EMaGer: A Wearable Full-Circumference HD-EMG Sensor and Data Augmentation Method for Robust Hand Gesture Recognition

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Wearable Technology

Background:

• Myoelectric prostheses require robust gesture recognition for effective control.
• Calibration and electrode placement variability pose significant challenges in electromyography (EMG) systems.
• Existing EMG sensors often lack the adaptability needed for consistent inter-session performance.

Purpose of the Study:

• To develop a novel high-density electromyography (HD-EMG) system for improved gesture recognition robustness.
• To introduce an original data augmentation technique to enhance system performance and reduce calibration burden.
• To demonstrate the synergistic benefits of co-designing EMG sensors with gesture inference algorithms.

Main Methods:

• Development of EMaGer, a 360° 64-channel HD-EMG bracelet with homogeneous electrode density.
• Implementation of an Array Barrel Shifting Data Augmentation (ABSDA) technique.
• Utilizing deep learning, specifically convolutional neural networks, for gesture classification.

Main Results:

• Achieved 76.98% inter-session classification accuracy for 6 gestures, a significant improvement over baseline intra-session accuracy of 93.75%.
• Demonstrated rotation invariance around the arm axis, increasing robustness to electrode movement.
• Showcased superior performance compared to state-of-the-art sensors when applying the same methods.

Conclusions:

• The co-design of EMG sensors and gesture inference algorithms is crucial for overcoming state-of-the-art challenges.
• The EMaGer system and ABSDA technique substantially reduce the need for frequent recalibration in EMG-based control systems.
• This approach offers clinical relevance by simplifying the setup and calibration of myoelectric prosthetic devices.