Non-Intrusive Sleep Staging with Integration of Ballistocardiography and Audio Signals

Area of Science:

• Sleep Science and Technology
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Clinical polysomnography (PSG) for sleep evaluation is invasive and uncomfortable, potentially skewing results.
• There is a need for convenient, non-intrusive methods to accurately assess sleep quality and stages.
• Wearable sensors and audio analysis offer potential for remote and unobtrusive physiological monitoring.

Purpose of the Study:

• To investigate the efficacy of combining audio analysis with non-intrusive sensors for sleep stage classification.
• To develop and validate a minimal-contact system for measuring cardiorespiratory parameters during sleep.
• To assess the feasibility of a fully non-contact sleep detection environment.

Main Methods:

• Cardiorespiratory signals were captured using a ballistocardiogram (BCG) sensor (mattress vibrations) and an Apple Watch Ultra 2 (optical wrist sensors).
• Signals were processed using non-linear methods and autocorrelation functions to generate a dataset.
• A long short-term memory (LSTM) model was trained on this dataset for sleep stage classification, with validation against PSG.

Main Results:

• The system achieved 81% agreement with PSG for classifying REM and non-REM sleep stages.
• Audio and BCG data alone showed approximately 76% agreement for sleep stage classification.
• Incorporating audio features significantly improved sleep stage evaluation in a non-intrusive setup.

Conclusions:

• A fully non-contact system using audio and BCG sensors is viable for accurate sleep detection and classification.
• This method offers a comfortable and minimally disruptive alternative to clinical PSG, improving patient experience.
• The findings support the use of non-intrusive technologies for more naturalistic sleep disorder diagnosis.