Data analyses¶
Bezmaternykh et al. (2021) discarded the first five volumes of each series to ensure the steady state. The preprocessing of the images was performed with the Matlab and SPM12 software. The batch included motion correction, slice timing, normalization of the images to MNI space and smoothing with Gaussian kernel of 8 mm. They used the default setting. A Shift to 2 mm or rotation to 2 degrees were considered as excessive head movements. Two patients were excluded. One because of head movement and another one due to prominent artifacts. The GIFT 3.0.a software was used to perform the spatial independent components analysis (ICA). The ICA was performed using Infomax algorithm with the option to reduce the stochasticity namely ICASSO and intensity normalization. The authors reconstructed the individual dynamics from the group data with the GICA, procedure of reverse reconstruction, for each participant. The extracted components in spatial domain were described by z-scores of weight coefficients, which indicated the degree of presence of the component time course in a particular voxel. The average group activation maps for each component and the coefficients of their spatial correlation with gray, white matter and cerebrospinal fluid masks were constructed. Components correlating with mask of either white matter or cerebrospinal fluid more than with one of grey matter were considered as artifacts and excluded from analysis. After that, the correlations with masks of classical resting state networks were calculated for the remaining components. The primary set of components’ maps we used was FMRIB/RIC one https://www.fmrib.ox.ac.uk/datasets/brainmap+rsns/, in cases no strong match to FMRIB/RIC set was found Stanford maps were also tried http://findlab .stanford.edu/functional_ROIs.html. The composition of the components was determined at the threshold t = 2. Afterwards, the FNC toolbox (http://mialab.mrn.org/software/fnc) was used to calculate temporal correlations between the dynamics of the selected components. With the Lag-Shift algorithm, the coefficients and the lag times for each pair of networks were computed. The time shift was selected to maximize an absolute value of correlation coefficient. Intergroup differences were estimated with the Student’s t test for independent samples, the dynamics of the networks intercorrelations from the first to the second recording with the t test for paired samples. In each participant, the Pearson correlation coefficient between time series of the components and 6 rigid body head motion parameters were calculated. Additional tests were performed excluding all pairs involving components with correlation coefficient with any motion parameter above margin in certain participant. R > 0:5, r > 0:4, and all p < 0:05 were tested, and r > 0:4 was empirically chosen as a tradeoff between reducing maximal allowed correlation and preserving the majority of data units. Thus, t test was repeated with exclusion of components correlating with motions to degree of 0.4 or higher, so results significant at first test with all data included and nonsignificant at second test with some data excluded were considered as motion related false positive (Bezmaternykh et al. 2021). With this information we recreated the analysis with slightly different tools.
In the following scripts you can find our preprocessing and ICA analysis with creating the correlation matrices and connectomes.