Estimating Sleep-Stage Distribution from Respiratory Sounds via Deep Audio Segmentation

Area of Science:

• Biomedical Engineering
• Sleep Science
• Signal Processing

Background:

• Polysomnography (PSG) is the gold standard for sleep studies but is costly and intrusive.
• Non-invasive biomarkers are needed for routine and home-based sleep monitoring.
• Respiratory dynamics correlate with sleep stages, offering potential for passive monitoring.

Purpose of the Study:

• To develop and validate a framework for estimating sleep stage distribution (Wake, Light, Deep, REM) using respiratory audio.
• To leverage respiratory rate and cycle regularity as non-invasive biomarkers for sleep staging.
• To provide a transparent, contact-free method for sleep monitoring.

Main Methods:

• A Transformer-based architecture was fine-tuned for segmenting respiratory cycles from audio.
• Statistical, spectral, and distributional features were extracted from segmented respiratory patterns.
• Stage-specific regression models, including TabPFN, were employed to predict sleep stage proportions.

Main Results:

• The segmentation module improved accuracy in predicting respiratory rate and cycle duration compared to baseline methods.
• The framework achieved favorable Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) for predicting proportions of all sleep stages.
• The TabPFN model demonstrated consistent top performance in sleep stage proportion prediction.

Conclusions:

• The proposed audio-based framework effectively estimates sleep stage distribution using respiratory signals.
• This approach offers a promising, non-invasive, and transparent method for sleep monitoring.
• The system's interpretable features and avoidance of black-box modeling enhance its potential for clinical support.