Abstract
BACKGROUND: Pancreas non-utilization affects 25%-30% of deceased donor offers in the US. The European-derived Pre-Procurement Pancreas Suitability Score (P-PASS) has performed inconsistently, is not validated in North America, and requires less accessible inputs, highlighting need for a robust prediction tool for non-utilization. METHODS: We developed a predictive model for pancreas non-utilization using procured offers from the US SRTR (2003-2023). A multivariable logistic model used variables selected by causal diagram, LASSO, and univariable analyses. Model performance was evaluated using discrimination (AUC), calibration, and net reclassification improvement (NRI) versus a modified P-PASS. Internal validation used 1000 bootstrap resamples. Sensitivity analyses included recent-era offers, an all-comer cohort including non-procured offers, and excluding non-donor-related non-utilizations. RESULTS: Among 30 757 pancreas offers, 26% were non-utilized. The model incorporated ten factors: older age, male sex, higher BMI, hypertension, gastrointestinal disease, smoking, donation after cardiac death, stroke as cause of death, elevated terminal creatinine, and abnormal lipase. This model achieved moderate discrimination (AUC 0.699; optimism-corrected 0.698), accurate calibration, Brier score 0.090, and 9.2% NRI. Sensitivity analyses confirmed robust performance, including AUC 0.8 among all-comer offers. CONCLUSIONS: This model provides a practical, data-driven tool that improves identification of high-risk pancreata and may support more consistent, efficient utilization decisions. By demonstrating modifiable center-level variation and outperforming P-PASS in its first North American external validation, our findings offer a strong evidence base for targeted quality-improvement and policy initiatives to enhance pancreas transplant activity.