Machine Learning for Early Detection and Progression Prediction of Global Chronic Kidney Disease: A Systematic Review of Externally Validated Models

Background. Machine learning (ML) models for chronic kidney disease (CKD) are published in large numbers, but most are tested only on the data used to build them. Performance in new populations is what matters for clinical use. Methods. Following PubMed/MEDLINE, IEEE Xplore and Google Scholar were searched. Eligible studies, published between January 2021 and June 2026, developed or evaluated ML models predicting incident CKD or CKD progression in adults from structured clinical data. They also had to report an area under the curve (AUC) or C-statistics for an ML model in external data, either geographic or temporal. Results. Of 1,087 records, 155 full texts were assessed and 16 studies were included. Three predicted incident CKD and 13 predicted progressions. Thirteen final models were tree-based. Headline external AUCs or C-statistics ranged from 0.69 to 0.97 and were usually lower than internal estimates. Calibration in external data was reported in some form by eight studies, but mostly as a Brier score alone. Three studies compared ML with the Kidney Failure Risk Equation (KFRE) in the same data, with mixed results. Twelve studies were at high risk of bias and none at low risk. No study used Central Asian data or linked predictions to clinical action. Conclusions. Externally validated ML models for CKD exist and can perform well, but small external samples, sparse calibration and few fair comparisons with the KFRE limit confidence. Local validation is needed before use in settings such as Uzbekistan.