172 - Ensemble Learning for Predicting Neonatal Birth Trauma Using High-Dimensional Data in the NICU

Publication Number: 172.4696

Collins Odhiambo, University of Illinois College of Medicine, Peoria, IL, United States; Nirzar S. Parikh, OSF Healthcare Children's Hospital of Illinois, Peoria, IL, United States; Gretchen L. Kopec, Children's Hospital of the University of Illinois, Edwards, IL, United States; Adam Cross, University of Illinois College of Medicine, Peoria, IL, United States

Background: Birth trauma continues to be a significant global health challenge, contributing to infant morbidity and mortality. While many birth traumas are associated with identifiable risk factors, some cases arise without clear predictors, complicating prevention efforts.

Objective: To develop and evaluate an ensemble learning model capable of predicting birth trauma in NICU settings. The model leverages high-dimensional data to account for known and unknown risk factors, while addressing the challenge of class imbalance in the dataset.

Design/Methods: A retrospective cohort analysis was conducted using data collected between January 1, 2019, and March 31, 2023, from nine medical centers in the state of Illinois. The dataset consisted of 711 patients and included a wide range of variables such as patient demographics, maternal and neonatal characteristics, vital signs, comorbidities, specific correlating factors for neonatal birth trauma (e.g., death, mechanical ventilation, seizures, meconium aspiration syndrome, culture proven sepsis/meningitis and pneumothorax). The ensemble model combined logistic, random forest, gradient boosting and support vector machine methods using soft voting. SHAP (SHapley Additive exPlanations) values were applied to identify the most important features. Model performance was assessed using various metrics such as sensitivity, specificity, positive predictive value, and negative predictive value.

Results: The classification model achieved an accuracy of 79%. For class 0 (no trauma), precision was 0.81, recall was 0.92, and the F1-score was 0.86. For class 1 (trauma), the model yielded a precision of 0.64, recall of 0.41, and an F1-score of 0.50. The macro-average F1-score was 0.68, highlighting stronger performance for class 0. The weighted average metrics reflected an overall precision of 0.77, indicating the impact of class imbalance. SHAP analysis identified the top three most influential features: antenatal corticosteroids (S-value: 0.4348), APGAR score at 10 minutes (S-value: 0.4044), and meconium-stained fluid (S-value: 0.3868).

Conclusion(s): The ensemble learning model, combined with SHAP analysis, effectively identified key risk factors and predicted birth trauma using high-dimensional data. The results demonstrate the potential of this model to assist healthcare professionals in early identification of at-risk neonates, offering an opportunity for timely intervention and improved patient outcomes in the NICU.