What question did this study set out to answer?

The research aims to develop machine learning models to predict early clinical outcomes in very low birth weight infants using prenatal predictors.

April 19, 2026Open Access

Machine learning-based prognostic modeling of early clinical outcomes in very low birth weight infants: insights from a nationwide cohort on delivery room resuscitation, prematurity-related complications, and mortality

Puntos clave

The research aims to develop machine learning models to predict early clinical outcomes in very low birth weight infants using prenatal predictors.
Included 21,231 very low birth weight infants from the Korean Neonatal Network (2013–2023).
Utilized nineteen prenatal and perinatal predictors for binary classification of 27 clinical outcomes.
Employed three tree-based algorithms optimized with stratified five-fold cross-validation.
Applied the Synthetic Minority Oversampling Technique to address class imbalance.
Evaluated model performance using F1-score and area under the receiver operating characteristic curve (AUC).
Models achieved strong predictive performance for resuscitation (F1-score ≥ 0.90, AUC ≥ 0.92).
Moderate performance was observed for mortality prediction (F1-score 0.56, AUC 0.89).
Key predictors identified included gestational age and birth weight.
Models showed potential for early identification of high-risk infants.

Resumen

Very low birth weight infants (VLBWIs) are at high risk of requiring immediate neonatal resuscitation and developing serious complications or mortality. Early prediction based on prenatal information could improve perinatal preparedness and clinical outcomes. We aimed to develop and evaluate machine learning models that predict the need for delivery room resuscitation, major morbidities, and mortality in VLBWIs using prenatal variables. In this population-based cohort study, we included 21,231 VLBWIs from the Korean Neonatal Network (2013–2023). Nineteen prenatal and perinatal predictors were used to construct binary classification models for 27 clinical outcomes. Three tree-based algorithms (Random Forest, eXtreme Gradient Boosting, and Light Gradient Boosting Model) were optimized using stratified five-fold cross-validation with the Synthetic Minority Oversampling Technique applied for class imbalance. Model performance was evaluated using F1-score, area under the receiver operating characteristic curve (AUC), and other standard metrics. SHapley Additive exPlanations (SHAP) analysis was used to interpret feature contributions. Final models were implemented in a Streamlit-based web application. The models showed strong predictive performance for key outcomes, including resuscitation (F1-score ≥ 0.90, AUC ≥ 0.92), respiratory distress syndrome, and bronchopulmonary dysplasia. Mortality prediction showed moderate performance (F1-score 0.56, AUC 0.89). Predictions for rare complications (such as intraventricular hemorrhage ≥ grade III, necrotizing enterocolitis) yielded lower F1-scores despite high AUCs. Gestational age and birth weight were identified in SHAP analysis as the most influential predictors across outcomes. Machine learning models using prenatal information quantitatively predicted neonatal outcomes in VLBWIs. These models show potential as practical tools for early identification of high-risk infants and for supporting delivery planning and clinical assessment.

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo