A robust feature adaptation approach against variation of muscle contraction forces for myoelectric pattern recognition-based gesture characterization

The lack of a robust scheme that can withstand the muscle contraction force variations (MCFV) in pattern recognition (PR)-based myoelectric prosthesis is a major challenge that prevents it from being fully realized in clinical settings. To overcome this issue, a novel feature adaptation scheme which partially leverages the non-Euclidean space concept based on Riemann manifold was proposed in this study. The scheme is comprised of three logically connected stages. The first stage leverages the symmetric positive definite (SPD) matrices as features. The second stage reduces the discrepancy between SPDs of different force levels by projecting all the SPDs towards a Riemann mean, while the third stage reinforces the robustness against MCFV by projecting the features toward a common distribution drawn from the training set. While considering the three force levels, the scheme was validated on in-house and public datasets obtained from amputees who performed different wrist and finger movements. The results of the evaluation revealed that the suggested method could greatly address the issue of MCFV with an increment in movement decoding greater than 15.02% accuracy and 16.50% F1-score against other state-of-the-art techniques. Additional investigation on the suitable force level that could be a benchmark for training showed that the moderate force level would give an optimal performance compared to low, or high force level in the presence of MCFV. The findings of the study revealed that the suggested control scheme could be used to adapt to MCFV, which could improve the overall robustness of myoelectric systems in both commercial and clinical applications.