Big data in myoelectric control: large multi-user models enable robust zero-shot EMG-based discrete gesture...

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Myoelectric control using electromyogram (EMG) signals offers intuitive device control for ubiquitous computing.
• Widespread adoption is hindered by the need for personalized machine learning models due to user variability.

Purpose of the Study:

• To demonstrate the feasibility of true zero-shot cross-user myoelectric control without user-specific training.
• To analyze the effectiveness of discrete classification approaches for EMG gesture recognition.
• To investigate factors influencing cross-user model performance and generalizability.

Main Methods:

• Utilized the 612-user EMG-EPN612 dataset for training and testing cross-user models.
• Employed a discrete classification approach, treating entire dynamic gestures as single events.
• Developed and evaluated models on unseen users and a novel dataset with covariate factors.

Main Results:

• Achieved 93.0% classification accuracy for six gestures on 306 unseen users.
• Demonstrated that large-scale data enables robust, generalizable cross-user myoelectric control.
• Showcased model effectiveness in mitigating confounding factors like cross-day variability and limb position.

Conclusions:

• Zero-shot cross-user myoelectric control is viable, challenging the necessity of user-specific models.
• Big data and discrete classification pave the way for more accessible and robust EMG-based interfaces.
• Future research can explore demographic representation and transfer learning for further model enhancement.