Using Multiple Decomposition Methods and Cluster Analysis to Find and Categorize Typical Patterns of EEG Activity in...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) rely on accurate electroencephalography (EEG) signal decomposition.
• Understanding the neural sources of EEG activity is crucial for improving BCI performance in motor imagery tasks.

Purpose of the Study:

• To compare sixteen linear EEG source separation methods for motor imagery BCI.
• To evaluate methods based on information reduction, physiological plausibility (dipolarity, task specificity), and component clustering.

Main Methods:

• Applied sixteen linear decomposition techniques to EEG data from motor imagery experiments.
• Assessed dipolarity of source topography and task specificity of source activity.
• Clustered components using Attractor Neural Network with Increasing Activity.

Main Results:

• Independent Component Analysis (ICA) methods AMICA and PWCICA yielded the most dipolar components and highest information reduction.
• AMICA and PWCICA identified the most task-specific EEG patterns among blind source separation algorithms.
• Common Spatial Pattern (CSP) methods showed superior pattern specificity compared to blind methods.
• Clustering revealed frequent patterns including eye movements, sensorimotor rhythm suppression, and activations in motor-related brain areas.

Conclusions:

• Multi-method EEG decomposition combined with clustering and specificity analysis is effective for processing electrophysiological data.
• ICA methods like AMICA and PWCICA are highly suitable for identifying physiologically plausible and task-specific sources in motor imagery BCIs.