Potential pitfalls of widely used implementations of common spatial patterns

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Common Spatial Patterns (CSP) is a widely used algorithm for feature extraction in brain-computer interfaces (BCIs).
• CSP relies on the assumption that signal covariance matrices have full rank, which may be violated by preprocessing steps like Independent Component Analysis (ICA) for artifact removal.

Purpose of the Study:

• To investigate the impact of rank-deficient covariance matrices on CSP implementations.
• To evaluate the performance of CSP in binary motor imagery classification tasks when rank reduction occurs.
• To review common open-source EEG analysis toolboxes for their handling of rank-deficient CSP.

Main Methods:

• Simulated rank reduction in EEG data.
• Applied CSP to binary motor imagery classification tasks using various open-source toolboxes (FieldTrip, BBCI, BioSig, EEGLAB, BCILAB, MNE).
• Analyzed classification accuracy and the nature of spatial filters generated by CSP.

Main Results:

• CSP implementations that do not account for rank reduction can significantly decrease classification accuracy, by up to 32%.
• Unprotected CSP can result in spatial filters with complex numbers, lacking clear neurophysiological interpretation.
• Variability in CSP implementation across different toolboxes was observed.

Conclusions:

• Researchers must be cautious about potential rank reduction in EEG signal preprocessing when using CSP.
• Ensuring CSP implementations correctly handle rank deficiency is crucial for reliable BCI performance.
• Verification of analysis pipelines and CSP implementations is recommended for accurate EEG signal processing.