Researchers trained deep learning models to identify where the upper airway collapses in obstructive sleep apnea patients, using only snoring sounds. The study compared a ResNet-50 convolutional network with an Audio Spectrogram Transformer (AST) on recordings from 37 patients who underwent drug-induced sleep endoscopy, the current reference standard for locating obstructions.
The AST model significantly beat ResNet-50 after statistical correction, achieving F1-scores of 0.59–0.87 and AUCs of 0.87–0.97 for obstructions at the velum, oropharynx, and retropalatal levels. Both models struggled with tongue-base, epiglottis, and retroglossal sites, which the authors attribute to small sample sizes; oversampling did not resolve the issue.
A separate analysis of full-night recordings showed that acoustic profiling could track airway dynamics over time. The authors suggest such models could become a non-invasive tool for overnight monitoring and for supporting treatment planning, though they note the need for larger datasets to improve weaker detection sites.