Improvement of EMG Pattern Recognition Model Performance in Repeated Uses by Combining Feature Selection and...

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Electromyography (EMG) pattern recognition is crucial for controlling prostheses and rehabilitation robots.
• Electrode shifts degrade EMG classification accuracy, limiting clinical use in repeated applications.
• Adaptive learning offers a solution but incurs significant time costs.

Purpose of the Study:

• To develop an efficient EMG pattern recognition scheme to overcome classification decline caused by electrode shifts.
• To compare the performance of various feature selection and classification methods for robust EMG pattern recognition.
• To identify a method suitable for improving EMG pattern recognition in repeated use scenarios.

Main Methods:

• Evaluated 12 combinations of three feature selection methods (NFS, SFS, PSO) and four classification methods (N-SVM, I-SVM, T-SVM, TI-SVM).
• Tested classification performance on EMG data from 12 subjects over 5 consecutive days.
• Compared the developed scheme against CNN with fine-tuning on small datasets.

Main Results:

• The TI-SVM method achieved the highest classification accuracy (p < 0.05).
• TI-SVM combined with SFS demonstrated superior classification accuracy and efficiency compared to NFS and PSO methods.
• TI-SVM with SFS outperformed CNN with fine-tuning on small datasets (p = 0.001).

Conclusions:

• The TI-SVM combined with the SFS method is highly suitable for enhancing EMG pattern recognition performance in repeated use.
• This efficient scheme addresses the challenge of classification decline due to electrode shifts.
• The findings support the clinical application of EMG pattern recognition in long-term rehabilitation and prosthetic use.