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

State-of-the-art upper-limb myoelectric prostheses are typically controlled using classification-based models that do not offer simultaneous control of wrist and hand movements (degrees of freedom or DOFs). Regression-based alternatives are being studied because they do offer simultaneous DOF control, yielding more natural movements, but generally require longer training routines. We investigated methods to reduce the training burden for regression-based myoelectric control. Five different methods to train regression models were tested using electromyographic (EMG) data collected from the forearms of 10 able-bodied participants. First, models were either trained traditionally with data from elbows at 90° position, with data from 3 limb positions, or trained by few-shot learning (with fewer data from 3 limb positions). Then, transfer learning was employed to pre-train models using data from all other users, with the models subsequently fine-tuned using either traditional or few-shot learning with new end-user data. The resulting five models were evaluated using linear regressor-, Convolutional Neural Network-, and Transformer-based approaches. Interestingly, the transfer learning pre-trained model in conjunction with few-shot fine-tuning achieved the second-highest median $\mathbf{R}^{2}$ of 0.76 across all participants. Our findings offer a proof of concept for regression-based myoelectric control of multiple DOFs that may more closely resemble natural limb function.