How Does Sensor-Space Group Blind Source Separation Face Inter-individual Neuroanatomical Variability? Insights from...

Because of volume conduction and inter-individual neuroanatomical variability, similar sources in different brains may lead to variable topographies. This represents a major limitation for sensor-space group level decomposition of electroencephalographic data, a technique which introduces potential biases when aggregating individual data. To which extent this impedes subsequent source separation and localization was quantified in the present study. To this end, several simulations using an atlas of human cerebral cortex that takes into account the variability of cortical morphology (Van Essen in NeuroImage 28:635-662, 2005) were performed. For each virtual subject (up to n = 160), the orientation and location of each single simulated dipole was randomly modified as a function of the variability of the cortical shape of a given point in the brain provided by the probabilistic atlas. The resulting activity was projected on the scalp, and topographical shifts were estimated. Then, different algorithms based on second order statistics (SOS) or higher order statistics were used to recover the simulated sources from sensor space information with group blind source separation (gBSS) procedures (based on UWSOBI or EFICA, respectively). As expected, the variability of orientation of the cortical surface across subjects was found to induce substantial variability in scalp potential maps, especially if the sources originate from the dorsolateral prefrontal cortex or the temporoparietal junction. These biases could be compensated for by increasing drastically the number of subjects included in the topographical analyses. By contrast, gBSS was found to be insensitive to inter-individual differences of neuroanatomy. Rather, the estimation of the spatial filters seems to be optimized for the population of interest. Thus, optimal performance of source separation and subsequent source localization did not require the inclusion of a large sample of subjects (n < 20), at least when applying SOS-based statistics that use source spectral diversity to identify and gather similar sources with variable location and orientation. The resulting conclusion that inter-individual neuroanatomical variability is not a major limitation to sensor-space gBSS methods provides boosting perspectives for this promising approach, especially for the detection and localization of task/population related neural sources.