Topography-time-frequency atomic decomposition for event-related M/EEG signals

We present a method for decomposing MEG or EEG data (channel x time x trials) into a set of atoms with fixed spatial and time-frequency signatures. The spatial part (i.e., topography) is obtained by independent component analysis (ICA). We propose a frequency prewhitening procedure as a pre-processing step before ICA, which gives access to high frequency activity. The time-frequency part is obtained with a novel iterative procedure, which is an extension of the matching pursuit procedure. The method is evaluated on a simulated dataset presenting both low-frequency evoked potentials and high-frequency oscillatory activity. We show that the method is able to recover well both low-frequency and high-frequency simulated activities. There was however cross-talk across some recovered components due to the correlation introduced in the simulation.