Incorporating respiratory signals for machine learning-based multimodal sleep stage classification: a large-scale...

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Machine Learning

Background:

• Poor sleep quality is linked to numerous diseases, necessitating accurate sleep monitoring.
• Traditional sleep laboratories are expensive and inaccessible for widespread, long-term use.
• Wearable sensors offer a feasible alternative for unobtrusive, at-home sleep tracking.

Purpose of the Study:

• To systematically compare actigraphy-based sleep staging with multimodal approaches.
• To investigate the impact of incorporating respiration rate variability (RRV) alongside actigraphy (ACT) and heart rate variability (HRV).
• To evaluate the performance of machine and deep learning algorithms on a large-scale sleep dataset.

Main Methods:

• Utilized a public sleep dataset with over 1,000 recordings.
• Compared actigraphy (ACT) alone against multimodal approaches including HRV and RRV.
• Extracted respiratory information using ECG-derived respiration (EDR) features and compared with respiration belts.
• Employed state-of-the-art machine and deep learning algorithms, including Long Short-Term Memory (LSTM).

Main Results:

• Incorporating RRV features significantly improved sleep stage classification accuracy (Matthews Correlation Coefficient - MCC).
• LSTM algorithms demonstrated superior performance, achieving a median MCC of 0.51 for AASM standard sleep staging.
• Respiratory information notably enhanced the detection of Wake and Rapid Eye Movement (REM) sleep epochs.

Conclusions:

• Multimodal sleep monitoring, particularly including respiratory data, enhances the accuracy of sleep stage classification.
• Deep learning models like LSTM show promise for robust sleep analysis using wearable sensor data.
• These findings support the transition of sleep monitoring from clinical settings to home-based, unobtrusive solutions.