A Recurrent Deep Network for Gait Phase Identification from EMG Signals During Exoskeleton-Assisted Walking

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Lower limb exoskeletons assist patients with movement disorders to regain autonomous gait.
• Accurate gait phase detection is crucial for exoskeletons to provide timely joint torque assistance.
• Electromyography (EMG) signals precede limb kinematics, offering a potential input for gait phase identification.

Purpose of the Study:

• To develop and evaluate a deep learning model for identifying stance and swing gait phases using EMG signals.
• To assess the model's accuracy and processing speed for real-time application in lower limb exoskeleton control.
• To investigate the model's generalization capabilities across different walking conditions (overground and treadmill) and data sampling rates.

Main Methods:

• A bidirectional Long Short-Term Memory (LSTM) deep learning model was employed.
• EMG data from four leg muscles of 26 healthy subjects were recorded during overground and treadmill walking with an exoskeleton.
• Data were labeled with gait phases using inertial motion sensor kinematics; models were trained and tested under various scenarios including different sampling rates and retraining.

Main Results:

• The model achieved high accuracy, reaching 92.43% for overground and 91.16% for treadmill walking at 500 Hz.
• Online operation simulation showed a processing time of 127 ms per sequence.
• Testing with 1500 Hz data yielded slightly lower accuracies but faster processing (11 ms reduction); retraining improved performance on 1500 Hz data to 87.17% and 89.64%.

Conclusions:

• The proposed deep learning model effectively identifies gait phases from EMG signals.
• The model demonstrates strong potential for integration into lower limb rehabilitation exoskeleton control systems.
• The balance of accuracy and rapid processing times makes it suitable for real-time exoskeleton assistance.