 0.47 (p = 0.076), 0.54, 0.59 (both p < 0.05). NSCLC showed 0.23 (0.15-0.27), 0.18 (0.13-0.22), 0.12 (0.08-0.18) versus planned 0.12 (0.07-0.19) with correlations 0.54-0.64 (p < 0.02). Conclusions: This AI-enabled framework integrates optimization and probabilistic modeling to recommend sites and project enrollment behavior. Rank agreement with benchmarks supports improved site feasibility assessment by identifying trials with higher site requirements and enrollment risk and enabling risk-adjusted planning.