AI Models for Predicting Acute Kidney Injury (AKI) and Post-AKI Mortality: Systematic Review and Meta-Analysis

Machine learning (ML) models are increasingly used to predict acute kidney injury (AKI), but validation quality and clinical readiness remain uncertain.

This systematic review and meta-analysis aimed to summarize discrimination performance and implementation-relevant gaps for AKI occurrence and post-AKI mortality prediction.

We searched the Cochrane Library, Embase, PubMed, and Web of Science through January 23, 2025. Eligible studies developed or validated ML-based prediction models and reported the area under the receiver operating characteristic curve (AUC). Two reviewers screened studies, extracted data, and assessed risk of bias using the PROBAST+AI (Prediction Model Risk Of Bias Assessment Tool+Artificial Intelligence). Logit-transformed AUCs were pooled using restricted maximum likelihood random-effects meta-analysis with Hartung-Knapp-Sidik-Jonkman-adjusted inference.

We included 219 studies with 7,343,170 participants and 101 modeling approaches. Primary analyses included 188 AUC estimates for AKI occurrence and 31 for post-AKI mortality. Pooled AUCs were 0.834 (95% CI 0.821-0.846) for AKI occurrence prediction and 0.830 (95% CI 0.807-0.851) for post-AKI mortality prediction. For AKI occurrence, nonlinear approaches, especially deep learning and tree-based or ensemble methods, generally showed higher pooled AUC point estimates than linear or generalized linear models in exploratory subgroup analyses. At the study level, PROBAST+AI rated 129 (58.9%) studies as having low risk, 84 (38.4%) studies as having high risk, and 6 (2.7%) studies as having unclear risk. External validation was uncommon: it was reported in 30 (16.0%) AKI occurrence records and 9 (29.0%) post-AKI mortality records.

ML models have shown high average discrimination for AKI occurrence and post-AKI mortality, supporting their potential value for AKI risk stratification and early warning. However, high levels of heterogeneity, limited external or prospective validation, and inconsistent reporting of model calibration and clinical utility mean that substantial barriers remain before routine clinical deployment. Pooled AUC estimates revealed that nonlinear models have considerable clinical translational potential. Further refinements to modeling frameworks are warranted to explore feasible strategies for real-world clinical implementation. Future studies should prioritize standardized definitions, robust validation, clinically meaningful thresholds, assessment of alert burden, and evidence that model-guided care improves kidney-protective management or patient outcomes.