A Subject-Transfer Framework Based on Single-Trial EMG Analysis Using Convolutional Neural Networks.

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Signal Processing

Background:

• Electromyography (EMG)-based myoelectric interfaces are crucial for restoring function in individuals with physical disabilities.
• Decoding user movement intention from EMG signals is vital for controlling external devices effectively.
• High intra-user variability in EMG signals presents a significant challenge for interface performance.

Purpose of the Study:

• To propose a novel subject-transfer framework for robust hand movement decoding from EMG signals.
• To address and mitigate the impact of intra-user variability on myoelectric interface performance.
• To enhance the accuracy and reliability of EMG-based control systems.

Main Methods:

• A subject-transfer framework utilizing pre-trained convolutional neural network (CNN) classifiers is proposed.
• Supportive CNN classifiers are selected and fine-tuned for a target subject using single-trial analysis.
• Hand movements are classified by aggregating (voting) the outputs from multiple supportive CNN classifiers.

Main Results:

• The proposed framework demonstrated improved hand movement decoding performance compared to self-decoding methods.
• Validation on NinaPro databases 2 and 3 showed enhanced accuracy in both healthy and amputee subjects.
• The framework proved robust against intra-user variability in EMG signal patterns.

Conclusions:

• The developed subject-transfer framework offers a significant advancement for EMG-based myoelectric interfaces.
• This approach provides a valuable tool for improving the control of external devices for users with physical impairments.
• The method effectively enhances decoding accuracy by leveraging data from multiple subjects to overcome individual variability.