Deep Cross-User Models Reduce the Training Burden in Myoelectric Control

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Electromyography (EMG) pattern recognition systems face adoption barriers due to extensive data collection and training requirements.
• Current cross-user models lack sufficient performance and require impractical training protocols, hindering commercialization.
• Robust machine learning models necessitate training data that includes confounding factors and multiple repetitions.

Purpose of the Study:

• To develop and evaluate a novel cross-subject EMG pattern recognition framework requiring minimal end-user training data.
• To improve the performance and practicality of myoelectric control systems for both intact-limb and amputee populations.
• To extend the adaptive domain adversarial neural network (ADANN) for effective cross-subject generalization.

Main Methods:

• An adaptive domain adversarial neural network (ADANN) was adapted into a cross-subject framework.
• Performance was compared against single-repetition within-user training and canonical correlation analysis (CCA).
• The study evaluated models on both intact-limb and amputee populations.

Main Results:

• ADANN significantly outperformed CCA in cross-subject performance for both intact-limb (86.8-96.2%) and amputee (64.1-84.2%) populations.
• The adaptation computation time for ADANN was substantially lower than traditional within-subject training protocols.
• The proposed ADANN framework requires very little end-user specific training data.

Conclusions:

• Cross-user models, powered by deep-learned adaptations like ADANN, show promise for generalized pattern recognition-based myoelectric control.
• ADANN offers a viable solution to reduce the training burden associated with EMG pattern recognition systems.
• This approach may overcome current limitations and facilitate wider commercial adoption of myoelectric control technologies.