Predicting long-term allograft outcomes in kidney transplant recipients using a machine learning approach: a 5-year retrospective cohort study.

Long-term outcomes of kidney allografts vary significantly among deceased donor kidney transplant recipients, and current prediction tools struggle to integrate comprehensive pre- and post-transplant factors. Extended longitudinal follow-up data beyond five years remains particularly scarce in kidney transplantation research despite being crucial for understanding true long-term outcomes. To address this, we developed and validated machine learning models to predict 5-year allograft survival using a distinctive cohort of 940 adult deceased donor kidney transplantation recipients with extended follow-up exceeding 5 years. Two predictive models were developed: a pre-transplant model (Kidney Allograft Prediction of Transplant Outcome Risk, KAPTOR-pre) using pre-transplant donor-recipient matching data, and a 1-year landmark conditional prediction model (KAPTOR-full) incorporating both pre- and post-transplant parameters, pathological data, and laboratory markers from the first year. KAPTOR-full achieved excellent discrimination with area under the receiver operating characteristic of 0.904, while KAPTOR-pre performed well at 0.813. In internal validation, both models showed higher C-index and improved risk stratification compared with established prognostic tools including KDPI. The extended follow-up period allowed internal assessment of model performance for 5-year outcomes. Ultimately, our models integrating routine clinical variables demonstrated excellent predictive performance for long-term graft survival. While the pre-transplant model achieved good discrimination, the addition of first-year post-transplant data significantly enhanced predictive accuracy. Both models outperformed existing tools in internal validation and may support personalized risk assessment, pending independent multicenter validation.