Early prediction of acute kidney injury (AKI) in the intensive care unit may support prevention, but AI models often lose calibration when transported across institutions. We developed and externally validated an explainable AI model for early ICU AKI prediction and evaluated deployment-readiness components.

We performed a retrospective prediction study using MIMIC-IV v3.1 for development/internal testing and eICU v2.0 for external validation. The prediction time was ICU admission plus 6 hours. The primary outcome was KDIGO stage 2-3 AKI or renal replacement therapy through 48 hours. Predictors were restricted to pre-index information and included demographics, ICU type, creatinine trajectory, vital signs, labs, vasopressor exposure, and mechanical ventilation. An XGBoost model was calibrated using out-of-fold development predictions and frozen before holdout and external validation. Deployment-readiness analyses included eICU recalibration, risk tiers, patient-level SHAP explanations, and silent-mode monitoring triggers.

The primary analysis included 61,656 MIMIC-IV stays with 2,961 events and 97,951 eICU stays with 3,238 events. In the MIMIC-IV holdout set, the frozen model achieved AUROC 0.861, AUPRC 0.311, Brier score 0.039, and observed-to-expected (O/E) ratio 1.030. In eICU external validation, AUROC was 0.847, AUPRC 0.229, Brier score 0.029, and O/E ratio 0.860, indicating preserved discrimination with calibration shift. At the 10% development alert threshold, eICU sensitivity was 48.8%, positive predictive value 20.4%, and false alerts 6.28 per 100 stays. In a split-sample eICU deployment-readiness experiment, Platt recalibration preserved AUROC 0.840 and AUPRC 0.220 while improving O/E from 0.859 to 0.997. The eICU very-high risk tier had observed AKI/RRT risk 27.4%. Patient-level SHAP explanations identified plausible drivers, including creatinine trajectory, mechanical ventilation, vasopressor exposure, lactate, respiratory rate, and acid-base variables.

An explainable AI model trained in MIMIC-IV retained strong discrimination during multicenter eICU external validation, but calibration shifted across settings. A deployment layer using local recalibration, risk-tier workflow, patient-level explanations, and silent-mode monitoring is a practical next step before live clinical alerting.