An interpretable machine learning model using routine clinical indicators for early pediatric sepsis risk stratification: a multicenter retrospective cohort study

Background Sepsis remains a leading cause of pediatric morbidity and mortality worldwide. The 2024 Phoenix Sepsis Criteria emphasize early recognition of organ dysfunction. Conventional diagnostics like blood cultures are time-consuming, and existing machine learning models lack early pathogen-level classification. A machine learning framework for early pediatric sepsis warning and Gram-stain pathogen classification was developed and validated using routine clinical indicators. Methods This multicenter retrospective study included 784 children: 432 from Fujian Children’s Hospital for model development and 352 from the Pediatric Intensive Care database of the Children’s Hospital, Zhejiang University School of Medicine for external-cohort evaluation. From 56 clinical variables and 9 derived indices, features were selected in the training cohort by univariable analysis, least absolute shrinkage and selection operator (LASSO) regression, and area under the receiver operating characteristic curve (AUC) evaluation. Eight algorithms were compared, with external-cohort accuracy also informing hyperparameter selection. Shapley additive explanations (SHAP) were used to interpret predictions. The Gram-stain nomogram was developed and evaluated in the pathogen-confirmed training sub-cohort. Results The reported models used 12 features. Random forest achieved an AUC of 0.887 (95% confidence interval: 0.850–0.924) and accuracy of 82.1% in the external cohort. Decision curve analysis suggested potential clinical net benefit. C-reactive protein, procalcitonin, alanine aminotransferase, and white blood cell count contributed most to model output. SHAP plots described nonlinear associations within the fitted model. The exploratory Gram-stain nomogram achieved an apparent AUC of 0.753 (95% confidence interval: 0.661–0.844) in the development sub-cohort, with sensitivity of 57.1% and specificity of 80.3% for Gram-positive infection. Conclusion The random forest model showed potential for early pediatric sepsis risk stratification and warrants evaluation in an independent cohort. The Gram-stain nomogram provides exploratory findings that require external validation before clinical use. Antibiotic selection and de-escalation should be guided by microbiological evidence and comprehensive clinical assessment.