Machine learning to predict late respiratory support in preterm infants: a retrospective cohort study

Bronchopulmonary dysplasia (BPD) has been a critical morbidity in preterm infants. To improve our definition and prediction of BPD is challenging yet indispensable. We aimed to apply machine learning (ML) to investigate effective models by using the recently-proposed and data-driven definition to predict late respiratory support modalities at 36 weeks' post menstrual age (PMA). We collected data on very-low-birth-weight infants born between 2016 and 2019 from the Taiwan Neonatal Network database. Twenty-four attributes associated with their early life and seven ML algorithms were used in our analysis. The target outcomes were overall mortality, death before 36 weeks' PMA, and severity of BPD under the new definition, which served as a proxy for respiratory support modalities. Of the 4103 infants initially considered, 3200 were deemed eligible. The logistic regression algorithm yielded the highest area under the receiver operating characteristic curve (AUROC). After attribute selection, the AUROC of the simplified models remain favorable (e.g., 0.801 when predicting no BPD, 0.850 when predicting grade 3 BPD or death before 36 weeks' PMA, and 0.881 when predicting overall mortality). By using ML, we developed models to predict late respiratory support. Estimators were developed for clinical application after being simplified through attribute selection.

Automated summary

This study aimed to improve the definition and prediction of bronchopulmonary dysplasia (BPD) by using machine learning (ML) models to predict respiratory support modalities in preterm infants at 36 weeks' post menstrual age. Data from very-low-birth-weight infants born between 2016 and 2019 were collected, and 24 attributes and 7 ML algorithms were used for analysis. Logistic regression algorithm was found to have the highest accuracy, and simplified models were developed for clinical application.