Feature clustering for robust frequency-domain classification of EEG activity

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Electroencephalogram (EEG) analysis is commonly performed in the frequency-domain.
• Traditional methods use pre-defined frequency bands or numerous narrow ranges, both with limitations.
• These limitations include ignoring temporal/participant variability or overlooking power redundancy.

Purpose of the Study:

• To introduce an unsupervised feature extraction method for EEG data.
• To develop data-driven, participant-specific frequency bands using feature clustering.
• To identify the most useful features for classification tasks.

Main Methods:

• An unsupervised feature extraction method using clustering to agglomerate narrow-band spectral powers.
• A fast correlation-based filter to select relevant features.
• Comparison against wide-band and narrow-band frequency-domain algorithms.

Main Results:

• The feature clustering algorithm achieved balanced classification accuracies over 70% for detecting mental states and tasks.
• The new algorithm demonstrated statistical superiority over traditional wide-band and narrow-band methods.
• This indicates improved performance in EEG-based classification.

Conclusions:

• The developed feature clustering algorithm offers a significant advancement in EEG analysis.
• It provides a promising and effective alternative to conventional frequency-domain approaches.
• The method enhances the accuracy and reliability of EEG data interpretation.