Attention-Based Deep Recurrent Neural Network to Estimate Knee Angle During Walking from Lower-Limb EMG

Area of Science:

• Biomedical Engineering
• Neuroscience
• Robotics

Background:

• Surface electromyography (sEMG) is a non-invasive technique to measure muscle electrical activity.
• Accurate prediction of joint angles from sEMG can enable adaptive control of wearable robotic systems.
• Cerebral palsy (CP) often involves motor impairments affecting gait and requiring assistive devices.

Purpose of the Study:

• To develop and validate a recurrent neural network (RNN) with gated recurrent units (GRUs) and an attention mechanism for estimating knee joint angles from sEMG during walking.
• To assess the impact of the attention mechanism on estimation accuracy.
• To evaluate the feasibility of this approach in a child with CP using an exoskeleton.

Main Methods:

• Utilized a GRU-RNN architecture incorporating an attention mechanism.
• Trained and validated the model using sEMG data collected during overground walking in healthy adolescents.
• Performed sensitivity analysis to identify key muscles for joint angle estimation.
• Tested the model's performance in a child with CP receiving exoskeleton assistance.

Main Results:

• The attention mechanism significantly improved the accuracy of knee angle estimation by focusing on relevant muscle activation patterns.
• Knee extensor and flexor muscles were identified as most crucial for accurate joint angle prediction.
• The GRU-RNN model successfully estimated knee angle in a child with CP during assisted walking.

Conclusions:

• The proposed GRU-RNN model with an attention mechanism shows initial feasibility for estimating user movement from sEMG.
• This approach holds promise for enhancing the control of robotic exoskeletons, particularly for children with neuromuscular disorders like CP.
• Accurate sEMG-based movement intention detection is vital for developing responsive and effective assistive robotic technologies.