Forecasting ICU Acute Kidney Injury with Actionable Lead Time Using Interpretable Machine Learning: Development and Multi-Center Validation

Background: Acute kidney injury (AKI) is common in intensive care and is frequently recognized late. We aimed to develop an interpretable, dynamic early warning score (EWS) to predict incident AKI within the next 24 h in ICU adult patients and to test its transportability temporally and geographically. Methods: We performed a retrospective cohort multi-center study using hospitalized ICU patients from MIMIC-IV and eICU-CRD databases. The outcome was incident AKI (KDIGO stage 1–3). Gradient-boosted trees (XGBoost) were trained with 10-fold cross-validation. Predictors were prespecified from the literature and finalized as 61 routinely available EHR predictors selected using SHAP ranking (spanning demographics/comorbidities, laboratory markers, vital-sign dynamics, and ICU therapies/procedures). Prespecified validations included (i) temporal validation in the COVID-19 era and (ii) geographic validation in eICU Northeast hospitals. Results: The development cohort included 51,833 ICU stays; temporal and geographic cohorts included 3346 and 2993 stays, respectively. Discrimination was high in internal validation (AUC 0.88, 95% CI 0.84–0.90) and remained robust temporally (AUC 0.84, 0.80–0.87) and geographically (AUC 0.82, 0.81–0.84). At a prespecified threshold, sensitivity/specificity were 0.76/0.79 (temporal) and 0.73/0.86 (geographic). Decision-curve analysis showed net benefit across clinically relevant thresholds. Key predictors reflected physiologic trajectories (e.g., mean arterial pressure dynamics), urine output, renal/metabolic markers (e.g., creatinine and BUN trends), and oxygenation dynamics (SpO2). Conclusions: A routinely updated, explainable EHR-based EWS can predict ICU AKI up to 24 h in advance with stable performance across a pandemic-era temporal shift and external geographic validation.