Mapping Language Networks Using the Structural and Dynamic Brain Connectomes

Language-specific ROIs used in this study. AG, angular gyrus; MFG, middle frontal gyrus; IFGt, IFG pars triangularis; IFGo, IFG pars opercularis; STP, superior temporal pole; STG, superior temporal gyrus; STS, superior temporal sulcus; MTG, middle temporal gyrus.

The connectivity weight analyses focused on a subnetwork of the whole brain connectome composed of all 28 unique possible connections between the eight language-specific ROIs. Each connection was independently assessed and their anatomic representations (in a cohort of healthy individuals; see Materials and Methods for details) are demonstrated in this figure. All deterministic streamlines are represented in the first column (colored in accordance with their main direction of displacement, as per tractography convention: red, lateral to lateral; blue, rostral to caudal; green, anterior to posterior). Each specific pair-wise connection is represented by a different color in the second column. The third column demonstrates the centers of mass (centroids) of each pair-wise connection (colored per tractography convention), and the fourth column demonstrates each connection centroid colored similarly to the second column. Note the comprehensive and intricate pattern of structural connectivity assessed in the connectivity weight analyses.

This figure illustrates the procedural steps used to assess connectome dynamics. The upper left subplot demonstrates a rendering of the cortical surface with the large cortical and subcortical poststroke lesion overlaid in blue. Next (second subplot on the first row), the superior temporal gyrus is seeded. Thereafter, based on the weight of structural connectivity between the superior temporal gyrus and the remaining whole brain connectome, the number of steps taken to reach each other language-specific ROI is calculated. The sequence of subplots demonstrated which ROIs are reached, in sequence. This process results in a vector denoting the inverse of the number of steps taken to reach each other ROI, when one ROI is seeded (illustrated in the bottom right, with the color bar illustrating the inverse of the number of steps). For each participant, this is repeated by seeding each ROI in turn, and 28 unique pair-wise connectome dynamics are calculated.

This diagram explains how the statistical analyses and out of sample SVR predictions were performed for models using individual measures or their combinations.

Statistical analyses comparing the distribution of Pearson correlation coefficients between real and predicted dependent measures for each model. Each subplot demonstrates the distribution of the subtraction of the correlation coefficient from one method minus another. This is possible since the training and testing split samples were identical for each model at every iteration, providing a direct comparison between models. If 95% of the subtractions felt above 0, the first test in the subtraction was considered statistically superior than the other at p < 0.05.

The individual coefficients are shown in color to illustrate which ROIs were more influential on the ROI model (left column) and which connections were more influential on the connectivity weights model (right most columns) for predicting WAB-AQ (first row) and WAB fluency (second row). The color bars indicate SVR coefficients.

This graph illustrates the OLS weights assigned to each measure when models using combined measures were constructed (x-axis). For each possible combination, the correlation coefficients obtained with out of sample testing (i.e., when applied to the test data) are listed and the OLS weights are stacked to demonstrate their relative values.

This figure illustrates the statistical comparisons between the model using all three measures (Ens, ensemble) versus each individual measure. Similar to Figure 5, the histograms demonstrate the distribution of the subtractions of the correlation coefficients from one method minus another for every possible run. If 95% of the subtractions fell above 0, the first test in the subtraction was considered statistically superior than the other at p < 0.05.