Early Warning Systems for Acute Kidney Injury Prediction Before Clinical Onset: A Systematic Review of 24-48 Hour Risk Stratification Models

Acute kidney injury (AKI) is usually recognised only after serum creatinine has risen, when the chance to prevent it has largely passed.Electronic early warning models that predict AKI 24-48 hours before it meets diagnostic criteria could leave time to act.Following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020, we searched PubMed, Web of Science, Scopus, Embase, and Cumulative Index to Nursing and Allied Health Literature (CINAHL) for peer-reviewed primary studies published between January 2021 and September 2026.Eligible studies developed or validated models predicting incident AKI at a fixed 24-or 48-hour horizon in unselected hospitalised or intensive care patients and reported discrimination at that horizon.Data were extracted using the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) checklist, and risk of bias was assessed with the Prediction model Risk Of Bias ASsessment Tool (PROBAST).Of 947 records, 13 studies were included.They came from the United States, China, Taiwan, Singapore, Sweden, and multinational consortia and together analysed about 623,000 admissions or encounters.Six studied intensive care patients, six studied general or hospital-wide inpatients (one of them in children), and one did not report its setting.Gradient-boosted trees and deep sequence models were the most common algorithms.Internal area under the receiver operating characteristic curve (AUROC) ranged from 0.72 to 0.99.External AUROCs ranged from 0.58 to 0.87 and were usually lower.Only two studies formally assessed calibration, only one included prospective validation, and none evaluated the effect of alerts on patient outcomes.Overall risk of bias was low in five studies, unclear in six, and high in two.Models that predict AKI 24-48 hours ahead achieve moderate-to-good discrimination, but performance falls on external validation.Reporting of calibration and evidence of clinical impact remain scarce.Prospective, impact-focused evaluation should come before routine deployment.