The localization of spontaneous brain activity: an efficient way to analyze large data sets

An efficient solution is presented of the problem to localize the electric generators of spontaneous magnetoencephalography (MEG) and electroencephalography (EEG) data for large data sets. When a data set contains more than 100,000 samples standard methods fail or become impractical. The method presented here is useful, for example, for the localization of (pathological) brain rhythms or the analysis of single-trial data. The problem is defined as finding the good fitting dipoles using the single-dipole model applied on each time sample. First, the data is bandpass filtered to select the rhythm of interest. Next, the empirical relationship between data power and probability of a dipole with a high goodness of fit (g.o.f.) is used to preselect data points. Then a global search algorithm is applied, based on precomputed lead fields on a fixed grid, to obtain a good initial guess for the nonlinear dipole search. Finally, the dipole search is applied on those samples that have a low initial guess error. In a group of five patients, it is found that 50% of the dipoles with a g.o.f. of at least 90% can be found by disregarding 90% of the data samples. Those dipoles can be found efficiently by disregarding all sample points with an initial guess relative residual error of 15% or lower. Finally, a simple empirical expression is found for the optimal mesh size of the global search grid. The method is completely automatic and makes it possible to study simple generators of large MEG and EEG data sets on a routine basis.