Exploration of Task-Specific Components in Different sEMG Features.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Human-Computer Interaction

Background:

• Surface electromyography (sEMG) signals contain intertwined subject-specific and task-specific components.
• Existing disentangling models require further investigation into the impact of different sEMG measures on performance.

Purpose of the Study:

• To evaluate the effectiveness of various sEMG signal processing measures as inputs for a disentangling model.
• To determine which sEMG features best isolate task-specific components for improved gesture classification.

Main Methods:

• A disentangling model with two encoders and one decoder was utilized.
• The model was tested using raw sEMG signals, EMG envelope, frequency-domain features (short-time Fourier transformation), and four time-domain features.
• Reconstruction of 2-D heatmaps representing task-specific information was performed.

Main Results:

• The combination of four time-domain features yielded the highest accuracy in task-specific component extraction for gesture classification.
• The heatmap derived from time-domain features closely matched physiological muscle activation patterns.
• This consistency highlights the optimal performance of time-domain features.

Conclusions:

• Time-domain features are superior inputs for disentangling sEMG signals into task-specific components.
• This method enhances gesture classification accuracy and provides physiologically relevant muscle activation insights.