Transfer learning in hand movement intention detection based on surface electromyography signals.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Electromyography (EMG) signals are used for human-computer interaction.
• Deep learning, specifically Convolutional Neural Networks (CNNs), shows promise for decoding hand movement intention from EMG signals.
• Training deep networks requires large datasets, which are time-consuming to create for individual subjects.

Purpose of the Study:

• To address the challenge of limited data for training deep learning models on EMG signals.
• To propose and evaluate a subject-transfer framework to leverage data from other subjects.
• To propose and evaluate a task-transfer framework to classify complex movements using knowledge from basic movements.

Main Methods:

• Developed two CNN-based architectures for hand movement intention detection.
• Implemented a subject-transfer learning approach using knowledge from multiple subjects.
• Implemented a task-transfer learning approach utilizing knowledge from basic hand movements to classify combined movements, incorporating few-shot learning and fine-tuning.

Main Results:

• Subject-transfer learning improved average classification accuracy on the Nearlab dataset from 92.60% to 93.30% and on NinaPro DB2 from 81.43% to 82.87%.
• The task-transfer approach demonstrated the ability to classify combined hand movements using knowledge from basic movements with minimal target data.
• Classification accuracy for combined hand movements improved by 10% using the task-transfer approach, leveraging learned information from basic movements.

Conclusions:

• Subject-transfer learning effectively compensates for limited data in individual subjects for EMG-based hand movement intention detection.
• Task-transfer learning enables the classification of complex hand movements by transferring knowledge from simpler, basic movements.
• These transfer learning frameworks significantly enhance the practicality and efficiency of deep learning models for EMG signal analysis.