Virtual lesions in MEG reveal increasing vulnerability of the language network from early childhood through adolescence

Abstract

In childhood, language outcomes following brain injury are inversely related to age. Neuroimaging findings suggest that extensive representation and/or topological redundancy may confer the pediatric advantage. Here, we assess whole brain and language network resilience using in silico attacks, for 85 children participating in a magnetoencephalography (MEG) study. Nodes are targeted based on eigenvector centrality, betweenness centrality, or at random. The size of each connected component is assessed after iterated node removal; the percolation point, or moment of dis-integration, is defined as the first instance where the second largest component peaks in size. To overcome known effects of fixed thresholding on subsequent graph and resilience analyses, we study percolation across all possible network densities, within a Functional Data Analysis (FDA) framework. We observe age-related increases in vulnerability for random and betweenness centrality-based attacks for whole-brain and stories networks (adjusted-p < 0.05). Here we show that changes in topology underlie increasing language network vulnerability in development.

Fig. 1. Results for the Random attack strategy on the whole-brain parcellation.

Data are presented as the beta estimate across densities +/− 95% CIs (shaded). After testing for significance of the overall model (see, Supplementary Table 1), analyses showed Age (negative) as the only significant regressor at all densities below 15% (F = 31.44, p < 0.0001). Sex and Handedness did not meet the p-value threshold (0.01) and 95% CIs contained 0 throughout the whole density range (Supplementary Fig. 3, panels c and d). Source data are provided as a Source Data file.

Fig. 2. Results for betweenness centrality-based attacks on the whole-brain parcellation.

Data are presented as the beta estimate across densities +/− 95% CIs (shaded). After testing for significance of the overall model (see, Supplementary Table 1), analyses showed Age (negative) as the only significant regressor at all densities (F = 126.38, p < 0.0001). Sex and Handedness did not meet the p-value threshold (0.01) and 95% CIs contained 0 throughout the whole density range (Supplementary Fig. 4, panels c and d). Source data are provided as a Source Data file.

Fig. 3. Distribution of nodes removed prior to failure in whole-brain attacks.

Distribution of nodes removed in the whole-brain for betweenness centrality-based attacks for the 1^st (panel a) and 4^th (panel b) quartiles of age. Regions are colored by the percentage of participants for which the node was removed prior to network failure (dark to light). These results were displayed for 5% density across participants. Results show more consistently removed regions, i.e., group level hubs, in the younger quartile (brighter regions in panel a) compared to older participants (panel b). Also, distribution of critical hubs becomes much more focal in the older group as evidenced by greater heterogeneity among neighboring regions. To show consistency across densities, results are also plotted at 2.5% (Supplementary Fig. 9, left panel) and 7.5% (Supplementary Fig. 10, left panel) density.

Fig. 4. Results for random attacks on the stories network parcellation.

Data are presented as the beta estimate across densities +/− 95% CIs (shaded). After testing for significance of the overall model (see, Supplementary Table 1), analyses showed Age (negative) as the only significant regressor at densities between 1 and 15% (F = 44.51, p < 0.0001). Sex and Handedness did not meet the p-value threshold (0.01) and 95% CIs contained 0 throughout the whole density range (Supplementary Fig. 6, panels c and d). Source data are provided as a Source Data file.

Fig. 5. Results for betweenness centrality-based attacks on the stories network parcellation.

Data are presented as the beta estimate across densities +/− 95% CIs (shaded). After testing for significance of the overall model (see, Supplementary Table 1), analyses showed Age (negative) as the only significant regressor at all densities (F = 86.25, p < 0.0001). Sex and Handedness 95% CIs contained 0 throughout the whole density range (Supplementary Fig. 7, panels c and d). Source data are provided as a Source Data file.

Fig. 6. Distribution of nodes removed prior to failure in stories network attacks.

Distribution of nodes removed in the stories network for betweenness centrality-based attacks for the 1st (a) and 4th (b) quartiles of age. Regions are colored by the percentage of participants for which the node was removed prior to network failure (dark to light). These results were displayed for 5% density across participants. Like whole-brain analyses, we found that the younger children had more regions that were removed consistently, i.e., group-level hubs, across subjects (panel a). To show consistency across densities, results are also plotted at 2.5% (Supplementary Fig. 9, panels c and d) and 7.5% (Supplementary Fig. 10, panels c and d) density.

Fig. 7. Distribution of language network nodes.

Nodes were derived from NeuroSynth by using a uniformity test (FDRq < 0.01) on all resulting activation maps with the search term “language”. Panel a represents the volumetric regions contained in the parcellation. Panel b is a 3D-rendered representation showing the centroid of each node.