Binary CorNET: Accelerator for HR Estimation From Wrist-PPG.

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Signal Processing

Background:

• Wrist-worn photoplethysmography (PPG) sensors are widely used for lifestyle monitoring, but motion artifacts (MA) challenge accurate heart rate (HR) estimation.
• Existing deep learning models like CorNET (CNN-LSTM) show promise but are computationally intensive for on-node processing in resource-constrained devices.

Purpose of the Study:

• To develop a low-complexity, energy-efficient hardware implementation for HR estimation from MA-corrupted PPG signals.
• To demonstrate a fully binarized neural network (bCorNET) for real-time HR monitoring on wearable devices.

Main Methods:

• A novel fully binarized network (bCorNET) topology was designed and mapped to hardware.
• Algorithm-to-architecture mapping and energy-efficient implementation techniques were employed.
• The framework was evaluated on 22 subjects from the IEEE SPC dataset.

Main Results:

• The bCorNET framework achieved a Mean Absolute Error (MAE) of 6.67 ± 5.49 bpm for HR estimation.
• The synthesized design operates at 3 GOPS @ 1 MHz, consuming 56.1 μJ per window.
• The system achieved a low latency of 32 ms for HR estimation every 2 seconds.

Conclusions:

• The proposed binarized network (bCorNET) offers a viable solution for low-power, on-node HR estimation from PPG signals.
• This approach significantly reduces computational complexity and energy consumption, making it suitable for wearable applications.
• The efficient hardware implementation demonstrates the potential for real-time, artifact-robust HR monitoring in ambulatory settings.