Automatic scoring of drug-induced sleep endoscopy for obstructive sleep apnea using deep learning

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Sleep medicine

Background:

• Obstructive sleep apnea (OSA) treatment is vital for health and economic reasons.
• Drug-induced sleep endoscopy (DISE) assesses upper airway collapse patterns.
• The VOTE classification system (Velum, Oropharynx, Tongue, Epiglottis) quantifies obstruction severity.

Purpose of the Study:

• To develop and validate a deep learning approach for automated VOTE obstruction degree scoring from DISE videos.
• To improve the efficiency and objectivity of DISE analysis.

Main Methods:

• A deep learning model was trained on 281 DISE videos from two sleep clinics.
• Videos were segmented into 5-second clips with VOTE site obstruction annotations.
• Model performance was evaluated against surgeon-annotated VOTE degrees.

Main Results:

• The model achieved a mean F1 score of 70% across all DISE examinations.
• Individual site performance varied: Velum (85%), Oropharynx (72%), Tongue (57%), Epiglottis (65%).
• Performance was consistent across different clinicians and hospitals.

Conclusions:

• Automated scoring of DISE examinations using deep learning is valid and feasible.
• This AI tool can aid in characterizing upper airway collapse for OSA treatment planning.