Spatio-temporal matching pursuit (SToMP) for multiple source estimation of evoked potentials

Decomposing the multichannel recorded signals into their generators' temporal activity patterns is an important step towards a feasible solution of the bioelectric inverse problem. Matching Pursuit with time-frequency dictionaries is a well known method for signal decomposition and feature extraction. In the current work this method is generalized into Spatio-Temporal Matching Pursuit (SToMP) and adapted for multiple source estimation of bioelectrical activity. In the first stage of the presented algorithm, the multichannel signals are decomposed into the best-matched spatio-temporal waveforms selected from a physiologically motivated time-frequency dictionary. This spatio-temporal decomposition enables fully linear exhaustive search for the optimal sources of each waveform in the second stage of the algorithm, avoiding non-linear optimization. The linear exhaustive search is constrained to a three-dimensional non-uniform grid (or voxels) of all the anatomical candidates for sources. The SToMP algorithm for multiple source localization was evaluated by simulation. It exhibits better results than other spatio-temporal multiple source localization methods, that are based on eigenvector decomposition, like MUSIC. Real data results of Visual Evoked Potentials source localization, with MRI data constrains and visualization, demonstrates physiological feasible solution of the bioelectric inverse problem. The SToMP decomposition algorithm is robust, and can be also used for spatio-temporal inverse filtering, or for any other sensor-array inverse problems (like ECG source estimation or radar direction estimation).