Feature stability and setup minimization for EEG-EMG-enabled monitoring systems

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Home healthcare delivery is crucial for reducing costs and infection risks, as highlighted during the SARS-CoV2 pandemic.
• Wearable devices for motor training applications require minimized setups for movement and brain signal monitoring.
• High-dimensional electroencephalography (EEG) and electromyography (EMG) data present challenges in data processing and sharing for gesture recognition.

Purpose of the Study:

• To develop a minimized and interpretable EEG-EMG monitoring system for hand gesture recognition.
• To reduce the dimensionality of EEG and EMG data while maintaining classification accuracy.
• To enhance the interpretability of machine learning models in wearable health applications.

Main Methods:

• Fused high-dimensional EEG and EMG data into magnitude squared coherence (MSC) signals.
• Extracted features from MSC signals using various algorithms for binary classification.
• Implemented a mapping-and-aggregation strategy to improve the interpretability of machine learning results.

Main Results:

• Achieved very low misclassification errors ($$) with a significantly reduced feature set ($$).
• Identified consistent patterns of centro-parietal brain area and arm muscle activation in the 8-80 Hz frequency band across algorithms.
• Demonstrated the effectiveness of the proposed approach for reliable gesture recognition.

Conclusions:

• The study presents a significant step towards a minimized and reliable EEG-EMG setup for gesture recognition.
• The developed system is suitable for remote motor training and other home healthcare applications.
• The findings align with previous literature, validating the identified neural and muscular activation patterns.