Efficient Multi-Positioned Training for Regression-Based Myoelectric Control: Exploring Transformer Models with...

Area of Science:

• Biomedical Engineering
• Rehabilitation Robotics
• Neuroprosthetics

Background:

• Current upper-limb myoelectric prostheses often use classification models, limiting simultaneous control of multiple degrees of freedom (DOFs).
• Regression-based models offer simultaneous DOF control for more natural movements but typically require extensive training.
• Reducing the training burden for regression-based myoelectric control is crucial for clinical adoption.

Purpose of the Study:

• To investigate and compare different methods for reducing the training burden of regression-based myoelectric control.
• To evaluate the effectiveness of transfer learning and few-shot learning techniques in this context.
• To enable more intuitive and natural control of upper-limb prostheses.

Main Methods:

• Electromyographic (EMG) data were collected from 10 able-bodied participants.
• Five training strategies were tested: traditional training (single/multiple limb positions) and few-shot learning.
• Transfer learning was applied, pre-training models on data from other users and fine-tuning with end-user data.
• Models were evaluated using linear regressor, Convolutional Neural Network, and Transformer architectures.

Main Results:

• Transfer learning combined with few-shot fine-tuning achieved the second-highest median R² of 0.76 across all participants.
• This approach demonstrated significant potential for improving regression-based myoelectric control.
• The study provides a proof of concept for efficient training of multi-DOF prosthetic control.

Conclusions:

• Reduced training routines are feasible for regression-based myoelectric control.
• Transfer learning and few-shot learning are effective strategies for minimizing training data requirements.
• The developed methods pave the way for more natural and intuitive prosthetic limb function.