Graph-based Dimensionality Reduction of EEG Signals via Functional Clustering and Total Variation Measure for BCI Systems

In this paper, we propose a novel and intuitively pleasing graph-based spatio-temporal feature extraction framework for classification of motor imagery tasks from elec- troencephalography (EEG) signals for brain-computer interface systems (BCIs). In particular, to account for the observation that measurements obtained from the EEG channels form a non-uniformly distributed sensor field, a representation graph is constructed using geographical distances between sensors to form connectivity neighborhoods. By capitalizing on the fact that functionality of different connectivity neighborhoods varies based on the intensity of the performed activity and concentration level of the subject, we formed an initial func- tional clustering of EEG electrodes by designing a separate adjacency matrix for each identified functional cluster. Using a collapsing methodology based on total variation measures on graphs, the overall model will eventually be reduced (collapsed) into two functional clusters. The proposed framework offers two main superiorities over its state-of-the-art counterparts: (i) First, the resulting dimensionality reduction is subject-adaptive and respects the brain plasticity of subjects, and; (ii) Second, the proposed methodology identifies active regions of the brain during the motor imagery task, which can be used to re-align EEG electrodes to improve accuracy during consecutive data collection sessions. The experimental results based on Dataset IVa from BCI Competition III show that the proposed method can provide higher classification accuracy as compared to the other existing methods.