Sound-Based Sleep Staging using Pretrained Speech Foundation Models.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Sleep Science

Background:

• Traditional polysomnography (PSG) for sleep staging is expensive and inconvenient.
• Wearable devices have limitations like motion artifacts and poor skin contact.
• Non-contact sleep monitoring is needed for large-scale and home-based applications.

Purpose of the Study:

• To develop a non-contact sleep staging method using sound analysis.
• To leverage transfer learning with pretrained speech foundation models for sleep staging.
• To assess the effectiveness of these models in capturing respiratory patterns for sleep staging.

Main Methods:

• Utilized transfer learning with pretrained speech foundation models (e.g., HuBERT).
• Analyzed temporal attention weights within HuBERT embeddings to identify respiratory patterns.
• Focused on sound-based analysis for sleep staging, avoiding physical contact.

Main Results:

• Speech foundation models effectively captured respiratory patterns relevant to sleep staging.
• Demonstrated the potential of sound-based analysis for accurate sleep staging.
• Showcased the capability of repurposed AI models in a novel biomedical application.

Conclusions:

• Repurposing speech foundation models offers a scalable, non-contact sleep staging solution.
• This approach presents a promising alternative to polysomnography for clinical and home settings.
• The method enhances accessibility and adherence for sleep disorder diagnosis and monitoring.