Machine learning for predicting graft rejection in kidney transplant recipients: Predictive performance, methodological quality and barriers to clinical use

Allograft rejection remains a major threat to kidney graft survival, and machine learning (ML) has been proposed to improve its prediction. This study appraises the predictive performance and clinical readiness of ML models for rejection prediction in kidney transplant recipients through a structured narrative review of systematic reviews and cohort studies published in 2019–2026. Pooled evidence from 14 studies (109,317 recipients) reported mean areas under the curve (AUC) of 0.795 for decision trees and 0.782 for neural networks. In a 987-patient cohort, logistic regression (AUC 0.717) performed as well as or better than random forest, elastic net and XGBoost for early subclinical rejection. Among 74 related graft-outcome models, 49 had a high risk of bias and only 13 reported calibration. ML offers moderate discrimination but inconsistent added value; multicentre external validation and transparent reporting are required before clinical adoption.