Predicting CPAP failure after less invasive surfactant administration (LISA) in preterm infants by machine learning...

Area of Science:

• Neonatal Medicine
• Respiratory Physiology
• Machine Learning in Healthcare

Background:

• Less invasive surfactant administration (LISA) is used for preterm infants with respiratory distress syndrome (RDS) on continuous positive airway pressure (CPAP) to avoid intubation.
• A significant proportion of infants fail LISA and require mechanical ventilation, impacting survival chances.
• Predicting LISA failure is crucial for optimizing respiratory management in neonates.

Purpose of the Study:

• To predict CPAP failure (CPAP-F) after LISA using machine learning (ML) analysis of vital parameter monitoring data.
• To identify key physiological features associated with LISA treatment success or failure.

Main Methods:

• Retrospective analysis of vital parameter data (HR, RR, SpO2, temperature) from preterm infants (<32 weeks GA) receiving LISA.
• Calculation of physiological features in defined time windows around the LISA procedure.
• Evaluation of logistic regression (LR) and support vector machine (SVM) ML models for CPAP-F prediction.

Main Results:

• 18 out of 51 (35%) infants experienced CPAP-F.
• Lower SpO2, temperature, and heart rate variability (HRV) were observed in CPAP-F infants.
• ML models achieved high prediction accuracy (AUC 0.90 for LR, 0.93 for SVM) using GA, HRV, RR, and SpO2 in the first 0.5 hours post-LISA.

Conclusions:

• Machine learning models can effectively predict CPAP-F after LISA in preterm infants using readily available vital signs.
• The predictive performance is highest in the initial 0.5-hour window post-LISA.
• ML-driven prediction offers insights into modifiable factors and supports personalized respiratory care strategies for neonates.