Acute kidney injury (AKI) is a common complication among critically ill patients and is associated with increased mortality, prolonged hospitalization, and long-term kidney dysfunction. Early identification remains challenging, particularly in resource-limited intensive care units (ICUs) where advanced biomarkers may not be available. We investigated whether AKI could be accurately predicted using only variables available at ICU admission and assessed model transparency using explainable AI.

We analyzed 849 adult ICU admissions from a South African tertiary hospital, including 497 patients who developed KDIGO-defined AKI. To simulate a clinically deployable model, only variables available at ICU admission were included. Variables reflecting AKI occurrence or collected after AKI onset, including KDIGO stage, AKI etiology, dialysis-related variables, creatinine trajectory, renal outcomes, and length of stay, were excluded to prevent data leakage. The final dataset comprised 27 admission-time variables, including demographics, comorbidities, illness severity scores (SOFA and SAPS 3), sepsis, vasopressor use, mechanical ventilation, and acute respiratory distress syndrome. Logistic regression, random forest, and gradient boosting models were developed using an 80:20 stratified train-test split with five-fold cross-validation. Model interpretability was evaluated using SHAP.

Logistic regression achieved the best performance, with a test ROC-AUC of 0.834 (five-fold cross-validation AUC 0.811 ± 0.031), accuracy of 75.3%, sensitivity of 78.0%, and specificity of 71.4%. Random forest (AUC 0.828) and gradient boosting (AUC 0.829) showed comparable discrimination, suggesting that predictive performance was primarily limited by the information available at admission rather than algorithm complexity. SHAP consistently identified SOFA and SAPS 3 scores as the strongest predictors, followed by vasopressor use, sepsis, and predicted mortality. These explanations aligned with established clinical knowledge of AKI risk factors.

Admission-time clinical information enables reliable and interpretable prediction of AKI without requiring longitudinal creatinine measurements or post-admission variables. The consistency between explainable AI and established clinical risk factors supports transparent bedside risk stratification and highlights the potential of explainable machine learning as a practical decision-support tool for early AKI identification in resource-limited ICUs.