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

Birth trauma remains a significant global health challenge and contributes substantially to neonatal morbidity and mortality. Although machine learning (ML) approaches have shown promise in predicting birth trauma, persistent limitations related to class imbalance, model interpretability, and incomplete characterization of risk hinder their clinical utility. This study aimed to develop and evaluate an ensemble ML framework capable of leveraging high-dimensional clinical data to capture both established and latent risk factors while maintaining transparency for clinical interpretation. We implemented a Super Learner ensemble integrating multiple base learners, class-balancing strategies, and SHAP-based explainability. On an independent test set, the ensemble correctly classified 102 true negatives and 12 true positives, with 25 false negatives and 2 false positives, yielding an overall accuracy of 80.9% (95% CI: 73.4%-87.0%). Model performance exceeded the no-information rate (p = 0.031) and demonstrated fair agreement beyond chance (Cohen's κ = 0.38). Diagnostic performance was characterized by high specificity (98.1%) and positive predictive value (85.7%), but limited sensitivity (32.4%), resulting in a balanced accuracy of 65.3%. These findings highlight the inherent trade-off between sensitivity and specificity in imbalanced neonatal datasets while underscoring the potential of ensemble ML methods to support early identification of neonates at elevated risk for birth trauma. Integrating robust data science methodologies with clinical domain expertise provides a pathway toward more accurate, interpretable, and actionable decision-support tools in neonatal care.