Self-Supervised Speech Representations for Sleep Apnea Severity Prediction.

Abstract

Obstructive sleep apnea (OSA) is a prevalent yet underdiagnosed condition characterized by repetitive upper airway obstruction during sleep. Current gold standard diagnostic standards rely on polysomnography (PSG), which is resource-intensive. Since upper airway characteristics impact both OSA and speech production, speech processing has emerged as a promising alternative for OSA screening. However, prior work has focused primarily on acoustic features. This study aims to develop a speech-based screening and severity estimation pipeline for OSA using self-supervised learning (SSL) and multimodal acoustic features. We proposed a novel fusion framework combining SSL-derived speech representations from pre-trained neural networks with traditional acoustic features and time-frequency representations of speech phase and magnitude. Elongated vowels recorded during wakefulness were used to screen for OSA at two apnea-hypopnea index (AHI) thresholds (10 and 30 events/hour) and to estimate AHI. Data were collected across three research sites, comprising participants of varied sex, race, and OSA severity. For OSA screening, the models achieved balanced accuracies of 0.79 (AHI $\geq$10) and 0.74 (AHI $\geq$30) in females, and 0.80 and 0.78 in males, respectively. AHI estimation yielded mean absolute errors of 12.0 events/hour (r = 0.63) in females and 14.7 events/hour (r = 0.52) in males. Our results demonstrate the feasibility of using speech, especially vowel phonation during wakefulness, as a biomarker for OSA risk and severity estimation. The approach generalizes well across diverse demographic groups. This study presents a significant step toward accessible, low-burden, and cost-effective OSA screening, with broad implications for scalable sleep health assessments.