Construction and validation of neonatal asphyxia risk prediction model based on machine learning

Background: Neonatal asphyxia is a leading cause of neonatal mortality and long-term neurodevelopmental impairment, particularly in low-resource settings, underscoring the need for improved predictive tools. Methods: This study developed a machine learning (ML) model using multi-source clinical data from 250 neonates (including mild/moderate asphyxia cases) at Zhejiang University Women’s Hospital. After meta-analysis identified 11 key risk factors, data preprocessing involved Random Forest imputation, standardization, and SMOTE for class balancing. Six ML algorithms (XGBoost, RF, Bagging, SVM, MLP, TabPFN) were evaluated, with SHAP analysis for interpretability. Results: XGBoost demonstrated superior performance (recall=0.82, precision=0.82, F1-score=0.82), with nuchal cord, assisted delivery, and prolonged labor emerging as top predictors. Ensemble methods (RF, Bagging) followed, while traditional models (SVM, MLP) and TabPFN showed lower efficacy. Conclusions: This study presents a validated ML framework for neonatal asphyxia prediction, offering clinical utility for early risk stratification and informed decision-making, particularly in resource-constrained environments.