A fully-automatic semi-supervised deep learning model for difficult airway assessment.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anesthesiology

Background:

• Difficult airway management poses significant clinical challenges.
• Current diagnostic methods for predicting difficult airways have limited accuracy.
• Accurate prediction is crucial for effective treatment planning in anesthesia.

Purpose of the Study:

• To develop a rapid, non-invasive, and cost-effective deep learning model for identifying difficult airway conditions.
• To assess the diagnostic accuracy of an AI-based photographic image analysis system.
• To overcome the limitations of low diagnostic accuracies in current prediction methods.

Main Methods:

• A semi-supervised deep learning model was trained and tested on 1000 patient images captured from 9 viewpoints.
• The model utilized 30% labeled training samples and 70% unlabeled samples.
• Performance was evaluated using accuracy, sensitivity, specificity, F1-score, and AUC.

Main Results:

• The semi-supervised model achieved high performance metrics (e.g., 90.00% accuracy, 0.9435 AUC) using only 30% labeled data.
• Performance was comparable to a fully supervised model (90.50% accuracy, 0.9457 AUC) and expert anesthesiologists (91.00% accuracy, 0.9497 AUC).
• The approach offers a favorable balance between performance and cost, with reduced sample labeling requirements.

Conclusions:

• This is the first study to apply semi-supervised deep learning for predicting difficult mask ventilation and intubation.
• The AI-based image analysis system is a promising tool for identifying patients with difficult airway conditions.
• The developed method demonstrates the potential of AI in improving airway management safety and efficiency.