Enhanced Prediction of Cardiac Risk in Neonates Using Calibrated Ensemble Learning Approaches.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Neonatal Intensive Care

Background:

• Cardiac arrest in newborns presents a critical challenge in neonatal intensive care.
• Early detection is crucial for improving patient outcomes and survival rates.
• Existing risk prediction models often lack the necessary accuracy and generalization capabilities.

Purpose of the Study:

• To develop and evaluate a Heterogeneous Hybrid Machine Learning Model (HHMM) for early cardiac arrest detection in neonatal intensive care.
• To assess the performance of HHMM compared to individual classifiers and conventional ensembles.
• To enhance the reliability and interpretability of cardiac risk prediction models in neonates.

Main Methods:

• A stacking ensemble approach combining XGBoost, Random Forest, and Multilayer Perceptron classifiers.
• Utilizing probability outputs of base classifiers as meta-features for a Random Forest meta-learner.
• Employing a weighted stacking method with performance-based weights and probability calibration.
• Training and testing the model on clinical cardiovascular risk data from neonatal intensive care units.

Main Results:

• The HHMM achieved superior performance over individual classifiers and conventional ensembles.
• Achieved 95% accuracy, 96% precision, 94% recall, and 95% AUC-ROC.
• Demonstrated enhanced ability to capture complex patterns, reduce overfitting, and improve generalization.
• The model's interpretability was enhanced through probability calibration.

Conclusions:

• The HHMM offers a significant advancement in early cardiac arrest prediction for newborns.
• Calibrated ensemble learning and stacking provide a robust framework for neonatal cardiac risk assessment.
• The HHMM is suitable for clinical decision-support systems, paving the way for real-time monitoring and improved neonatal patient outcomes.