Snoring-based obstructive sleep apnea screening and AHI estimation with an adapted pretrained audio model

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence

Background:

• Obstructive sleep apnea (OSA) is a prevalent sleep disorder linked to severe health issues.
• Snoring is a key symptom of OSA, driving research into non-contact detection methods.
• Limited snoring data and patient variability challenge the generalizability of current detection models.

Purpose of the Study:

• To propose a novel snoring-based framework for OSA detection and Apnea-Hypopnea Index (AHI) estimation.
• To adapt the Wav2vec 2.0 model for efficient and accurate snoring analysis.
• To improve the generalization of OSA detection models to new subjects.

Main Methods:

• Utilized Wav2vec 2.0, a pretrained audio model, for OSA detection.
• Implemented layer dropping to reduce computational load while preserving acoustic features.
• Introduced a cross-layer attention mechanism to integrate multi-level acoustic information.
• Developed a regression model for AHI estimation using simulated clinical statistics.

Main Results:

• The proposed framework demonstrated superior performance compared to existing methods on subject-dependent and independent datasets.
• Achieved higher accuracy and reduced computational cost.
• The AHI estimation model yielded a Mean Absolute Error (MAE) of approximately 11 events/hour and a Pearson Correlation Coefficient (PCC) of 0.81.

Conclusions:

• Task-adapted pretrained models improve generalization to unseen subjects.
• The study provides evidence for the feasibility of using snoring for OSA initial screening and AHI estimation.
• This approach offers a promising avenue for non-contact sleep apnea monitoring.