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

Area of Science:

• Medical Informatics
• Machine Learning
• Neonatal Health

Background:

• Neonatal asphyxia poses a significant threat to newborns, causing mortality and long-term neurological issues.
• Accurate prediction is challenging due to complex risk factors and imbalanced medical datasets.
• Early identification is crucial for timely clinical intervention in high-risk newborns.

Purpose of the Study:

• To develop an imbalance-sensitive machine learning framework for predicting neonatal asphyxia.
• To enhance the representation of informative asphyxia cases in medical datasets.
• To improve the accuracy and reliability of neonatal asphyxia prediction models.

Main Methods:

• Proposed a novel synthetic oversampling framework named HEM-SMOTE.
• Integrated local Euclidean-distance and Mahalanobis-distance for neighbor selection.
• Generated informative synthetic asphyxia samples to address class imbalance.

Main Results:

• The proposed HEM-SMOTE framework achieved superior performance compared to traditional classifiers.
• Attained 95.35% accuracy, outperforming conventional SMOTE and class-weighted baselines.
• Demonstrated more balanced predictive performance across various discrimination and classification metrics.

Conclusions:

• HEM-SMOTE effectively improves minority-class representation for neonatal asphyxia prediction.
• The framework offers a promising solution for accurate and reliable prediction of neonatal asphyxia.
• This approach can aid in timely clinical interventions for at-risk newborns.