Q-PPG: Energy-Efficient PPG-Based Heart Rate Monitoring on Wearable Devices

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning

Background:

• Photoplethysmography (PPG) sensors in wrist-worn devices monitor heart rate (HR).
• Motion Artifacts (MAs) degrade PPG signal quality, impacting HR accuracy.
• Existing methods often use hand-tuned parameters, limiting real-world applicability.

Purpose of the Study:

• To develop accurate and lightweight deep learning models for HR monitoring resistant to MAs.
• To address the complexity of deep learning models for deployment on resource-constrained wearable devices.

Main Methods:

• A design space exploration methodology was used to generate Temporal Convolutional Networks (TCNs) from a seed model.
• Two Neural Architecture Search (NAS) tools and a hardware-friendly quantizer were employed.
• Models were tested on the PPG-Dalia dataset and deployed on an embedded STM32WB55 microcontroller.

Main Results:

• The most accurate model achieved state-of-the-art Mean Absolute Error (MAE) on the PPG-Dalia dataset.
• Quantized TCNs demonstrated suitability for real-time execution on embedded platforms.
• The best quantized network achieved 4.41 BPM MAE with low energy consumption (47.65 mJ) and memory footprint (412 kB).
• A smaller network achieved 8 BPM MAE with a minimal memory footprint (1.9 kB) and energy consumption (1.79 mJ).

Conclusions:

• The proposed methodology successfully generates highly accurate and extremely lightweight TCNs for HR monitoring.
• The developed models are suitable for real-time deployment on embedded wearable devices, overcoming limitations of traditional methods.