Source Reconstruction Quality

Following the rationale of previous studies (Liu et al. 2017, 2018; Abreu et al. 2020b), a spatial ICA step similar to that applied to the fMRI data for identifying RSNs was then performed on the reconstructed source dynamics S, with the purpose of separating those potentially associated with RSNs and/or other regions of interest in our tasks. This can be formulated as:

where U∈RT×I is the mixing matrix, with each column ui∈RT×1 the time-course of the source component (SC) i; and SIC∈RI×D represents the spatial maps in the source space associated to each of the I SCs. Because the EEG data is submitted to a temporal reduction step prior to solving the inverse problem in order to reduce noise while guaranteeing a temporally continuous estimation of sources (López et al. 2014), the rank of S is reduced accordingly, being then defined an upper bound on the number of SCs to be estimated. Such maximum allowed number of SCs was then estimated, which was between 50 and 60 in all cases. Finally, the SCs were converted into z-scores, and the deformation field estimated while solving the forward problem was applied to transform them from the source space into the MNI space.

Model-based quality metrics were first considered, namely the variance explained (VE) of the reconstructed EEG data Y~=L∙S relative to the actual EEG data Y (see Eq. 3), and the expectation of the posterior probability P(model | data) of the inversion models. The latter is obtained from the associated log-evidence values of all models, for all subjects, as described in (Rigoux et al. 2014), and reflects the probability of obtaining a given model when randomly selecting a subject. These probability values were normalized to sum to one, over the models under analysis. Other quality metrics reflecting more directly the presence of neuronal activity of interest in the SCs were also considered, as described next.

First, because the perception of motion in general, and of biological motion in particular, is known to elicit certain brain regions (Chang et al. 2018), the following four spherical regions of interest (ROIs) of 10 mm centered at specific MNI coordinates (indicated in square brackets) were considered: anterior insula (aINS) at [± 36, 24, 2], extrastriate body area (EBA) at [left –46, –75, –4; right 47, –71, –4], fusiform body area (FBA) at [left –38, –38, –27; right 43, –43, –28], and fusiform gyrus (FFG) at [± 42, –56, –14]. Four additional task-related brain regions (Chang et al. 2018) were obtained from FSL atlases (threshold applied to the probability maps is indicated in square brackets), namely: inferior frontal gyrus (IFG) [0.25], posterior superior temporal sulcus (pSTS) [0.25], visual area V3 [0.25] and visual area hMT + /V5 [0.10]. After binarizing the ROIs and the SC maps, the Dice coefficient d, and the proportion of the ROIs contained in the SC maps p_RS, were then quantified according to (Dice 1945):

where N_ROI and N_SC denote the number of non-zero voxels in the ROIs and SC maps, respectively, and N_ov the number of overlapping non-zero voxels between the two images; both measures range from 0 (no overlap) to 1. These same two measures, d and p_RS, were also computed between the SC maps and 10 RSN templates described in (Smith et al. 2009), in order to assess which, if any, SCs represented RSNs (similar to the identification of RSNs on fMRI data described previously).

All these measures were computed for each subject, run, inversion algorithm, set of covariance components (CCs), SC maps and maps of interest (8 ROIs and 10 RSN templates). Because only a subset of the SC maps is expected to be associated with those maps of interest, the SC map yielding the highest dice coefficient for each map was identified, and the associated d∗ and pRS∗ maximum values kept for subsequent analyses. The d∗ and pRS∗ values were further summarized by computing their average within each map type (ROIs and RSN templates), thus yielding the final set of 13 [subjects] × 4 [runs] × 4 [inversions] × 3 [sets of CCs] × 2 [map types] = 1440 values of d∗ and pRS∗.

The main effects of the population group (MS patients and healthy subjects), inversion algorithm, the set of CCs and the type of map of interest, as well as interaction effects, were evaluated by means of a 4-way repeated measures Analysis of Variance (ANOVA) for the VE, d∗ and pRS∗ measures treated separately as the dependent variables. Multiple comparisons between the inversion algorithms, sets of CCs and interactions between the two were performed by means of a post-hoc statistical test with the Tukey–Kramer correction. A level of statistical significance p < 0.05 was considered.