An explainable machine learning model for early pediatric sepsis prediction using cytokine and routine laboratory data.

Abstract

Early sepsis diagnosis in children remains challenging due to nonspecific presentations. This study aimed to develop an interpretable machine learning (ML) model to improve early prediction. We conducted a retrospective cohort study of pediatric patients with infections. Using clinical and cytokine data, key predictors were selected to construct and compare several machine learning models. The logistic regression model demonstrated the best overall performance for early sepsis prediction. Key predictive factors included specific interleukins (IL-10, IL-33) and routine infection markers. Model interpretability was achieved using Shapley Additive Explanations (SHAP) analysis. We established an interpretable, high-performance model for early pediatric sepsis prediction. Its implementation as an online calculator facilitates real-time risk assessment, bridging the gap between predictive analytics and bedside clinical utility. Key Message: this study developed and validated a machine learning model that integrates pediatric cytokine profiles with routine infection markers, achieving an AUC of 0.908 for early sepsis diagnosis. This study uniquely integrates the patient's immune signature with machine learning and SHAP interpretation, translating it into a clinically accessible web-based calculator. this model holds immediate potential for deployment in emergency or outpatient settings to aid in early sepsis identification and treatment decision-making, thereby potentially improving patient outcomes.