Sleep stage classification based on multi-level feature learning and recurrent neural networks via wearable device

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Sleep Medicine

Background:

• Automatic sleep stage classification is crucial for long-term sleep monitoring.
• Wearable devices offer advantages over polysomnography for home sleep studies.
• This research focuses on developing a novel sleep staging method using wearable sensor data.

Purpose of the Study:

• To propose and validate a new algorithm for automatic sleep staging using data from wearable devices.
• To leverage heart rate and wrist actigraphy for accurate sleep classification.
• To assess the algorithm's performance in both resting and comprehensive sleep scenarios.

Main Methods:

• A two-phase approach involving multi-level feature learning and recurrent neural networks (RNNs) was developed.
• Low- and mid-level features were extracted from raw signals to capture temporal and frequency properties.
• Bidirectional long short-term memory RNNs were utilized to learn sequential patterns in sleep data.

Main Results:

• The algorithm achieved weighted precision, recall, and F1 scores ranging from 60.5% to 67.7% across different experimental groups.
• Performance was evaluated using leave-one-out cross-validation in resting and comprehensive sleep datasets.
• Comparative experiments confirmed the algorithm's effectiveness in sleep stage classification.

Conclusions:

• The proposed method is efficient and effective for scoring sleep stages.
• This algorithm is well-suited for application in wearable devices for at-home sleep monitoring.
• The findings support the use of wearable technology for accessible and continuous sleep analysis.