Spatio-temporal warping for myoelectric control: an offline, feasibility study

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electromyographic (EMG) signal classification is crucial for myoelectric control.
• Current feature extraction methods often analyze EMG signals in the time domain, channel-by-channel, neglecting multi-channel spatio-temporal relationships.

Purpose of the Study:

• To develop a novel feature extraction method for multi-channel EMG signals that captures spatio-temporal dynamics.
• To evaluate the efficacy of the new feature compared to traditional and deep learning methods in myoelectric control applications.

Main Methods:

• A new feature, spatio-temporal warping (STW), was developed by combining Long Short-Term Memory (LSTM) and Dynamic Temporal Warping.
• The STW feature was designed to capture the spatio-temporal relationships within multi-channel EMG signals.

Main Results:

• The STW feature demonstrated superior performance over traditional features, reducing average classification error by 5%-17%.
• STW outperformed Convolutional Neural Networks (CNNs) by 5%-18% and CNN + LSTM models by 2%-14%.
• All performance improvements were statistically significant with large effect sizes.

Conclusions:

• The STW feature enhances EMG signal classification accuracy in a explainable manner compared to current deep learning methods.
• This study provides a foundation for real-time implementation and prosthesis control applications.