Reproducible Neuronal Components found using Group Independent Component Analysis in Resting State Electroencephalographic Data

Abstract

AbstractObjectiveEvaluate the reliability of neural components obtained from the appli-cation of the group ICA (gICA) methodology to resting-state EEG datasets acquired from multiple sites.MethodsFive databases from three sites, covering a total of 292 healthy subjects, were analyzed. Each dataset was segmented into groups of 15 subjects, for a total of 19 groups. Data were pre-processed using an automatic pipeline leveraging robust average referencing, wavelet-ICA and automatic rejection of epochs. On each group, stable gICA decompositions were calculated using the ICASSO methodology through a range of orders of decompositions. Each order was characterized by reliability and neuralness metrics, which were evaluated to select a single order of decomposition. Finally, using the decompositions of the selected order, a clustering analysis was performed to find the common components across the 19 groups. Each cluster was characterized by the mean scalp map, its dipole generator with its localization in Talairach coordinates, the spectral behavior of the associated time-series of the components, the assigned ICLabel class and metrics that reflected their reproducibility.ResultsLower order of decompositions benefits the gICA methodology. At this, using an order of ten, the number of reproducible components with high neuronal information tends to be around nine. Of these, the bilateral motor, frontal medial, and occipital neuronal components were the most reproducible across the different datasets, appearing in more than 89% of the 19 groups evaluated.ConclusionWe developed a workflow that allows finding reproducible spatial filters between different data sets. This contributes to the improvement of the spatial resolution of the EEG as a brain mapping technique.