Convolutional Neural Network Approach for Elbow Torque Estimation during Quasi-dynamic and Dynamic Contractions.

Area of Science:

• Biomechanics
• Neuroscience
• Robotics

Background:

• Accurate torque estimation is crucial for robotics, prosthesis control, and clinical diagnostics.
• Estimating elbow joint torque dynamically presents significant challenges.
• Surface electromyogram (EMG) signals offer a non-invasive method for muscle activity assessment.

Purpose of the Study:

• To accurately estimate elbow joint torque during flexion and extension under quasi-dynamic and dynamic conditions.
• To evaluate the impact of incorporating mechanical data (position and velocity) on torque estimation accuracy.
• To compare the performance of convolutional neural network (CNN) models using different input data combinations.

Main Methods:

• High-density surface electromyogram (HD-EMG) signals were recorded from biceps brachii, brachioradialis, and triceps brachii muscles of five participants.
• A convolutional neural network (CNN) was employed for torque estimation.
• Models were trained using (1) HD-EMG only, (2) HD-EMG and position, and (3) HD-EMG, position, and velocity data.

Main Results:

• Combining HD-EMG with position data improved R² values by 2.35% (isotonic), 37.50% (isokinetic), and 16.67% (dynamic) compared to HD-EMG alone.
• Further incorporating velocity data alongside HD-EMG and position enhanced R² values by an additional 2.29% (isotonic), 12.12% (isokinetic), and 20.50% (dynamic).
• The inclusion of mechanical data significantly improved the accuracy of dynamic elbow torque estimation.

Conclusions:

• Incorporating mechanical data, specifically position and velocity, alongside HD-EMG signals substantially improves the accuracy of elbow torque estimation.
• CNN models demonstrate effectiveness in leveraging multi-modal data for precise dynamic torque prediction.
• This approach holds promise for enhancing applications in robotics, prosthesis control, and clinical diagnostics.