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

Area of Science:

• Neonatal Medicine
• Medical Informatics
• Computational Biology

Background:

• Bronchopulmonary dysplasia (BPD) is a significant cause of morbidity in preterm infants.
• Accurate prediction of BPD and its severity is crucial for optimizing respiratory support and outcomes.
• Existing BPD definitions and prediction models require improvement for clinical utility.

Purpose of the Study:

• To apply machine learning (ML) algorithms to predict late respiratory support modalities in preterm infants.
• To utilize a recently proposed, data-driven definition of BPD for improved prediction accuracy.
• To develop simplified ML models for potential clinical application in predicting BPD severity and mortality.

Main Methods:

• Retrospective analysis of very-low-birth-weight infants from the Taiwan Neonatal Network database (2016-2019).
• Utilized 24 early-life attributes and seven ML algorithms to predict outcomes at 36 weeks' postmenstrual age (PMA).
• Evaluated model performance using the area under the receiver operating characteristic curve (AUROC), focusing on logistic regression.

Main Results:

• Logistic regression demonstrated the highest predictive performance (AUROC).
• Simplified ML models after attribute selection maintained favorable AUROCs: 0.801 (no BPD), 0.850 (grade 3 BPD or death before 36 weeks' PMA), and 0.881 (overall mortality).
• Developed and simplified ML estimators for predicting late respiratory support and outcomes.

Conclusions:

• ML models, particularly logistic regression, can effectively predict BPD severity and mortality in preterm infants.
• The data-driven BPD definition aids in developing accurate predictive tools.
• Simplified ML models show promise for clinical implementation to guide respiratory support strategies.