Explainable machine learning using urinary metabolomics to predict pediatric sepsis-associated acute kidney injury: a two-center prospective observational study

Sepsis-induced acute kidney injury (S-AKI) is a common and serious complication in critically ill children with a poor prognosis, and its early and accurate prediction remains challenging due to the lack of reliable biomarkers. Urinary small-molecule metabolomics offers a promising approach to capture the dynamic metabolic changes during the progression of S-AKI. In this two-center prospective observational study, we enrolled 360 children from both centers. Urine samples were collected within 24h after hospitalized children diagnosed with sepsis, stored at -80 °C, and analyzed using gas chromatography-mass spectrometry (GC-MS). Based on urinary metabolic fingerprints (U-MF), we developed and validated a machine learning model for early prediction of S-AKI. The Shapley Additive Explanations (SHAP) algorithm was applied to visually explain the optimal model. A panel of 10 metabolites was selected as common discriminative features. Among the 4 machine learning models evaluated, the support vector machine (SVM) demonstrated the best performance in both the discovery cohort (AUC 0.94, 95% CI: 0.91-0.98) and the external validation cohort (AUC 0.89, 95% CI: 0.82-0.96), enabling early prediction of S-AKI within 24 h. Furthermore, the U-MF panel was integrated into an open-access online platform to facilitate clinical translation. Our findings suggest that U-MF combined with machine learning holds promise as a robust and noninvasive approach with potential utility for early prediction of S-AKI in pediatric patients.