Table 1_Machine learning-based time-to-event survival analysis in pediatric patients with severe sepsis.xlsx

BackgroundPediatric sepsis remains a leading cause of mortality in critically ill children worldwide. Current approaches to sepsis prognosis rely on clinical criteria and biomarkers with variable performance. This study aimed to develop and validate time-to-event survival prediction models for pediatric sepsis using survival analysis machine learning algorithms.
MethodsWe conducted a retrospective cohort study of 223 pediatric sepsis patients from a pediatric intensive care database (2010–2018). Five survival analysis machine learning algorithms were evaluated: CoxPHSurvivalAnalysis, HingeLossSurvivalSVM, GradientBoostingSurvivalAnalysis, RandomSurvivalForest, and ExtraSurvivalTrees. These algorithms predict survival time rather than binary outcomes. Model performance was assessed using time-dependent area under the curve (td-AUC), concordance index (c-index), Brier score, and calibration curves. SHapley Additive exPlanations (SHAP) analysis was performed for model interpretability, and zero-crossing point analysis identified clinically actionable thresholds.
ResultsAmong 223 patients, 200 (89.7%) survived with median ICU stay of 12.2 days for survivors vs. 2.3 days for non-survivors. RandomSurvivalForest achieved the highest performance with td-AUC of 0.97, while CoxPHSurvival and HingeLossSurvivalSVM showed comparable c-indices of 0.87. SHAP analysis identified calcium total and RDW as the strongest mortality predictors. Zero-crossing point analysis established clinical thresholds: calcium total <1.10 mmol/L, RDW >15.07%, sodium <131.68 mmol/L, and pH <7.32 were associated with increased mortality risk, with U-shaped relationships observed for creatinine and lymphocytes.
ConclusionsRandomSurvivalForest demonstrated superior time-to-event prediction performance for pediatric sepsis. The survival analysis approach provides dynamic risk assessment and precise timing for clinical interventions. A web-based prediction calculator was developed to facilitate clinical implementation.