Meta-Learning for Fast Adaptation in Intent Inferral on a Robotic Hand Orthosis for Stroke.

Area of Science:

• Rehabilitation Robotics
• Machine Learning
• Neuroscience

Background:

• Collecting labeled training data is a major challenge in machine learning for assistive and rehabilitative robotics.
• Muscle tone, spasticity, and hand function vary significantly in stroke subjects, even within the same individual across sessions.

Purpose of the Study:

• To investigate meta-learning for fast adaptation in intent inferral for robotic hand orthosis control in stroke survivors.
• To mitigate the data collection burden required for adapting neural networks to new subjects or sessions.

Main Methods:

• Proposed MetaEMG, a meta-learning framework tailored for electromyography (EMG) signal processing.
• Applied meta-learning to adapt high-capacity neural networks using small, subject- or session-specific datasets.
• Utilized clinical data from five stroke subjects for experimentation.

Main Results:

• MetaEMG demonstrated improved intent inferral accuracy with limited fine-tuning epochs.
• The approach showed effective adaptation to new sessions or subjects.
• Achieved significant improvements using small, personalized datasets.

Conclusions:

• MetaEMG successfully formulates intent inferral for stroke subjects as a meta-learning problem.
• This meta-learning approach enables rapid adaptation for controlling robotic hand orthoses using EMG signals.
• Paves the way for more personalized and efficient assistive robotic technologies.