Creation of an artificial intelligence model for intubation difficulty classification by deep learning (convolutional...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anesthesiology

Background:

• Tracheal intubation is crucial for airway management but can be challenging in emergency settings.
• There is a need for objective tools to assess intubation difficulty for less experienced medical personnel.
• Artificial intelligence (AI) shows potential in medical applications, including image analysis.

Purpose of the Study:

• To develop an AI model utilizing convolutional neural networks (CNNs) to predict tracheal intubation difficulty.
• To classify intubation difficulty based on patient facial images.
• To link facial image data with actual intubation difficulty assessments.

Main Methods:

• Facial images were collected from patients undergoing surgery.
• Anesthesiologists rated intubation difficulty as "Easy" or "Difficult".
• A deep learning AI model (CNN) was trained using facial images and corresponding difficulty ratings. Performance was evaluated using receiver operating characteristic curves (AUC).

Main Results:

• The best AI model achieved an Area Under the Curve (AUC) of 0.864.
• The model demonstrated 80.5% accuracy, 81.8% sensitivity, and 83.3% specificity.
• Class activation heat maps indicated the AI focused on facial contours and neck area for classification.

Conclusions:

• This study demonstrates the first application of deep learning (CNN) for classifying tracheal intubation difficulty from facial images.
• The developed AI model shows significant potential for aiding medical staff, particularly those with limited experience.
• The AI model could be a valuable tool in emergency situations and under general anesthesia for improving intubation success rates.