Ultrasound Predicts Drug-Induced Sleep Endoscopy Findings Using Machine Learning Models

Area of Science:

• Medical Imaging
• Sleep Medicine
• Otolaryngology

Background:

• Drug-induced sleep endoscopy (DISE) is standard for assessing airway collapse but has limitations.
• Ultrasound offers a low-risk, objective method for airway measurements.
• Predicting collapse patterns using ultrasound could improve upon current diagnostic methods.

Purpose of the Study:

• To identify ultrasound-derived anatomical measurements for predicting airway collapse.
• To correlate ultrasound measurements with the VOTE (Velum, Oropharynx, Tongue, Epiglottis) criteria for collapse.
• To evaluate the accuracy of ultrasound in assessing airway collapse severity during sleep.

Main Methods:

• Ultrasonography was performed on 20 adult patients while awake and sedated.
• Concurrent endoscopy was performed during drug-induced sleep (DISE).
• Generalized Least Squares (GLS) regression and machine learning (ML) models analyzed ultrasound measurements to predict collapse (Pc) and VOTE scores.

Main Results:

• Ultrasound measurements showed significant associations with endoscopic airway collapse (Kendall Tau correlation).
• GLS models demonstrated moderate to strong correlations (R²adj 0.53–0.82) between ultrasound features and collapse severity.
• ML models accurately predicted binary VOTE scores in four of five subsites (F1 score >0.65), with lateral velum collapse being most accurate (F1=0.93).

Conclusions:

• Ultrasound is a reliable tool for identifying airway collapse features during drug-induced sleep.
• Regression (GLS) and machine learning (ML) models show promise in predicting collapse severity using ultrasound measurements.
• Ultrasound-based airway assessment may offer a valuable alternative or adjunct to DISE.