Contrastive Principal Component Analysis in the Time-Frequency Domain for EEG Data Reduction

Electroencephalogram (EEG) is a non-invasive approach to measuring neural oscillations. As EEG signals are known to be non-stationary, one common approach to analyze them has been time-frequency analysis. Time-frequency distributions allow us to represent the time-varying energy content of the signals across different frequency bands. However, analyzing multichannel EEG data across multiple experimental conditions and subjects in the time-frequency domain results in highly redundant representations. In order to reduce the dimensionality and extract relevant features in the time-frequency domain, principal component analysis (PCA) has been widely utilized. Time-Frequency PCA (TF-PCA) provides a data-driven way to reduce the time-frequency surfaces into a few meaningful components. In many clinical studies, it is not only sufficient to reduce the EEG data into a few interpretable components, it is also important to discriminate between different experimental conditions or subject groups. In this paper, we extend TF-PCA by employing recently introduced contrastive PCA (cPCA) to discover low-dimensional structure that discriminates between two datasets. The proposed TF-cPCA approach is evaluated on an EEG study to discriminate between experimental conditions and brain regions from event-related potentials recorded during a cognitive control task.