Deep Multi-Scale Fusion of Convolutional Neural Networks for EMG-Based Movement Estimation.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Robotics

Background:

• Electromyography (EMG)-based motion estimation is crucial for advanced applications like myoelectric control.
• Simultaneously estimating joint angle and velocity from EMG signals presents significant challenges.

Purpose of the Study:

• To develop and validate a novel method for accurate simultaneous estimation of joint angle and velocity using EMG data.
• To assess the model's performance under isotonic, isokinetic, and dynamic contraction conditions.

Main Methods:

• A novel two-stream Convolutional Neural Network (TS-CNN) architecture was employed.
• TS-CNN learns features from raw EMG data at different scales for motion prediction.
• The model was evaluated for elbow flexion and extension movements.

Main Results:

• The TS-CNN demonstrated robust performance, outperforming conventional CNN and other literature methods.
• High R² values were achieved for both joint angle (up to 0.81) and velocity (up to 0.79) estimation.
• Model performance varied with experimental conditions, with velocity estimation excelling in isokinetic and angle estimation in dynamic contractions.

Conclusions:

• The proposed TS-CNN method offers accurate and robust EMG-based simultaneous joint angle and velocity estimation.
• Experimental conditions significantly influence the accuracy of EMG-based motion prediction.
• The findings have implications for improving myoelectric control systems and prosthetic devices.