Continuous Joint Kinematics Prediction Using GAT-LSTM Framework Based on Muscle Synergy and Sparse sEMG

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Surface electromyography (sEMG) signals are crucial for motion prediction in rehabilitation, sports, and human-computer interaction.
• Achieving high prediction accuracy with sparse sEMG electrodes is challenging, limiting system reliability.
• Existing methods struggle to effectively utilize the coordinated muscle activity captured by sEMG.

Purpose of the Study:

• To introduce a novel framework, MSGAT-LSTM, for accurate continuous motion prediction using sparse sEMG data.
• To leverage muscle synergy (MS) theory and graph attention networks (GAT) with LSTM for improved prediction.
• To address the limitations of sparse sEMG electrode setups in motion prediction tasks.

Main Methods:

• Developed a novel MSGAT-LSTM framework integrating GAT for relational feature learning and LSTM for temporal dependencies.
• Utilized muscle synergy theory to calculate cosine similarity between sEMG features for assigning edge weights, capturing coordinated muscle contributions.
• Employed graph-based learning to effectively compensate for sparse sEMG electrode limitations.

Main Results:

• MSGAT-LSTM demonstrated superior performance over state-of-the-art methods (e.g., 3DCNN, GCN-LSTM, CNN-LSTM) on Ninapro DB2 and a self-collected dataset.
• The framework achieved significant improvements in motion prediction accuracy, measured by RMSE and R2.
• Incorporating MS into GCN reduced training time by 13% compared to GCN-LSTM, enhancing efficiency.

Conclusions:

• The MSGAT-LSTM framework offers a robust solution for continuous motion prediction using sparse sEMG signals.
• Integrating muscle synergy theory with graph-based deep learning significantly enhances prediction accuracy and computational efficiency.
• This approach holds substantial potential for advancing sEMG-based applications in various fields.