MOVE

Localizing brain interactions from imaginary parts of cross-spectral matrices

Independent sources do not contribute systematically to imaginary parts of cross-spectra, which are therefore promising quantities to study brain interactions. While it is possible to decompose the imaginary cross-spectrum into subsystems of pairwise interactions using dynamical arguments only, it is not possible to decompose these subsystems further into sources without making additional spatial assumptions. We suggest to map channel data into source space using a linear source estimate and then to exploit the fact that different sources are spatially distinct. As a prerequisite we formulate a PCA in source space (sPCA) and show that spatial orthogonality is an adequate assumption for linear estimates of random current dipoles. Similar to ICA as typically used in the context of fMRI we additionally impose a spatial non-overlap criterion. When combining the two criteria it is possible to uniquely identify two interacting source distributions from interacting brain subsystems. This is demonstrated in simulations and for occipital alpha- and central mu-rhythm of real EEG data taken from a BCI experiment on imagined foot movement.