Deep Learning Models Optimization for Gait Phase Identification from EMG Data During Exoskeleton-Assisted Walking.

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Exoskeletons offer promising solutions for motor dysfunction rehabilitation.
• Accurate online control of assistive exoskeletons is crucial for patient gait assistance.
• Surface electromyography (sEMG) signals reflect muscle activity related to movement.

Purpose of the Study:

• To explore Deep Learning (DL) models for online gait phase prediction using sEMG data during exoskeleton-assisted walking.
• To optimize DL models for reduced size and computational cost while maintaining high accuracy.
• To assess the feasibility of online implementation for real-time exoskeleton control.

Main Methods:

• Utilized sEMG data and joint kinematics to predict gait phases (stance/swing).
• Employed a cross-subject design for model generalization.
• Implemented hyperparameter optimization to reduce DL model size and computational demands.
• Simulated a use-case scenario to evaluate online implementation feasibility.

Main Results:

• Identified a DL model achieving approximately 95% accuracy in gait phase classification.
• Significantly reduced the number of parameters in the optimized DL models.
• Achieved a mean computational time of less than 10 ms for the optimized models.
• Introduced the Trade-off Score (TOS) metric for evaluating model cost-performance.

Conclusions:

• The proposed DL approach enables accurate, online gait phase prediction using sEMG data for lower-limb exoskeletons.
• Optimized DL models offer a feasible solution for real-time control, enhancing exoskeleton-assisted rehabilitation.
• The findings support the potential of sEMG-based DL models for improving assistive exoskeleton functionality and patient outcomes.