Building a Machine Learning Model to Predict the Early Mortality Risk in Pediatric ICU Sepsis Patients

Objective: This study aims to construct a risk prediction model for early mortality in pediatric intensive care unit (PICU) sepsis patients using machine learning methods. The model is intended to assist clinicians in identifying high-risk patients to enable timely interventions, and to help prevent early deaths in PICU sepsis patients. Methods : A single-center retrospective cohort study design was used; clinical data of sepsis patients admitted to the PICU of Chongqing Medical University Affiliated Children's Hospital from January 2015 to December 2021 were included. The data comprised demographic information, vital signs, complications, laboratory indicators, diagnoses, and treatments. Patients were divided into early mortality and survival groups based on whether death occurred within 14 days of PICU admission. Seventy percent of the data were randomly assigned to the training set, and 30% were assigned to the validation set. Seven machine learning models, including logistic regression (LR), decision tree (DT), random forest (RF), XGBoost (XGB), LightGBM (LGBM), Naive Bayes (NB), and support vector machine (SVM), were used to build the early mortality prediction model for pediatric sepsis patients. The model's predictive performance was evaluated using sensitivity, specificity, receiver operating characteristic curve (ROC curve), and calibration curve. The clinical application value of the model was assessed through decision curve analysis (DCA). Results: A total of 1,559 pediatric sepsis patients were included, with 198 cases of early mortality, resulting in an early mortality rate of 12.7%. After feature selection using LASSO regression, recursive feature elimination (RFE), and Boruta algorithms, six optimal predictive variables were identified: blood transfusion, pediatric Sequential Organ Failure Assessment (pSOFA) score, cardiopulmonary resuscitation, days of mechanical ventilation, secondary infection(s), and septic shock. These variables were used to construct the early mortality prediction model. ROC curve analysis showed that the area under the curve (AUC) values of the seven models ranged from 0.88 to 0.94. The XGBoost and LGBM models performed the best, both achieving an AUC of 0.94. Among all models, XGBoost and LGBM had the highest accuracy, 0.929 and 0.925 respectively; sensitivity was 0.534 and 0.483; specificity was 0.985 and 0.988; and F1 scores were 0.653 and 0.615, respectively. Calibration curves indicated that the XGBoost and LGBM models performed best among the seven machine learning models, showing excellent calibration performance. Decision curve analysis demonstrated that, compared to other models, XGBoost and LGBM exhibited greater net benefits across a wider range of threshold probabilities, indicating good clinical application value and confirming them as the optimal models. Conclusion: Machine learning models are reliable tools for predicting early mortality risk in pediatric patients with sepsis in the PICU. The XGBoost and LGBM models were built using six key variables: blood transfusion, Pediatric Sequential Organ Failure Assessment (pSOFA) score, post-cardiopulmonary resuscitation (CPR), days of mechanical ventilation, secondary infection, and septic shock. These models demonstrate stable performance, high discriminatory ability, and accuracy. Moreover, they aid clinicians in identifying high-risk patients and implementing early interventions. Trial Registration: The trial was registered on the Chinese Clinical Trial Registry with the number ChiCTR2300068260 on February 13, 2023.