Performance of machine learning models for CKD progression and kidney failure: a systematic review
Louise Aubiniere-Robb, Joshua Tandy, Bhautesh Dinesh Jani, Jennifer S Lees, Jesse DawsonAbstract
Background
Machine learning (ML) is increasingly applied to predicting chronic kidney disease (CKD) progression and kidney failure using routinely collected healthcare data, but the methodological quality and clinical applicability of these models remain uncertain. We hypothesised that ML models would show strong discrimination but be limited by inadequate external validation, inconsistent calibration, and high risk of bias (ROB), restricting clinical implementation.
Methods
We conducted a systematic review according to PRISMA guidance. Ovid MEDLINE, Embase, and CINAHL were searched from inception for studies evaluating ML models predicting CKD progression, kidney failure, or kidney replacement therapy using routinely collected healthcare data. Studies were synthesised by modelling framework, and ROB was assessed using PROBAST.
Results
Fifty-two studies were included. ML models demonstrated favourable discrimination, with 37/52 studies reporting AUROC or concordance estimates ≥ 0.80. Where directly comparable, gains over conventional statistical models and established risk scores, including the Kidney Failure Risk Equation (KFRE), were often modest. External validation and calibration assessment were each reported in 19/52 studies, while 41/52 studies were at high ROB. Kidney function, proteinuria-related variables, and age were the most influential predictors. Substantial heterogeneity was observed in cohort design, outcomes, prediction horizons, and validation methodology. Inconsistent handling of proteinuria, missing data, and model-development workflows further limited assessment of transportability and real-world applicability.
Conclusions
ML models for CKD prognostic prediction demonstrate favourable discrimination across kidney outcomes. However, limited external validation, inconsistent calibration, and methodological heterogeneity restrict clinical implementation. Future progress requires rigorous validation, transparent methodology, and demonstration of real-world clinical utility.