Replacing Attention With Modality-Wise Convolution for Energy-Efficient PPG-Based Heart Rate Estimation Using...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Continuous heart rate (HR) monitoring via photoplethysmography (PPG) is crucial for wearables.
• Motion artifacts (MA) from arm movements degrade PPG signal accuracy.
• Existing deep learning (DL) fusion methods are often too large for resource-constrained wearable devices.

Purpose of the Study:

• To develop a novel, lightweight deep learning architecture (PULSE) for improved sensor fusion in HR tracking.
• To create a knowledge distillation mechanism for deploying efficient DL models on microcontrollers.
• To evaluate the performance and resource efficiency of the proposed models against state-of-the-art methods.

Main Methods:

• A novel lightweight DL architecture, PULSE, utilizing a multi-head cross-attention layer for temporal feature fusion.
• A relation-based knowledge distillation technique to train a smaller student network with modality-wise convolutions.
• Post-training quantization applied to the student model for microcontroller deployment.

Main Results:

• PULSE achieved near state-of-the-art performance on the PPG-DaLiA dataset and reduced mean absolute error (MAE) by 22.6% on WESAD.
• The distilled student model demonstrated MAE of 4.81 BPM on PPG-DaLiA, comparable to state-of-the-art.
• The student model achieved a 10.9× lower memory footprint (37.9 kB) and 45.9× lower energy consumption (0.577 mJ).

Conclusions:

• The proposed PULSE architecture and its distilled student model offer an effective and efficient solution for accurate HR monitoring in wearable devices.
• The lightweight design and reduced resource requirements make the models suitable for real-time deployment on microcontrollers.
• This work addresses the challenge of deploying complex DL algorithms on edge devices for physiological signal processing.