Online Unsupervised Adaptation of Latent Representation for Myoelectric Control During User-Decoder Co-Adaptation

Area of Science:

• Biomedical Engineering
• Neuroscience
• Robotics

Background:

• Myoelectric control interfaces map electromyographic (EMG) signals to control external devices, crucial for active prosthesis control.
• Signal characteristics drift over time, destabilizing static interfaces and necessitating online user-decoder co-adaptation.
• Existing adaptation methods face challenges like costly data labeling and slow convergence.

Purpose of the Study:

• To develop a rapid, unsupervised decoder adaptation method for myoelectric control interfaces.
• To address the limitations of current online adaptation techniques, focusing on efficiency and reduced data requirements.

Main Methods:

• Utilized an autoencoder to extract motor intent representations in latent manifold space.
• Implemented an online unsupervised adaptation scheme using Moore-Penrose Inverse for rapid network re-training.
• Tracked evolving signal manifolds to maintain stable control despite signal changes.

Main Results:

• The proposed adaptation scheme demonstrated approximately 50% faster convergence compared to state-of-the-art methods.
• Online experiments showed comparable robustness to supervised methods in cursor and prosthetic hand control tasks.
• Prosthetic hand cup relocation task completion time post-adaptation with electrode shift matched baseline performance.

Conclusions:

• The unsupervised adaptation method significantly enhances the accessibility and reliability of myoelectric control interfaces.
• This approach effectively bridges the translational gap by enabling robust co-adaptation during real-time operation.
• The method offers a promising solution for stable and intuitive control of advanced prosthetic devices.