2.6. fMRI analyses

fMRI data were analyzed using a conventional general linear model (GLM) approach in SPM12 (http://www.fil.ion.ucl.ac.uk/spm/software/spm12/) running in Matlab R2014b (MathWorks Inc., Natick, MA). Preprocessing entailed realignment, coregistration to individual subject's T1‐weighted images, spatial normalization to template space (MNI 152), and smoothing. Interscan movement was quantified during realignment; motion parameters (six per acquisition) were entered into first‐level analyses as nuisance regressors. Contrast maps were generated (stories minus noise) for each subject, and passed on for subsequent group comparisons (second‐level analysis). At the second level, we compared the spatial distribution of language maps for children born preterm and term controls.

To objectively identify the language network in our study cohort, we computed the joint activation map for children born preterm and typically developing controls. The resulting activation map was sectioned using a 200‐unit random parcellation scheme (Craddock, James, Holtzheimer, Hu, & Mayberg, 2012). Centroids of parcels with significant activation on fMRI served as network nodes for subsequent MEG virtual sensor extraction and connectivity analyses (Figure 1B) (Kadis et al., 2016).

Functional MRI activation maps and network nodes. (A) Eight axial slices showing typical bilateral temporoparietal activation for stories versus noise (p < .001, n = 30, 15 TC and 15 EPT) and 3D surface rendering showing activation from lateral perspective. (B) The activation map is parcellated to produce the “nodes” for network analysis