MorpheusNet: Resource efficient sleep stage classifier for embedded on-line systems.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Sleep Medicine

Background:

• Manual sleep stage classification (SSC) is time-consuming and limits therapeutic applications.
• Deep learning models for SSC exist but require significant computational resources, hindering real-time and edge deployment.
• Wearable devices offer potential for scalable sleep-based therapies if SSC can be automated efficiently.

Purpose of the Study:

• To develop a compact, power-efficient deep learning model for real-time, on-device sleep stage classification.
• To enable the deployment of sleep-based therapies on embedded systems with constrained hardware.
• To reduce the computational complexity of SSC models without compromising accuracy.

Main Methods:

• Developed a novel, compact deep learning architecture for sleep stage classification.
• Optimized the model using 8-bit quantization to reduce memory footprint and enhance power efficiency.
• Tested the model on three public sleep datasets and implemented it on an Arm Cortex-M4 processor.

Main Results:

• The compact model achieved performance comparable to state-of-the-art methods.
• Model complexity was reduced by up to 280 times compared to existing approaches.
• Quantized model showed only a 0.95% average drop in accuracy and achieved 1.6-second latency on an Arm Cortex-M4 for on-line SSC.

Conclusions:

• The developed compact deep learning model enables efficient, real-time, on-device sleep stage classification.
• This approach facilitates the integration of sleep analysis into wearable devices for scalable therapeutic interventions.
• The power-efficient and low-complexity design allows deployment on microcontrollers, overcoming previous limitations.