Abstract
Background/Objectives: Machine-learning (ML) models are promoted for predicting delayed graft function (DGF) and graft survival after kidney transplantation; we examined whether they outperform regression and how well either has been validated. Methods: We searched PubMed/MEDLINE, Embase, and Web of Science through 28 June 2026, searched citations for named DGF models, extracted data in duplicate, assessed risk of bias using the Prediction Model Risk of Bias ASsessment Tool (PROBAST), and pooled only external validations of the same published model. Results: We included 217 reports (145 on DGF and 91 on graft survival, 19 on both); across seven external cohorts, the Irish 2010 DGF model had a pooled area under the curve (AUC) of 0.73 (95% confidence interval 0.69–0.77), the median difference in AUC or C-statistic between ML and regression in 28 same-cohort comparisons was 0.01 (interquartile range −0.02 to 0.03), and 208 of the 212 non-pediatric reports were at high risk of bias. Conclusions: ML has not been shown to predict DGF or graft survival better than regression, and the comparisons are too heterogeneous to show that it is worse; a model should inform organ acceptance or recipient counseling only after it has shown discrimination and calibration in an independent cohort.