Optimized ensemble learning framework for neonatal asphyxia prediction using perinatal clinical features.

Background: Neonatal asphyxia is a life-threatening perinatal condition associated with neonatal mortality and long-term neurological impairment. Therefore, early identification of high-risk newborns is essential for timely clinical intervention; however, accurate prediction of neonatal asphyxia remains a key challenge due to heterogeneous maternal and perinatal risk factors, for example, overlapping clinical patterns, and class imbalance in medical datasets. Thus, to solve this issue, this research proposes an imbalance-sensitive machine-learning framework for the prediction of neonatal asphyxia using perinatal and maternal clinical variables. In contrast to conventional SMOTE, which interpolates minority samples without explicitly prioritizing their diagnostic difficulty, the proposed framework is designed to strengthen the representation of informative asphyxia cases located in uncertain or overlapping decision regions while reducing the influence of less representative synthetic generation. Methods: In summary, in this research, a novel synthetic oversampling framework, named HEM-SMOTE, is proposed to improve minority-class representation by generating more informative synthetic asphyxia samples through hybrid distance-guided neighbor selection. The proposed method integrates local Euclidean-distance-based similarity with Mahalanobis-distance-based covariance awareness to identify representative minority neighbors prior to the generation of synthetic samples. Results: To measure the performance, we compared our proposed framework with the traditional classifiers, for instance, logistic regression, support vector machine, random forest, balanced random forest, Extra Trees, gradient boosting, XGBoost, LightGBM, and CatBoost. The comparison is performed based on accuracy, balanced accuracy, precision, recall, specificity, F1-score, and AUROC. The experimental results show that the proposed framework achieved the strongest overall performance, and achieved 95.35% accuracy as compared with conventional SMOTE and class-weighted baselines. The proposed approach provided more balanced predictive performance across discrimination and classification metrics.