SR-FSL: Sample reconstruction enhanced few-shot learning for real-time motor unit identification from surface...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Deep learning (DL) shows potential for online motor unit (MU) identification from high-density surface electromyogram (HD-sEMG).
• Current DL methods require substantial data, limiting practical application in real-time scenarios.

Purpose of the Study:

• To develop a data-efficient few-shot learning method for online MU identification.
• To enhance DL model performance using a novel sample reconstruction strategy for minimal experimental data.

Main Methods:

• A spatio-temporal neural network was pre-trained on simulated HD-sEMG data.
• A sample reconstruction strategy generated synthetic data to fine-tune the model with limited experimental data.
• HD-sEMG signals were acquired from abductor pollicis brevis muscles using an 8x8 electrode array.

Main Results:

• The proposed method achieved approximately 93% accuracy in online MU identification.
• Effective model fine-tuning was accomplished using only 6 seconds of experimental data.
• The method significantly outperformed existing comparison techniques.

Conclusions:

• This research offers an efficient solution for real-time MU identification.
• The findings are expected to promote DL-based real-time HD-sEMG decomposition for neural-machine interfaces.
• Applications include robotic motor control and rehabilitation medicine.