Energy-efficient real-time 4-stage sleep classification at 10-second resolution.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Sleep Medicine

Background:

• Polysomnography is the gold standard for sleep stage classification but is expensive and impractical for home use.
• There is a need for cost-effective, long-term sleep monitoring solutions suitable for wearable devices.
• Electrocardiogram (ECG) signals contain valuable information for sleep stage detection.

Purpose of the Study:

• To develop an energy-efficient algorithm for classifying four sleep stages (wake, REM, light, deep) using single-lead ECG signals.
• To optimize machine learning and deep learning models for resource-constrained wearable applications.
• To evaluate the performance and energy efficiency of the proposed SleepLiteCNN model.

Main Methods:

• Two windowing strategies were employed: 5-minute windows for machine learning and 30-second windows for deep learning, enabling 10-second resolution.
• Various machine learning and deep learning models were evaluated, leading to the design of SleepLiteCNN.
• 8-bit quantization and Field-Programmable Gate Array (FPGA) deployment were used to further optimize energy consumption and resource usage.

Main Results:

• Deep learning models achieved high accuracy (92%) but were energy-intensive.
• SleepLiteCNN demonstrated strong performance with 89% accuracy and 89% F1-score while minimizing energy use.
• Quantization reduced energy consumption to 5.48 μJ per inference, maintaining 90% accuracy and F1-score, with significant resource reduction on FPGA.

Conclusions:

• SleepLiteCNN offers a practical and energy-efficient solution for continuous ECG-based sleep staging on wearables.
• The developed method enables real-time sleep monitoring, crucial for diagnosing sleep disorders.
• This approach facilitates the development of next-generation wearable devices for personalized health management.