Introducing block-Toeplitz covariance matrices to remaster linear discriminant analysis for event-related potential...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Covariance matrices from electroencephalogram (EEG) time series are crucial for brain signal decoding using machine learning.
• Estimating these matrices is challenging due to small datasets and high dimensionality in EEG data.
• Current state-of-the-art for event-related potential (ERP) based brain-computer interfaces (BCI) uses shrinkage regularization for covariance estimation in linear discriminant analysis (LDA).

Purpose of the Study:

• To improve linear discriminant analysis (LDA) for electroencephalogram (EEG) based brain-computer interfaces (BCI) by developing a more effective covariance estimation method.
• To leverage domain-specific characteristics of EEG data to enhance covariance regularization beyond general approaches.
• To introduce a block-Toeplitz structure for covariance matrices in LDA to exploit signal stationarity assumptions.

Main Methods:

• Proposed a novel 'ToeplitzLDA' method enforcing a block-Toeplitz structure on the covariance matrix for LDA.
• Assumed signal stationarity within short time windows for each EEG channel to justify the block-Toeplitz structure.
• Conducted offline re-analysis of data from 213 subjects across 13 different ERP BCI protocols.

Main Results:

• ToeplitzLDA significantly increased binary classification performance compared to shrinkage regularized LDA (up to 6 AUC points) and Riemannian classification (up to 2 AUC points).
• In an unsupervised visual speller task, this method led to an average 81% relative reduction in spelling errors for 25 subjects.
• ToeplitzLDA demonstrated lower memory and reduced time complexity for LDA training and robustness against increased temporal features.

Conclusions:

• The proposed block-Toeplitz covariance estimation method enhances classification rates in ERP-based BCI protocols.
• This approach can reduce calibration times for BCI applications, improving overall usability.
• ToeplitzLDA's reduced computational and memory requirements make it particularly suitable for mobile BCI systems.