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

Abstract

Objectives

Ultrasound is a promising low-risk imaging modality that can provide objective airway measurements that may circumvent limitations of drug-induced sleep endoscopy (DISE). This study was devised to identify ultrasound-derived anatomical measurements that could accurately predict collapse pattern and location based on the VOTE criteria (VOTE: Velum, Oropharynx, Tongue, and Epiglottis).

Methods

Ultrasonography was performed on 20 adult patients of various airway subsites while awake and sedated with concurrent endoscopy performed during drug-induced sleep. Measurements were obtained from the ultrasonographic images, and percent collapse (P_c) was estimated then graded using a standard VOTE score. Generalized Least Squares regression (GLS) was used to establish models predictive of P_c on a continuous scale, while multiple machine learning (ML) models were trained to predict each VOTE score (binary, >50% collapse) from ultrasound measurements.

Results

Measurements of multiple ultrasonographic airway subsites demonstrated associations with endoscopic collapse using Kendall Tau correlation. The GLS models showed moderate to strong correlation between multiple ultrasound features and P_c (R²_adj 0.53–0.82) across all VOTE subsites. ML models accurately predicted binarized VOTE scores from ultrasound measurements in four out of five VOTE subsites (F1 score >0.65), while the VOTE subsite with the most accurately predicted collapse was lateral velum collapse with an F1 score of 0.93 averaged across all models.

Conclusions

Ultrasound is a reliable imaging modality and can identify features of airway collapse during drug-induced sleep. Regression (GLS) and ML models show promise in predicting severity of collapse during DISE with analysis of airway ultrasonographic measurements.

Level of Evidence

3 Laryngoscope, 135:1642–1651, 2025