Development and validation of an explainable XGBoost model for early mortality prediction in pediatric sepsis.

Objective: To develop and internally validate an explainable machine learning model for early mortality prediction in pediatric sepsis using routinely available clinical variables, to support clinical risk stratification and decision-making. Methods: A total of 752 children with sepsis admitted to the Pediatric Intensive Care Unit (PICU) at the Children's Hospital of Soochow University between January 1, 2019, and June 30, 2023, were retrospectively enrolled. Twenty-six routinely available clinical variables obtained within 24 h after PICU admission were included. After data preprocessing, variables with excessive missing values, significant multicollinearity, or limited univariate predictive value were excluded. Clinically meaningful ratio-derived variables were constructed, and logarithmic transformation was applied to selected right-skewed continuous variables. Twelve supervised machine learning algorithms were developed and compared for mortality prediction. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), precision-recall (PR) curve, decision curve analysis (DCA), and other performance metrics. Model interpretability was assessed using SHapley Additive exPlanations (SHAP). Results: Among the 12 machine learning algorithms, extreme gradient boosting (XGBoost) achieved the best overall predictive performance. Following recursive feature elimination based on permutation importance, a simplified XGBoost model incorporating seven variables was established and interpreted using SHAP. In the internal validation cohort, the final model achieved an AUC of 0.803, with a sensitivity of 0.846, specificity of 0.696, accuracy of 0.722, and an F1 score of 0.512. The mean AUCs obtained from 5-fold and 10-fold cross-validation were 0.780 ± 0.044 and 0.771 ± 0.070, respectively, indicating good model stability. SHAP analysis identified pH, PaO₂, glucose, blood urea nitrogen (BUN), lactate dehydrogenase (LDH), international normalized ratio (INR), and the BUN-to-creatinine ratio (BUN/Cr) as the features that made the greatest contributions to the final model's mortality predictions. Conclusion: An explainable XGBoost-based prediction model was developed for early mortality prediction in pediatric sepsis using seven routinely available clinical variables collected within 24 h after PICU admission. The model demonstrated good discrimination, stability, and model-based interpretability, suggesting its potential as a practical tool for early risk stratification and individualized clinical management. Further multicenter external validation is warranted before routine clinical implementation.