Exploiting the Intertemporal Structure of the Upper-Limb sEMG: Comparisons between an LSTM Network and...

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Myoelectric interfaces offer significant potential for clinical and other applications by mapping neuromuscular activity to motion.
• Current surface electromyogram (sEMG) decoding methods often overlook the time-series nature of the signals, limiting performance.
• Effective decoding of movement intent is crucial for natural and intuitive control of prosthetic devices and other interfaces.

Purpose of the Study:

• To compare the performance of traditional cross-sectional pattern recognition methods with a novel approach utilizing Long Short-Term Memory (LSTM) neural networks for sEMG-based movement intent decoding.
• To evaluate the efficacy of LSTM networks in capturing the temporal dynamics of sEMG signals for improved gesture recognition.
• To assess the generalizability of the LSTM approach across different signal processing strategies and subject populations, including transradial amputees.

Main Methods:

• A comparative study was conducted evaluating traditional cross-sectional pattern recognition classifiers against a Long Short-Term Memory (LSTM) neural network classifier.
• The LSTM classifier was designed to leverage the inherent temporal ordering of surface electromyogram (sEMG) signals.
• Performance was evaluated using both conventional sEMG features and raw sEMG data across healthy subjects and transradial amputees.

Main Results:

• The Long Short-Term Memory (LSTM) neural network approach demonstrated superior performance compared to traditional cross-sectional gesture recognition techniques.
• The enhanced decoding accuracy was consistent regardless of whether the LSTM classifier utilized conventional features or raw sEMG data.
• The findings were validated across both healthy participants and individuals with transradial amputations, indicating broad applicability.

Conclusions:

• LSTM neural networks offer a significant advancement in decoding movement intent from sEMG signals by effectively utilizing temporal information.
• The LSTM approach provides a more robust and accurate method for myoelectric control compared to conventional pattern recognition techniques.
• This study highlights the potential of deep learning, specifically LSTMs, to enhance the functionality and user experience of natural myoelectric interfaces for diverse user groups.