Using Multiple Decomposition Methods and Cluster Analysis to Find and Categorize Typical Patterns of EEG Activity in Motor Imagery Brain–Computer Interface Experiments

In this study, the sources of EEG activity in motor imagery brain–computer interface (BCI) control experiments were investigated. Sixteen linear decomposition methods for EEG source separation were compared according to different criteria. The criteria were mutual information reduction between the source activities and physiological plausibility. The latter was tested by estimating the dipolarity of the source topographic maps, i.e., the accuracy of approximating the map by potential distribution from a single current dipole, as well as by the specificity of the source activity for different motor imagery tasks. The decomposition methods were also compared according to the number of shared components found. The results indicate that most of the dipolar components are found by the Independent Component Analysis Methods AMICA and PWCICA, which also provided the highest information reduction. These two methods also found the most task-specific EEG patterns of the blind source separation algorithms used. They are outperformed only by non-blind Common Spatial Pattern methods in terms of pattern specificity. The components found by all of the methods were clustered using the Attractor Neural Network with Increasing Activity. The results of the cluster analysis revealed the most frequent patterns of electrical activity occurring in the experiments. The patterns reflect blinking, eye movements, sensorimotor rhythm suppression during the motor imagery, and activations in the precuneus, supplementary motor area, and premotor areas of both hemispheres. Overall, multi-method decomposition with subsequent clustering and task-specificity estimation is a viable and informative procedure for processing the recordings of electrophysiological experiments.