Supervised machine learning models for predicting sepsis-associated acute kidney injury in children: a real-world evaluation.

backgroundSepsis-associated acute kidney injury (S-AKI) substantially increases mortality. The recent Phoenix criteria have redefined pediatric sepsis, yet AKI risk factors under this framework remain unclear. This study aimed to develop a machine learning (ML) model to identify key predictors of AKI in pediatric patients with sepsis as defined by the Phoenix criteria.

methodsThis retrospective cohort study included 2424 pediatric patients with a diagnosis of sepsis, of whom 484 developed AKI according to Kidney Disease: Improving Global Outcomes criteria. Data from the first 24 hours of admission were analyzed. The cohort was randomly divided into a training set (70%) and a test set (30%). A least absolute shrinkage and selection operator-random forest hybrid approach selected features, and ten ML algorithms were developed and evaluated on the training set, with their performance subsequently validated on the test set. Model performance was assessed using area under the curve (AUC), accuracy, Brier score, and decision curve analysis. Shapley additive explanations (SHAP) analysis interpreted the optimal model.

resultsFeature selection identified 13 key predictors, including SpO₂/FiO₂ ratio, international normalized ratio, procalcitonin, absolute neutrophil count, and mean arterial pressure. The categorical boosting (CatBoost) model achieved the best performance, with an AUC of 0.843 [95% confidence interval (CI) = 0.818-0.867] in the training set and 0.797 (95% CI = 0.755-0.839) in the test set. SHAP analysis highlighted procalcitonin and absolute neutrophil count as the most influential features. A web-based tool was created for clinical application.

conclusionsBased on the Phoenix criteria, we developed an interpretable CatBoost-based early prediction model and web tool for pediatric S-AKI using 13 clinical features. The model showed good predictive performance and provides a preliminary data-driven framework for S-AKI risk stratification. Future multi-center external validation is needed to confirm generalizability and explore feasibility of clinical integration.