Decomposing Task-Relevant Information From Surface Electromyogram for User-Generic Dexterous Finger Force Decoding.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Robotics

Background:

• Electromyographic (EMG) based motor intent detection algorithms are often user-specific, limiting their broad application.
• High inter-person variability and signal interference pose challenges for developing generic EMG models.

Purpose of the Study:

• To develop a user-generic algorithm for accurate motor intent detection from EMG signals.
• To enable continuous multi-finger force prediction for unseen users.

Main Methods:

• Implemented a feature disentanglement approach using an autoencoder-like architecture.
• Decomposed EMG amplitude features into user-invariant and user-sensitive representations.
• Evaluated the model on eight subjects for concurrent prediction of three-finger force using leave-one-subject-out validation.

Main Results:

• Achieved lower force prediction error (RMSE: 6.91 ± 0.45 % MVC) and higher classification accuracy (83.0 ± 4.5%) compared to conventional EMG amplitude and PCA methods.
• Demonstrated superior performance in user-generic force predictions against state-of-the-art neural networks.

Conclusions:

• The proposed feature disentanglement method offers accurate and user-generic neural decoding for myoelectric control.
• This approach provides significant advancements for developing adaptive and efficient assistive robotic hands.