TRANSPARENT AI-BASED EARLY WARNING FOR ACUTE KIDNEY INJURY: MULTI-CENTRE VALIDATION WITH A 24-HOUR ACTIONABLE HORIZON

Acute kidney injury (AKI) is common, costly, and often recognized only after injury begins; existing tools are reactive or weakly validated. Delayed recognition worsens outcomes, prolongs hospital stay, and increases costs—underscoring an urgent need for earlier detection. We therefore built a bedside-ready early warning score (EWS) that anticipates incident AKI within 24-h.

Retrospective intensive care unit (ICU) cohorts from the Medical Information Mart for Intensive Care (MIMIC-IV) and the eICU Collaborative Research Database (eICU-CRD) were expanded into rolling 12-h windows of routinely collected features, a 2-h gap to prevent leakage during training, and a 24-h prediction horizon, updated every 6 hours. Adult patients without baseline AKI or renal replacement therapy were included; stays were censored at first AKI. Extreme gradient-boosted trees (XGBoost) with prespecified monotonic constraints were trained under 10-fold, stay-level cross-validation; probabilities were calibrated with Platt scaling. Two prespecified validation sets tested transportability: temporal (MIMIC-IV 2020–2022, COVID-19 era) and geographic external (eICU Northeast region hospitals). Discrimination, calibration, decision-curve analysis (DCA), feature importance, and subgroup bias checks were performed. The full pipeline was fixed and packaged for further use in accordance with TRIPOD guidance.

After exclusions, development included 44,066 ICU stays; temporal and external validations included 3,346 and 2,993 stays, respectively. Event rates were 11.7% (development), 17.1% (temporal), and 14.2% (external). AUROC was 0.88 (95% CI 0.84–0.90) in development, 0.84 (0.80–0.87) in temporal validation, and 0.82 (0.80–0.85) externally; AUPRC was 0.60 (0.57–0.62), 0.60 (0.56–0.64), and 0.53 (0.50–0.57), respectively. Using a prespecified threshold, the temporal cohort achieved sensitivity 0.76 (0.73–0.80), specificity 0.79 (0.76–0.83), precision 0.47 (0.43–0.54); external validation achieved sensitivity 0.73 (0.67–0.78), specificity 0.86 (0.83–0.89), precision 0.48 (0.42–0.54). This means roughly 3 in 4 impending AKI cases can be anticipated early enough for intervention. Calibration was strong (Brier 0.07, 0.10, 0.09); post-calibration slopes ≈1.0 with minimal intercept shift (temporal slope 1.03, calibration-in-the-large [CITL] 0.008; external slope 1.00, CITL 0.057). DCA showed consistent net benefit versus treat-all and treat-none strategies, with durable advantage up to high thresholds. Subgroups showed stable discrimination without material differences by sex or race. SHAP explanations highlighted actionable drivers—declining mean arterial pressure, low urine output, rising creatinine and blood urea nitrogen, deteriorating SpO₂ and heart rate trajectories. No high-frequency waveforms or bespoke data feeds were required.

A transparent, reproducible EWS anticipates ICU AKI within 24 h, generalizes across time (COVID-19 era) and geography, and remains well-calibrated with proven clinical utility. At a prespecified threshold, it identifies approximately 3 in 4 impending AKI cases while maintaining high specificity, enabling early actions within routine workflows (optimize perfusion/volume, reassess nephrotoxins, reconsider contrast). Transportable design and interpretable outputs support immediate, pragmatic deployment and local recalibration.