Machine Learning-Based Risk Assessment Framework for Early Prediction and Nursing Intervention in Neonatal Hyperbilirubinemia

Abstract

In this study, XGBoost, an explainable machine learning framework for early newborn hyperbilirubinemia prediction and risk stratification, is compared to Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Deep Neural Networks. Prediction models used TSB, gestational age, postnatal age, birth weight, feeding style, and mother health. Performance was assessed using stratified 5-fold cross-validation, an independent 20% held-out test set, repeated random train-test splits (n = 5), ablation studies, statistical significance testing, hyperparameter tuning, SHAP and DiCE counterfactual analysis. In cross-validation, the XGBoost model surpassed LR (AUC 0.863), RF (0.880), SVM (0.890), and DNN (0.860) with an AUC of 0.97 (95% CI: 0.95–0.98), 94.2% accuracy, 96.1% sensitivity, and 92.8% specificity. Fold-wise validation revealed consistent performance with mean AUC 0.97 ± 0.015, accuracy 94.2 ± 0.5%, sensitivity 96.1 ± 0.6%, and specificity 92.6 ± 0.5%. XGBoost generalized well on the unknown test set (AUC 0.96, accuracy 93.5%, sensitivity 95.2%, specificity 91.8%). Robustness analysis revealed an average AUC of 0.968 ± 0.008 with ±0.2–0.6% variation across five random data splits. Ablation studies had significant improvements across all configurations ( p < 0.05), progressing from baseline LR (AUC 0.863) to default XGBoost (0.912), feature engineering (0.941), hyperparameter optimization (0.956), and proposed full model (0.970). TSB was the most significant predictor (SHAP contribution +0.42), followed by gestational age (+0.18) and postnatal age (+0.11). Hyperparameter change boosted XGBoost by 0.058 AUC. DiCE counterfactual research reduced risk score from 0.83 to 0.46, suggesting treatment. Risk-guided nurse interventions increased early intervention rates from 58% to 91% and hospital stays from 2.5 days in low-risk neonates to 6.8 days in high-risk infants, improving resource allocation. Explainable XGBoost produces accurate, robust, and clinically interpretable neonatal hyperbilirubinemia prediction.