We classified performance-related mental states from EEG-derived measurements. We investigated the usefulness of massively distributed source reconstruction, comparing scalp and cortical scales. This approach provides a more detailed picture of the functional brain networks underlying the changes related to the mental state of interest. Local and distant synchrony measurements (coherence, phase locking value) were used for both scalp measurements and cortical current density sources, and were fed into a SVM-based classifier. We designed two simulations where classification scores increased when our 60-electrode scalp measurements were reconstructed on 60 sources and on a 500-source cortex. Source reconstruction appeared to be most useful in these simulations, in particular, when distant synchronies were involved and local synchronies did not prevail. Despite the simplicity of the model used, certain flaws in accuracy were observed in the localization of informative activities, due to the relationship between amplitude and phase for mixed signals. Our results with real EEG data suggested that the phenomenon of interest was characterized merely by modulations in local amplitudes, but also in strength of distant couplings. After source reconstruction, classification rates also increased for real EEG data when seeking distant phase-related couplings. When reconstructing a large number of sources, the regularization coefficient should be carefully selected on a subject-by-subject basis. We showed that training classifiers using such high-dimension data is useful for localizing discriminating patterns of activity.
Change Detection and Visualization of Functional Brain Networks using EEG Data
