sEMG-Based Hand-Gesture Classification Using a Generative Flow Model.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Surface electromyography (sEMG) based hand-gesture classification faces challenges with conventional algorithms due to signal complexity and variability.
• Deep learning models offer improved accuracy and robustness but lack interpretability, hindering clinical adoption.

Purpose of the Study:

• To develop an interpretable deep learning model for accurate hand-gesture classification using sEMG data.
• To address the clinical requirement for comprehensible machine learning models in assistive technologies.

Main Methods:

• A generative flow model (GFM) integrated with a SoftMax classifier was employed for sEMG hand-gesture classification.
• The GFM was utilized to model the distribution of 53 distinct hand gestures from the NinaPro database 5.
• The reverse flow of the GFM was used to ensure the interpretability of learned features.

Main Results:

• The proposed GFM approach achieved an accuracy of 63.86 ± 5.12% in classifying 53 hand gestures.
• The learned features from the GFM demonstrated interpretability, with each dimension correlating to aspects of muscle synergy.
• This method overcomes the 'black box' problem associated with traditional deep learning models.

Conclusions:

• Generative flow models offer a promising solution for interpretable and accurate sEMG-based hand-gesture classification.
• The developed model enhances the potential for clinical applications requiring transparent decision-making processes.
• The findings suggest a link between interpretable deep learning features and underlying muscle synergy.