Ensemble learning for predicting birth trauma using high-dimensional data in the neonatal intensive care unit.

Area of Science:

• Neonatal Health
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Birth trauma is a major cause of neonatal morbidity and mortality globally.
• Existing machine learning (ML) models for birth trauma prediction face challenges like class imbalance and lack of interpretability.
• Accurate prediction of birth trauma is crucial for timely intervention and improved neonatal outcomes.

Purpose of the Study:

• To develop and evaluate an ensemble ML framework for predicting birth trauma using high-dimensional clinical data.
• To capture both known and hidden risk factors for birth trauma.
• To ensure the ML model's transparency and interpretability for clinical application.

Main Methods:

• Implemented a Super Learner ensemble integrating multiple base learners.
• Utilized class-balancing strategies to address data imbalance.
• Incorporated SHAP (SHapley Additive exPlanations) for model explainability.

Main Results:

• Achieved 80.9% overall accuracy on an independent test set.
• Demonstrated high specificity (98.1%) and positive predictive value (85.7%).
• Showcased a balanced accuracy of 65.3% due to a sensitivity of 32.4%, highlighting a common trade-off in imbalanced datasets.

Conclusions:

• Ensemble ML methods show promise in identifying neonates at risk for birth trauma.
• The developed framework offers a pathway to more accurate and interpretable clinical decision-support tools.
• Integrating data science with clinical expertise is key for advancing neonatal care and reducing birth trauma impact.