Real-Time Prediction for Neonatal Endotracheal Intubation Using Multimodal Transformer Network

Area of Science:

• Neonatal Medicine
• Artificial Intelligence in Healthcare
• Respiratory Care

Background:

• Neonates in NICUs face risks of respiratory decompensation, often necessitating endotracheal intubation.
• Delayed intubation in neonates is linked to increased morbidity and mortality, especially during urgent, unplanned events.

Purpose of the Study:

• To develop and validate a predictive model for endotracheal intubation in neonates.
• To enable timely preparation for intubation, thereby reducing risks associated with delayed procedures.

Main Methods:

• A deep neural network, specifically a multimodal transformer model, was employed.
• The model analyzed time-series data (vital signs, FiO2) and initial clinical information from neonates on noninvasive ventilation.

Main Results:

• The model accurately predicted the need for intubation 3 hours in advance with high performance metrics (AUROC 0.880, F1-score 0.864).
• Validation on an additional dataset confirmed the model's generalization ability (AUROC 0.890, F1-score 0.893).

Conclusions:

• The developed deep neural network model shows significant potential for real-time prediction of intubation needs in neonates.
• This predictive capability can enhance patient safety by facilitating proactive preparation and avoiding high-risk, late intubations.