Self-Supervised Speech Representations for Sleep Apnea Severity Prediction.

Area of Science:

• Sleep Medicine
• Biomedical Engineering
• Speech Processing

Background:

• Obstructive sleep apnea (OSA) is common but often undiagnosed.
• Polysomnography (PSG) is the gold standard but is resource-intensive.
• Speech characteristics are linked to OSA and may offer a screening alternative.

Purpose of the Study:

• To develop a speech-based pipeline for OSA screening and severity estimation.
• To utilize self-supervised learning (SSL) and multimodal acoustic features.
• To overcome limitations of prior speech-based OSA research focusing only on acoustic features.

Main Methods:

• A novel fusion framework combining SSL-derived speech representations and traditional acoustic features.
• Analysis of vowel phonation during wakefulness for OSA screening and apnea-hypopnea index (AHI) estimation.
• Data collected from diverse participants across three research sites.

Main Results:

• Balanced accuracies for OSA screening were 0.79 (AHI ≥10) and 0.74 (AHI ≥30) in females, and 0.80 and 0.78 in males.
• AHI estimation showed mean absolute errors of 12.0 events/hour (r=0.63) in females and 14.7 events/hour (r=0.52) in males.
• The models demonstrated generalization across diverse demographic groups.

Conclusions:

• Speech, particularly vowel phonation, is a feasible biomarker for OSA risk and severity.
• This speech-based approach offers a low-burden, cost-effective method for OSA screening.
• The findings have significant implications for scalable sleep health assessments.