Modeling Complex EEG Data Distribution on the Riemannian Manifold Toward Outlier Detection and Multimodal...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Riemannian geometry is increasingly used for brain-computer interfaces (BCIs).
• Existing methods often assume unimodal data distributions, which is limiting for electroencephalography (EEG) due to high data variability.
• Modeling complex, potentially multimodal, data distributions is crucial for improving BCI reliability.

Purpose of the Study:

• To propose a novel data modeling method for complex distributions on a Riemannian manifold of EEG covariance matrices.
• To enhance the reliability of brain-computer interfaces (BCIs) by addressing limitations of current machine learning techniques.
• To develop flexible methods for outlier detection and multimodal classification in EEG data.

Main Methods:

• Introduced Riemannian spectral clustering (RiSC) to represent EEG covariance matrix distributions on a manifold using a graph-based approach.
• Utilized geodesic distances for similarity measurement within the graph structure.
• Developed outlier detection (odenRiSC) and multimodal classification (mcRiSC) methods based on RiSC, with data-driven parameter selection.

Main Results:

• The proposed outlier detection method (odenRiSC) demonstrated superior accuracy in detecting EEG outliers compared to existing techniques.
• The multimodal classifier (mcRiSC) outperformed standard unimodal classifiers, particularly on datasets with high variability.
• RiSC-based methods effectively model both unimodal and multimodal distributions on the Riemannian manifold.

Conclusions:

• Riemannian spectral clustering (RiSC) provides a robust framework for EEG outlier detection and multimodal classification.
• The developed methods (odenRiSC and mcRiSC) are expected to enhance the robustness of real-world BCIs and neuroergonomics applications.
• These advancements facilitate the deployment of BCIs beyond laboratory settings.