Brain EEG Time-Series Clustering Using Maximum-Weight Clique.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Brain electroencephalography (EEG) data presents unique challenges for unsupervised learning due to its complex, multivariate, nonlinear, and nonstationary nature.
• Conventional unsupervised time-series learning methods struggle with unlabeled EEG data, limiting its application in areas like neurocognitive disorder diagnosis and brain-machine interfaces.

Purpose of the Study:

• To address the challenge of clustering unlabeled EEG time-series data.
• To propose a novel algorithm, mwcEEGc, for improved EEG data clustering.
• To evaluate the effectiveness and superiority of mwcEEGc compared to existing unsupervised learning methods.

Main Methods:

• Developed mwcEEGc, an EEG clustering algorithm that maps the problem to maximum-weight clique (MWC) searching.
• Constructed an improved Fréchet similarity-weighted EEG graph, considering both vertex and edge weights.
• Clustered EEG data based on similarity weights rather than cluster centroids.

Main Results:

• mwcEEGc achieves high-quality clusters with excellent intracluster compactness and intercluster scatter.
• Demonstrated the superiority of mwcEEGc over ten state-of-the-art unsupervised learning/clustering approaches on 14 real-world EEG datasets.
• Validated mwcEEGc's performance using standard clustering validity criteria.

Conclusions:

• mwcEEGc is the first algorithm to utilize MWC searching for clustering unlabeled EEG trials.
• The algorithm satisfies theoretical clustering properties, including richness, consistency, and order independence.
• mwcEEGc offers a significant advancement in unsupervised learning for complex EEG data analysis.