Decoding Gestures in Electromyography: Spatiotemporal Graph Neural Networks for Generalizable and Interpretable...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Deep learning has advanced electromyography (EMG) for upper-limb gesture recognition in various fields.
• Current methods often ignore sensor network topology, limiting feature extraction and model performance.
• This overlooks crucial relational information within EMG sensor networks.

Purpose of the Study:

• To develop novel graph structures for EMG sensor networks that capture spatial and temporal relationships.
• To present Graph Convolution Network (GCN) models for enhanced gesture recognition using these graph structures.
• To improve feature extraction, model generalizability, and interpretability in EMG-based systems.

Main Methods:

• Designed custom graph structures to represent spatial proximity of EMG sensors and temporal adjacency of signals.
• Employed Graph Convolutional Networks (GCNs) to extract and aggregate features from these graph structures.
• Validated the approach on five public EMG gesture recognition datasets.

Main Results:

• Achieved state-of-the-art performance in gesture recognition across multiple datasets.
• Demonstrated the model's ability to provide interpretable insights into muscular activation patterns.
• Showcased high accuracy even with reduced sensor configurations, highlighting potential for practical applications.

Conclusions:

• The proposed graph-based input structure and GCN classifier effectively enhance EMG-based gesture recognition.
• This methodology offers improved feature extraction and model interpretability.
• The approach shows promise for integration into AI-powered rehabilitation and human-computer interaction systems.