Deep learning of snoring to predict collapse site, treatment outcomes, and its clinical translation in obstructive sleep apnea

The National Heart, Lung, and Blood Institute awarded Kennesaw State University $235,520 on September 1, 2026, under the Cardiovascular Diseases Research program (CFDA 93.837) for a project grant to develop deep learning models that extract acoustic signatures from snoring sounds to identify the anatomical site of airway collapse in obstructive sleep apnea (OSA) patients and predict treatment outcomes.

The funded work represents a secondary analysis of existing, clinically collected datasets that include synchronized snoring recordings, drug-induced sleep endoscopy findings, and sleep-recording data. The research team proposes to train deep learning models to phenotype the airway by correlating acoustic features of snores with the structures that vibrate during upper airway obstruction, classified using the VOTE system (velum, lateral pharyngeal walls, tongue base, and epiglottis). The project aims to create a non-invasive, widely accessible biomarker to guide personalized therapy selection—including mandibular advancement devices, hypoglossal nerve stimulation, pharyngeal surgery, and positional therapy—for patients who cannot tolerate or do not respond to continuous positive airway pressure.

Performance takes place in Kennesaw, Georgia. The award period runs from September 1, 2026, through August 31, 2028.

Abstract

Continuous positive airway pressure (CPAP) is an effective treatment for OSA, in which the patient must wear an oxygen mask every night. Due to discomfort, only 30-50% of patients can tolerate CPAP. Alternatives to CPAP are available, such as mandibular advancement devices (MAD), hypoglossal nerve stimulation (HGNS), pharyngeal surgery, and positional therapy. The success of these treatments often depends on the site of UA obstruction. For example, HGNS therapy electrically activates tongue muscles. The eligibility to get HGNS treatment requires ruling out of collapse of the soft palate, as patients with this collapse pattern respond poorly to HGNS. Thus, IT is clear that treatments are highly site-of-collapse specific. The primary site of collapse is categorized using the vote classification system, which denotes the velum (V), lateral pharyngeal walls (O), tongue base (T), and epiglottis (E). Drug-induced sleep endoscopy (DISE) is the gold standard for determining vote, where an endoscope is used while the patient is under sedation to visualize where the collapse occurs. Due to its invasiveness and high cost, DISE is unavailable to most patients. Therefore, there is an urgent need for a widely accessible, non-invasive biomarker that can phenotype the airway and guide personalized therapy. Snoring can serve as such a biomarker, as the acoustic features of snores can reflect the structures that vibrate during UA obstruction. The central hypothesis of this project is that deep learning models can extract these acoustic signatures from snoring to identify the anatomical site of collapse and predict treatment outcomes. This work represents a secondary analysis of existing, clinically collected datasets that include synchronized snoring, DISE, and sleep-recording data. We propose a three-part strategy to test this hypothesis. Aim 1 will develop deep learning architectures that combine handcrafted acoustic features, self-supervised speech embeddings, and spectrogram-based models to classify the site of collapse from DISE snoring recordings. Aim 2 will extend the framework to predict HGNS treatment response using pre-treatment snore recordings. Aim 3 will systematically evaluate how much snore sampling frequency is needed for reliable classification and validate the models using routine polysomnography (PSG) snore sensors. Successful completion of this project will establish a scalable, non-invasive snore-based biomarker for identifying UA collapse sites and predicting treatment responsiveness.