Development of a deep neural network for automated electromyographic pattern classification.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Surface electromyography (sEMG) signal quality assessment is crucial but time-consuming.
• Manual evaluation of sEMG requires expert judgment, limiting processing speed.
• Automating sEMG quality control can significantly improve efficiency in research and clinical settings.

Purpose of the Study:

• To compare the performance of supervised and unsupervised artificial neural networks (ANNs) for automated sEMG quality evaluation.
• To determine if ANNs can achieve human-like accuracy in classifying sEMG signal quality.
• To develop a reliable automated tool for sEMG data processing.

Main Methods:

• Five artificial neural networks (two supervised, three unsupervised) were trained and tested using a large dataset (n=47,000) of manually classified sEMG recordings.
• sEMG data from various lower-limb muscles during motor tasks were utilized.
• Performance was evaluated based on classification accuracy into four categories.

Main Results:

• Unsupervised ANNs showed a 30-40% improvement in classification accuracy compared to supervised ANNs, exceeding 98% accuracy.
• AlexNet, an unsupervised ANN, achieved the highest accuracy at 99.55% with minimal misclassifications.
• The study confirmed that automated ANN-based evaluation matches human-like classification accuracy.

Conclusions:

• Automated sEMG quality evaluation using ANNs is feasible and highly accurate.
• Unsupervised ANNs, especially AlexNet, offer superior performance for this task.
• The developed ANN classifier provides a valuable, publicly available tool for electromyography analysis.