300 - Predicting severe intraventricular hemorrhage or early death using machine-learning algorithms in very-low-birth-weight infants: Analysis of a Korean Neonatal Network Database

Poster Number: 300
Publication Number: 300.2752

Background: Severe intraventricular hemorrhage (IVH) in premature infants can lead to serious neurological complications. IVH in preterm infants often occurs when multiple factors coincide or overlap, including changes in cerebral blood flow, abnormalities in autoregulation, and the vulnerability of the periventricular germinal matrix. IVH in preterm infants most frequently occurs shortly after birth, and therapeutic options to dramatically improve the prognosis of established hemorrhage remain limited.

Objective: This retrospective cohort study used the Korean Neonatal Network (KNN) dataset to develop prediction models for severe IVH or early death in very-low-birth-weight infants (VLBWIs) using machine-learning algorithms.

Design/Methods: The analysis included 16,343 VLBWIs with a birth weight of less than 1,500 g, born in Korea and registered in the KNN between 2013 and 2020. However, 6,501 infants with birth weights of less than 500 g, gestational ages less than 23 weeks, or missing data were excluded from the analysis (Figure 1). The outcome was the diagnosis of IVH Grades 3–4 or death within one week of birth. Predictors were categorized into three groups based on their observed stage during the perinatal period (Table 1). The dataset was divided into derivation and validation sets at an 8:2 ratio. Models were built using Logistic Regression with Ridge Regulation (LR), Random Forest (RF), and eXtreme Gradient Boosting (XGB).

Results: In the Stage 1 model, the LR algorithm had the highest AUROC value of 0.63, but its weighted F1 score of 0.63 was lower than that of other algorithms. In the Stage 2 model, the LR and XGB algorithms emerged as the top performers, with a consistent AUROC value of 0.86. In the Stage 3 model, the XGB algorithm exhibited the best overall performance—the AUROC 0.86 and the AUPRC 0.38, respectively. Notably, the XGB model demonstrated the highest performance, achieving a weighted F1 score of 0.91 (Table 2). Furthermore, the importance of variables in the XGB model was highest in terms of gestational age and neonatal resuscitation stage.

Conclusion(s): A machine-learning algorithm has successfully developed the XGB model for predicting IVH and death within one week for VLBWIs. If the model is incorporated into treatment and management protocols, such as neonatal resuscitation, it has the potential to reduce the occurrence of permanent brain injury caused by IVH during the early stages of birth. Furthermore, it is believed that the medical records of the NICUs can be effectively utilized as a clinical decision support system.