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

Comprehensive Summary

The study develops and validates an explainable machine learning model using urinary metabolomic profiling to predict sepsis-associated acute kidney injury (S-AKI) in critically ill pediatric patients. Researchers conducted a two-center prospective observational study involving 360 children with sepsis and collected urine samples within 24 hours of diagnosis for gas chromatography–mass spectrometry (GC-MS) analysis. Using statistical feature-selection methods, the researchers identified 10 key urinary metabolites as predictive biomarkers and used these variables to construct four machine learning models: support vector machine (SVM), logistic regression, gradient boosting machine, and naïve Bayes. Among the leading factors identified to be associated with S-AKI were metabolic disturbances in pyruvate metabolism. Among the four models, the SVM model demonstrated the strongest predictive performance, achieving excellent discrimination in both the discovery cohort (dataset used to build the machine learning model, n = 255) and external validation cohort (dataset from pediatric patients of another hospital to test generalizability, n = 105), along with high accuracy, sensitivity, specificity, and clinical utility. To improve interpretability, the researchers applied Shapley Additive Explanations (SHAP), which found that urinary lactic acid, homovanillic acid, and other metabolite signatures contributed significantly to prediction outcomes. These metabolites were individually weighted within the SVM model to generate patient-specific probabilities of developing S-AKI. Overall, this study built a robust predictive model for predicting pediatric sepsis-associated acute kidney injury. The model is released as an open-access tool that can be used by clinicians to obtain individualized risk estimates for their own patients as they see fit.

Outcomes and Implications

The authors developed a robust AI tool that, when combined with urinary metabolomic data, can detect acute kidney injury secondary to sepsis. The integration of metabolic biomarkers with machine learning algorithms, as seen by the strong performance of the support vector machine model, demonstrates how data-driven tools can improve how we predict and manage septic patients. AI interventions may enable earlier clinical intervention and potentially mitigate disease progression and improve patient outcomes. Importantly, the use of explainable machine learning methods, such as SHAP, increases transparency by identifying key metabolites contributing to risk, which uncovers the potential mechanisms behind the pathogenesis of AKI. For clinicians, the value of incorporating metabolomic data and AI-assisted tools into decision-making processes is clear. As for healthcare systems, this study further supports the benefits of AI-powered infrastructure to allow for enhanced detection of poor health outcomes before they even precipitate. Nonetheless, the study also has limitations, including potential selection bias, limited ability to predict AKI severity or timing, and challenges in standardizing metabolite measurements for routine clinical use. Despite these limitations, the findings demonstrate that combining noninvasive biomarkers with explainable machine learning holds promise for advancing early diagnosis and personalized management of S-AKI in critical care.