Development of a machine learning-based multicenter study for predicting severe intraventricular hemorrhage in...

To develop and validate machine learning models for predicting severe intraventricular hemorrhage (SIVH) in preterm infants using only prenatal and intrapartum variables, and to identify the optimal model.

This multicenter retrospective cohort study included preterm infants (gestational age ≤34 weeks) who required noninvasive or invasive respiratory support, using data from the Jiangsu Neonatal Respiratory Failure Collaborative Network (2019-2021). Data were split 7:3 into training and validation sets. Eight machine learning algorithms were compared. Performance was assessed by area under the curve (AUC), calibration, and decision curve analysis. An independent cohort served for temporal validation.

Among 3,320 infants, the incidence of SIVH was 6.66%. Six predictors were identified: gestational age, birth weight, endotracheal intubation in the delivery room, chest compressions in the delivery room, surfactant administration in the delivery room, and chorioamnionitis. In the training set, logistic regression achieved the highest mean AUC based on 10-fold cross-validation. Bootstrap internal validation of the LR model yielded a corrected AUC of 0.809. In the independent hold-out validation set, the LR model again demonstrated the best discriminative ability among the eight algorithms (AUC: 0.804), with satisfactory calibration and favorable net benefit. Sensitivity analyses showed that the LR model maintained robust performance in temporal validation using an independent cohort (AUC: 0.822) and across geographic subgroups (AUC range: 0.729-0.885). An online interactive tool was also developed to facilitate clinical application.

A LR model using six readily available perinatal variables provides good discrimination and calibration for predicting SIVH in preterm infants, enabling immediate risk stratification at birth for SIVH occurring during the neonatal hospitalization period.