ChatEMG: Synthetic Data Generation to Control a Robotic Hand Orthosis for Stroke.

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning

Background:

• Intent inferral for stroke patient hand orthoses is hindered by data collection challenges and electromyography (EMG) signal variability.
• Traditional methods require extensive labeled data for new conditions, sessions, or subjects, proving time-consuming and burdensome.

Purpose of the Study:

• To introduce ChatEMG, an autoregressive generative model for creating synthetic EMG signals.
• To enable context-specific EMG data expansion using prompts, reducing the need for new labeled data.
• To improve the generalizability and accuracy of intent inferral classifiers for hand orthoses in stroke survivors.

Main Methods:

• Developed ChatEMG, a generative model that synthesizes EMG signals conditioned on provided prompts (EMG sequences).
• Utilized generative training on a large dataset, enabling context-specific generation via prompting.
• Evaluated the utility of synthetic samples for training intent inferral classifiers.

Main Results:

• Synthetic EMG samples generated by ChatEMG are classifier-agnostic.
• The use of synthetic data significantly improved intent inferral accuracy across various classifiers.
• The complete approach was successfully integrated into a single patient session for functional orthosis control.

Conclusions:

• ChatEMG effectively generates context-specific synthetic EMG data, overcoming limitations of traditional data collection.
• This approach enhances the performance of intent inferral classifiers, enabling more accurate control of hand orthoses for stroke survivors.
• This represents the first deployment of an intent classifier trained on synthetic data for functional orthosis control in stroke survivors.