Score-Based Data Generation for EEG Spatial Covariance Matrices: Towards Boosting BCI Performance.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) classifier efficacy is data-dependent.
• Geometric deep learning utilizes spatial covariance matrices from EEG as input.
• Generating synthetic EEG-derived matrices can improve classifier performance.

Purpose of the Study:

• To propose a generative modeling technique for synthesizing spatial covariance matrices from EEG data.
• To enhance geometric deep learning classifiers for motor imagery tasks.
• To evaluate the quality and utility of generated spatial covariance matrices.

Main Methods:

• Utilized state-of-the-art score-based generative models.
• Generated spatial covariance matrices from a left/right-hand-movement motor imagery EEG dataset.
• Assessed generated samples using visual and quantitative methods, including classifier prediction and neurophysiological alignment.

Main Results:

• Generated samples exhibited exceptional pixel-level resolution.
• The Fréchet mean of generated samples aligned with known neurophysiological patterns (Mu and Beta bands at C3/C4).
• A pre-trained classifier predicted 84.3% of generated samples accurately, with up to 8.7% accuracy improvement in a holdout experiment.

Conclusions:

• Score-based generative modeling is a powerful technique for synthesizing EEG-derived spatial covariance matrices.
• The generated data supports the improvement of geometric deep learning classifiers for motor imagery.
• The findings demonstrate the potential for data augmentation in EEG-based machine learning.