Machine Learning Model Predicts Pediatric Sepsis Risk Using Routine Clinical Indicators

A machine learning random forest model utilizing routine clinical indicators achieved an area under the receiver operating characteristic curve of 0.887 and an accuracy of 82.1 percent in an external cohort for early pediatric sepsis risk stratification, according to a multicenter retrospective study. The authors developed and evaluated the computational framework to address limitations in conventional diagnostics such as time-consuming blood cultures and the absence of early pathogen-level classification in existing models, in the context of the 2024 Phoenix Sepsis Criteria emphasizing early recognition of organ dysfunction.

The retrospective investigation analyzed records from 784 pediatric patients across two clinical centers. The model development cohort consisted of 432 children from Fujian Children’s Hospital, while an external validation cohort included 352 children sourced from the Pediatric Intensive Care database of the Children’s Hospital, Zhejiang University School of Medicine.

To construct the risk stratification models, researchers initially examined 56 clinical variables and 9 derived indices within the training cohort. Feature selection was conducted using univariable analysis, least absolute shrinkage and selection operator regression, and area under the receiver operating characteristic curve evaluation, resulting in 12 features utilized by the reported models. The investigators compared eight machine learning algorithms, utilizing external-cohort accuracy to inform hyperparameter selection.

In the external cohort evaluation, the random forest algorithm achieved an area under the receiver operating characteristic curve of 0.887, with a 95 percent confidence interval of 0.850 to 0.924, and an accuracy of 82.1 percent. Decision curve analysis indicated potential clinical net benefit. To interpret the predictions, Shapley additive explanations were applied to describe nonlinear associations within the fitted model, identifying C-reactive protein, procalcitonin, alanine aminotransferase, and white blood cell count as the primary contributors to model output.

The investigators also developed and evaluated an exploratory Gram-stain nomogram within the pathogen-confirmed training sub-cohort. In this development sub-cohort, the nomogram demonstrated an apparent area under the receiver operating characteristic curve of 0.753, with a 95 percent confidence interval of 0.661 to 0.844. For identifying Gram-positive infection, the nomogram exhibited a sensitivity of 57.1 percent and a specificity of 80.3 percent.

The authors concluded that the random forest model demonstrated potential for early pediatric sepsis risk stratification and warrants further evaluation in an independent cohort. The team noted that the Gram-stain nomogram provides exploratory findings that require external validation prior to clinical application, emphasizing that antibiotic selection and de-escalation should continue to be guided by microbiological evidence and comprehensive clinical assessment.

Subject of Research: Machine learning for pediatric sepsis risk stratification

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

Article References: Li, Y., Li, X., Wang, D., Li, M., Pan, W., & Xue, H. (2026). An interpretable machine learning model using routine clinical indicators for early pediatric sepsis risk stratification: a multicenter retrospective cohort study. Frontiers in Cellular and Infection Microbiology, 16, Article 1933201. https://doi.org/10.3389/fcimb.2026.1933201

Image Credits: AI Generated

DOI: 10.3389/fcimb.2026.1933201

Keywords: pediatric sepsis, machine learning, random forest, clinical indicators, risk stratification, Gram-stain nomogram, Phoenix Sepsis Criteria, intensive care, interpretable, machine, learning, model