Transfer learning using low-dimensional subspaces for EMG-based classification of hand posture.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Robotics

Background:

• Muscle synergies are fundamental neural control strategies for movement.
• Understanding task invariance of muscle synergies is crucial for advanced prosthetic limb control.
• Electromyographic (EMG) signal analysis provides insights into muscle activation patterns.

Purpose of the Study:

• To evaluate the task invariance of muscle synergies using a transfer learning approach.
• To assess the feasibility of applying this invariance for prosthetic hand control.
• To investigate the robustness of muscle synergy representations across different tasks.

Main Methods:

• Decomposition of EMG data from voluntary tasks (finger spelling, grasp postures, unconstrained exploration) into muscle synergies using non-negative matrix factorization.
• Reconstruction of cross-task weights using base matrices from different tasks.
• Classification of reconstructed weights using Support Vector Machine (SVM) and Extreme Learning Machine (ELM) classifiers.

Main Results:

• Both SVM and ELM classifiers demonstrated significantly higher performance compared to randomized controls.
• Lower-rank EMG representations showed robust performance within and between tasks.
• Results support the hypothesis of functional invariance in multi-muscle synergies.

Conclusions:

• Muscle synergies exhibit functional invariance across various hand/forearm tasks.
• This invariance can be leveraged for efficient prosthetic hand calibration by transferring learned EMG patterns.
• The findings pave the way for more intuitive and adaptive prosthetic control systems.