Real-time fine finger motion decoding for transradial amputees with surface electromyography

Area of Science:

• Neuroscience and Biomedical Engineering
• Rehabilitation Technology
• Machine Learning for Healthcare

Background:

• Human-machine interfaces (HMIs) are crucial for advanced rehabilitation technologies.
• Individuals with limb loss require improved quality of life through better assistive devices.
• Decoding fine motor skills from biosignals remains a significant challenge.

Purpose of the Study:

• To develop a novel CNN-Transformer model for decoding continuous fine finger motions from surface electromyography (sEMG) signals.
• To enhance control of prosthetic devices for transradial amputees.
• To address challenges in acquiring labeled sEMG data for amputees.

Main Methods:

• Integrated Convolutional Neural Network (CNN) and Transformer architectures for sEMG signal processing.
• Designed a flexible and stretchable epidermal array electrode sleeve (EAES) for high-fidelity signal acquisition.
• Utilized a computer vision (CV) based multimodal data acquisition protocol and transfer learning with few-shot calibration.

Main Results:

• The CNN-Transformer model demonstrated superior performance in decoding finger motions across various evaluation scenarios (intra-session, inter-session, inter-subject).
• The system showed promising zero-shot and few-shot learning capabilities for personalized amputee calibration.
• The EAES ensured comfortable wear and robust signal capture, critical for long-term use.

Conclusions:

• The proposed HMI model effectively decodes fine finger motions from sEMG signals for transradial amputees.
• The combination of advanced hardware and machine learning techniques facilitates intuitive prosthetic control.
• This approach holds significant potential for real-time, personalized control of prostheses and assistive technologies.