What question did this study set out to answer?

This study aims to evaluate sleep quality non-intrusively by integrating ballistocardiography and audio signals for diagnosing sleep stages.

May 16, 2026Frontiers in Network Physiology0 citationsOpen Access

Evaluating sleep quality in a non-intrusive manner using contactless ballistocardiography and audio signals through a LSTM-TCN machine learning model

DJDominic JaworskiSimon Fraser University TKTae-Ho KimSimon Fraser University BCBohyung ChoiSimon Fraser University

Key Points

This study aims to evaluate sleep quality non-intrusively by integrating ballistocardiography and audio signals for diagnosing sleep stages.
Diagnostic tool combines ballistocardiogram and microphone for data collection during sleep.
Processed cardiorespiratory signals using time, frequency, and nonlinear analysis emphasizing nonlinear characteristics.
Trained a machine learning model utilizing LSTM and TCN to classify sleep stages with accuracy compared to PSG.
Achieved an accuracy of 80.51% for distinguishing between wake/non-REM/REM sleep stages.
Cohen’s κ of 0.65 indicated substantial agreement with PSG results.
Indicated cardiopulmonary coupling measurements enhanced the classification of sleep stages.

Abstract

Background Non-intrusive sleep detection devices are increasingly sought after because conventional sleep studies require multiple body-attached sensors, which are uncomfortable and impractical for routine use. Integrating sensors and devices into a system that measures a network of physiological signals and their interactions (e.g., cardiorespiratory coupling) non-intrusively can help analyze sleep quality without negatively affecting natural sleep. Effectively measuring natural sleep quality can provide awareness that helps individuals adjust daily habits to improve the amount of good-quality sleep and support earlier identification of sleep disorders without worsening them with the uncomfortable polysomnography (PSG) setup. Methods In this study, we introduce a diagnostic tool that combines a ballistocardiogram (BCG) and a microphone as a potential substitute for conventional PSG methods for collecting cardiorespiratory signals while participants sleep. The cardiorespiratory signals were processed in the time, frequency, and nonlinear domains, with an emphasis on nonlinear analysis because of the dynamic nature of physiological processes linked to the nervous system. Also, cardiopulmonary coupling (CPC) was measured to observe the relationship between cardiac and respiratory activity, consistent with network physiology perspectives on coupled subsystem dynamics, which would help the model classify sleep stages. Furthermore, the audio signals were processed through spectral and autocorrelation methods to obtain respiratory activity from sleep sounds. Results These features were used to train a model combining long short-term memory (LSTM) and a temporal convolutional network (TCN), achieving an accuracy of 80.51% (Cohen’s κ = 0.65) for wake/non-REM/REM sleep stages compared with PSG under leave-one-subject-out cross-validation. Conclusion A non-intrusive system for evaluating sleep stages can enable medical professionals to diagnose sleep disorders without negatively affecting physiological data from patients using PSG-based studies. By quantifying cardiorespiratory interactions from contactless sensing, this approach provides a network physiology–aligned framework for longitudinal, in-home monitoring of sleep-related subsystem dynamics.

Ask AI

Helpful

Bookmark

View Full Paper