Introduction: In resource-limited Pediatric Intensive Care Units (PICUs), where diagnostic and therapeutic capacities are constrained, accurately predicting mortality is essential for effective triage and resource allocation. While the Pediatric Risk of Mortality (PRISM) IV score is widely accepted, its performance in these settings is unclear, creating a need for more precise, cost-effective, and accessible tools. This study aimed to bridge this gap by testing the hypothesis that a Machine Learning (ML) model, developed using routinely collected local data, could predict pediatric mortality more accurately than the PRISM IV score. Methods: This retrospective cohort study included 300 patients (1 month to 15 years) admitted to the PICU between 2023-2024. We developed Random Forest classifier and an XGBoost (eXtreme Gradient Boosting) model using PRISM IV variables collected within 24 hours of admission. The dataset was divided into training (80%) and testing (20%) sets with model performance optimized via 5-fold cross-validation and hyperparameter tuning. The performance was measured using accuracy, area under the receiver operating characteristic curve (AUC-ROC), and precision F1 score. The SHapley Additive exPlanations (SHAP) method was used to evaluate the impact of the input variables. Results: The cohort’s median age was 3 years (IQR: 0.83 - 7.00), and the mortality rate of 6.34%. The XGBoost model was highly accurate (accuracy: 0.986, AUC-ROC: 0.985, precision: 0.800, and F1-score: 0.889), outperforming the PRISM IV score (accuracy: 0.663, AUC-ROC: 0.881, precision: 0.148, and F1-score: 0.255). SHAP analysis identified the use of ionotropes, need for mechanical ventilation, low Glasgow Coma Scale score, low heart rate, and a low systolic blood pressure as the most important predictors of mortality. Conclusions: ML models predicted mortality with higher accuracy than the PRISM IV score, demonstrating that powerful, high-precision predictive tools can be developed from local, routinely available data. By enabling more accurate risk stratification, these models can empower clinicians to make better-informed decisions, optimize limited resources, and improve patient outcomes. While these single-center results are highly encouraging, a multi-center trial is required before they can be broadly implemented.
Thapa et al. (Sun,) studied this question.