Abstract
BACKGROUND: Sepsis-associated acute kidney injury (S-AKI) represents a critical complication with high mortality rates in intensive care units. Current risk stratification tools lack precision and interpretability for clinical decision-making. This study aimed to develop and validate interpretable machine learning models for predicting hospital mortality in S-AKI patients. METHODS: This retrospective cohort study utilized five international critical care databases: Medical Information Mart for Intensive Care (MIMIC)-IV (n = 12 966), MIMIC-III-CareVue (n = 2209), eICU (n = 8210), Northwestern University Intensive Care Unit (NWICU) (n = 2207) and Salzburg Intensive Care database (SICdb) (n = 1893). Adult patients with S-AKI meeting sepsis-3.0 and acute kidney injury criteria were included. Feature selection used the Boruta algorithm on MIMIC-IV, MIMIC-III and eICU databases. Eleven machine learning algorithms were trained using MIMIC-IV data with external validation on all other datasets. Performance was evaluated using receiver operating characteristic (ROC) curve analysis, calibration plots and decision curve analysis. SHapley Additive exPlanations (SHAP) analysis provided model interpretability. RESULTS: Among 27 485 S-AKI patients, hospital mortality was 27.5%. Boruta identified 21 consensus features including severity scores [Simplified Acute Physiology Score II (SAPS II), Sequential Organ Failure Assessment (SOFA), OASIS], vital signs and laboratory parameters. Gradient Boosting Machine emerged as optimal with area under the curve (AUC) values of 0.770 (training), 0.731 (internal validation) and 0.732-0.778 across four external validation cohorts. The model demonstrated excellent calibration and minimal overfitting (3.9% AUC difference). Decision curve analysis revealed superior clinical utility across probability thresholds of 4%-82%. SHAP analysis identified SAPS II as the most important predictor, with scores >60 and SOFA >15 associated with substantially increased mortality risk. Complete case analysis confirmed model robustness (AUC 0.766-0.847). CONCLUSIONS: The interpretable machine learning model demonstrated excellent performance and robust generalizability for S-AKI mortality prediction across five international databases. SHAP analysis provided clinically meaningful insights supporting personalized risk stratification and evidence-based clinical decision-making.