Pediatric sepsis remains a leading cause of global mortality, necessitating early and accurate risk stratification. Traditional scoring systems often fail to capture the rapid, non-linear progression of the disease involving both immune dysregulation and metabolic derangement. This study aimed to develop and validate an interpretable machine learning (ML) prognostic model by integrating novel inflammatory indices, specifically the Systemic Immune-Inflammation Index (SII) and Lactate to predict 28-day mortality in children with sepsis. A single-center, retrospective cohort study was conducted involving 500 pediatric patients (300 sepsis and 200 controls), followed by an independent validation cohort of 200 patients. Clinical characteristics, inflammatory biomarkers (NLR, PLR, SII), and metabolic indicators (Lactate) were analyzed. LASSO regression was employed for robust feature selection, and ten ML algorithms were compared to construct the predictive model. SHAP (SHapley Additive exPlanations) analysis was utilized to visualize feature contributions and ensure model interpretability. Sepsis patients, particularly non-survivors, exhibited significantly elevated SII, NLR, D-dimer, and Lactate levels compared to survivors (p 0.87. SHAP interpretation revealed that elevated Lactate and younger age were strong drivers of mortality risk, while elevated SII and PCT also contributed positively to adverse outcome predictions. These findings were confirmed in the validation cohort, where SII and Lactate retained significant discriminative power. This study establishes a robust, interpretable ML model for pediatric sepsis prognosis. The integration of composite immune indices (SII) and metabolic markers (Lactate) significantly enhances early risk stratification beyond traditional metrics. This approach provides a clinically practical tool to identify high-risk patients at admission, potentially guiding timely, personalized interventions during the initial phase of care.
Liang et al. (Mon,) studied this question.