Machine learning-based risk stratification of early graft failure in simultaneous pancreas-kidney transplantation.

Simultaneous pancreas–kidney transplantation (SPK) restores insulin independence for patients with diabetes and end-stage renal disease, yet early graft failure remains a major obstacle. We analyzed 6,725 adult SPK procedures recorded by UNOS between 2014 and 2024 to build a peri-operative risk model using Random Survival Forests (RSF). After preprocessing, MissForest imputation and LASSO screening, 21 predictors were retained. The RSF, trained on 80% of the data and tested on the remainder, achieved a Harrell C-index of 0.73 and a Uno C-index of 0.58; time-dependent AUCs were 0.75 and 0.74 at one and two years, with Brier scores of 0.075 and 0.090. Maximizing Youden’s J produced a risk-score threshold of 105.24 that separated recipients into high- and low-risk strata with strongly divergent Kaplan–Meier curves (log-rank p = 2.4e−16). Decision-curve analysis demonstrated net clinical benefit across threshold probabilities 0–0.25 and, at a 10% threshold, predicted 41 unnecessary interventions avoided per 100 patients relative to treating all. Pre-transplant insulin status, cold-ischemia time and recipient age dominated variable importance. A compact RSF based on routinely collected peri-operative data provided internally validated graft-failure risk stratification that may support peri-operative risk assessment and targeted surveillance. Prospective multi-center validation is required before clinical implementation.