Machine Learning-based Prediction Models in Pediatric Polytrauma: A Retrospective Observational Pilot Study from a Resource-limited Setting

Abstract

A BSTRACT Pediatric polytrauma is a major cause of morbidity and mortality worldwide. Early risk stratification remains challenging, particularly in resource-limited settings where access to advanced imaging, intensive care facilities, and specialist workforce may be limited. The aim of this study was to evaluate the feasibility of developing artificial intelligence-based predictive models for early identification of high-risk pediatric polytrauma patients. A retrospective observational study screened 60 pediatric polytrauma patients aged 1–15 years. Twenty-three records were excluded because of incomplete clinical data required for model development, resulting in a final analytical cohort of 37 patients. Clinical variables included demographic characteristics, physiological parameters, metabolic indicators, and Injury Severity Score (ISS). Logistic Regression, Random Forest, and XGBoost models were developed to predict intensive care unit admission, inhospital mortality, morbidity, and composite high-risk status. Model performance was assessed as an exploratory feasibility analysis using receiver operating characteristic–area under the curve and classification metrics. The best-performing model was integrated into a Streamlit-based prototype interface. Machine learning models demonstrated promising discriminatory performance for predicting adverse outcomes. Glasgow Coma Scale, ISS, and serum lactate emerged as the most influential predictors. The prototype decision-support interface demonstrated the technical feasibility of translating model outputs into a real-time clinical tool. Machine learning models demonstrated preliminary feasibility for supporting early risk stratification in pediatric polytrauma. However, the small sample size, potential risk of overfitting, and lack of external validation necessitate cautious interpretation. Larger multicenter studies with rigorous validation are required before clinical implementation.