Tangent space spatial filters for interpretable and efficient Riemannian classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Riemannian geometry methods are effective for brain-computer interface (BCI) decoding.
• These methods face challenges with high-dimensional data and interpretability, potentially leading to artifact-driven performance.
• This limits their application in online, high-density BCI systems and critical areas like neurofeedback.

Purpose of the Study:

• To develop dimension reduction solutions for Riemannian geometry-based BCI classification.
• To establish the equivalence between linear functions on the tangent space and spatial filters.
• To improve the efficiency and interpretability of BCI classification methods.

Main Methods:

• Proved the equivalence between linear functions on the tangent space and derived spatial filters.
• Proposed dimension reduction solutions for Riemannian methods without intensive optimization.
• Validated pipelines against common spatial patterns and tangent space classification using a large, open-access BCI analysis framework (7 datasets, 200+ subjects).
• Verified framework robustness through visualization of spatial patterns.

Main Results:

• Proposed spatial filtering methods achieve competitive or superior performance compared to classic tangent space classification.
• Achieved up to 97% reduction in testing time cost.
• Demonstrated consistent performance with only 4-6 filter components, irrespective of channel count.
• Visualized spatial patterns confirmed underlying neuronal sources.

Conclusions:

• The proposed methods enhance theoretical understanding of Riemannian geometry in BCI classification.
• Enables more efficient classification and effective removal of artifact sources.
• Facilitates broader application of BCI technology, especially in sensitive patient populations and neurofeedback systems.