Predicting first-year kidney graft loss at the time of donor offer using machine learning and competing risks in a Latin American cohort

Abstract Introduction Kidney transplantation faces organ shortages, underscoring the need for early risk stratification of graft loss. We developed and validated a 12-month prediction model that treats death as a competing event. Methods We conducted a retrospective cohort study of 2030 adult kidney transplant recipients (2008–2023) from Colombia’s largest transplant network. Models included a Random Survival Forest for Competing Risks (RSF-CR) and Fine-Gray (FG) regression. Internal validation used stratified cross-validation. Model performance was evaluated via discrimination (C-index), calibration, and clinical utility (decision curve analysis). Results Key predictors included: Donor type, Stroke cause of death (Deceased donor), Recipient age, Donor creatinine, PRA I > 20, Expanded criteria donor, Donor age, Years on dialysis, PRA II > 20, Donor hypertension, DR compatibility, Retransplantation. The RSF-CR model outperformed the FG, achieving a C-index of 0.87 (vs. 0.72) and high sensitivity (88%). It accurately identified low-risk candidates (NPV: 98%) and showed a positive net benefit. Conclusion We developed and validated a predictive model for first-year graft loss in kidney transplant recipients using a machine learning for competing risks model. The model showed strong discriminative ability and moderate calibration. Further temporal validation in our population and external validation in other clinical contexts is required to ensure its applicability.