Development and validation of a machine learning prediction model for pediatric sepsis-associated acute kidney injury: a multicenter cohort study

Sepsis-associated acute kidney injury (SA-AKI) significantly increases mortality in pediatric sepsis patients. However, early warning tools for pediatric cases remain scarce. This study aimed to develop and validate an interpretable machine learning model for the early prediction of SA-AKI in pediatric sepsis. This multicenter retrospective cohort study analyzed data from pediatric sepsis patients admitted to pediatric intensive care units (PICUs) at five Chinese centers. Patients were stratified into training, internal temporal validation, and four external validation cohorts. A rigorous feature selection process identified the most predictive variables from 41 routinely available clinical parameters within 24 hours of admission. Ten machine learning algorithms were trained and optimized, with the model achieving the highest area under the curve (AUC) selected as the final model. Model interpretability was achieved using SHapley Additive exPlanations (SHAP). A total of 3,437 pediatric sepsis patients were included, with an overall SA-AKI incidence of 25.7%. The XGBoost model was selected as the optimal algorithm, demonstrating robust discrimination across all cohorts, with AUCs of 0.899 (95% CI: 0.88–0.91) in the training set, 0.854 (95% CI: 0.81–0.89) in the internal temporal validation set, and ranging from 0.794 to 0.844 across four external validation sets. The model significantly outperformed traditional clinical scores (pSOFA and Phoenix) in most cohorts (all p < 0.05), although its discrimination was comparable to that of pSOFA in the PIC validation cohort; calibration was acceptable in most cohorts but poorer in the PIC cohort, where the model under-predicted AKI risk. Decision curve analysis confirmed net clinical benefit within the clinically plausible threshold range of 5%–50%, and over the treat-all strategy at thresholds above approximately 10%. SHAP analysis revealed standard base excess (SBE), aspartate aminotransferase (AST), and potassium as the top predictors of SA-AKI, with dependence plots elucidating their nonlinear relationships with SA-AKI risk. This study developed and validated an explainable XGBoost model for early pediatric SA-AKI prediction. Utilizing eight routinely available parameters, the model demonstrated robust multicenter discrimination, superior or comparable performance to traditional scores, and potential clinical utility. By enabling timely, interpretable risk stratification for high-risk children, this tool addresses a critical gap in pediatric SA-AKI prediction and holds promise for guiding early kidney-protective strategies. Given the retrospective design and the all-China validation context, prospective implementation and impact studies are warranted before clinical deployment.