A comprehensive analysis of wrist electromyography for hand gesture recognition: Advancing from forearm to wrist wearable devices
Loading...
Date
2025-10
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of New Brunswick
Abstract
For several decades, electromyography (EMG) signals from the muscles of the upper forearm have been used for the myoelectric control of prosthetics in individuals with upper limb amputations. However, with advancements in wearable technologies, there is now increased motivation to integrate EMG sensors into wrist-worn devices such as smartwatches and fitness trackers, which are more familiar to general consumers. This could potentially broaden the scope of commercial applications, including the control of robotics, more immersive gaming consoles, and enhanced human-computer interaction within virtual/augmented reality environments.
This dissertation comprises a series of systematic studies; the first study examined the quality of wrist EMG signals and their viability for hand gesture recognition in direct comparison with the more commonly used forearm EMG signals. An additional study evaluated the stability of wrist EMG-based pattern recognition models across different days. It introduced a novel Inter-Day Feature Set (IDFS) and a novel Adaptive Maximum Independence Domain Adaptation (AdaptMIDA) technique to address natural EMG variations over time.
A final study investigated the generalizability of wrist EMG signals across multiple users and presented an enhanced Temporal Convolutional Network-Bidirectional Long Short-Term Memory (TCN-BiLSTM) deep-learning model aimed at reducing or removing the training burden on new users while maintaining high levels of pattern recognition performance.
The findings indicate that wrist EMG signals possess superior signal quality metrics, including increased signal strength and a higher signal-to-noise ratio compared to forearm EMG signals. Furthermore, wrist EMG signals are less susceptible to motion artifacts and spectral deformations. Wrist EMG-based models also demonstrated greater reliability and resilience to the adverse effects of EMG variations across days. Furthermore, results revealed that wrist EMG signals were more generalizable across users, enhancing the pattern recognition performance of wrist EMG-based deep-learning models and requiring less training data from new users. In all analyses, wrist EMG-based models outperformed their forearm EMG-based counterparts, suggesting significant potential for the broader adoption of wrist EMG-based wearable devices.
Overall, this dissertation advances the understanding of the potential of wrist EMG and proposes solutions to leverage its hand gesture pattern recognition capabilities, paving the way for the development of reliable wrist EMG-based wearables.