Changes in task-related functional connectivity across multiple spatial scales are related to reading performance

Abstract

Reading requires the interaction of a distributed set of cortical areas whose distinct patterns give rise to a wide range of individual skill. However, the nature of these neural interactions and their relation to reading performance are still poorly understood. Functional connectivity analyses of fMRI data can be used to characterize the nature of interactivity of distributed brain networks, yet most previous studies have focused on connectivity during task-free (i.e., "resting state") conditions. Here, we report new methods for assessing task-related functional connectivity using data-driven graph theoretical methods and describe how large-scale patterns of connectivity relate to individual variability in reading performance among children. We found that connectivity patterns of subjects performing a reading task could be decomposed hierarchically into multiple sub-networks, and we observed stronger long-range interaction between sub-networks in subjects with higher task accuracy. Additionally, we found a network of hub regions known to be critical to reading that displays increased short-range synchronization in higher accuracy subjects. These individual differences in task-related functional connectivity reveal that increased interaction between distant regions, coupled with selective local integration within key regions, is associated with better reading performance. Importantly, we show that task-related neuroimaging data contains far more information than usually extracted via standard univariate analyses--information that can meaningfully relate neural connectivity patterns to cognition and task.

Figure 1. Time course of each trial and experimental design.

Thirty-nine children were presented with 48 lexical trials interspersed pseudo-randomly with 36 fixation and 24 perceptual control trials in an event-related design. Each trial lasted 4200 ms, with 200 ms breaks in between, for a total of 4.4 sec each and 480 sec for the entire run. During lexical trials, words were presented sequentially for 800 ms, after which a red fixation cross appeared. The subject then had 2200 ms to provide a rhyming judgment response. During fixation trials, a black cross was displayed for 1800 ms and then was replaced by a red cross. The subject subsequently had 2200 ms to acknowledge the red cross with a button press.

Figure 2. Determination of task-responsive network and hierarchical partitioning to form three levels of functional hierarchy.

(A) We determine group-level GLM random effects contrasts of lexical minus fixation trials (task-positive) and fixation minus lexical trials (task-negative) to define two functional systems encompassing task-responsive brain regions. Voxels within these regions are coarse-grained to 6×6×6 mm³ ROIs to form the highest-resolution units of our network, called nodes. Time series of hemodynamic response are extracted from these nodes and are pairwise cross-correlated to define a weighted connectivity matrix for all subjects. (B) Schematic illustration and dendrogram of hierarchical partitioning. All nodes are pooled together and iteratively partitioned using modularity-based clustering algorithms to form a hierarchical network of relations. Each group on a higher level is partitioned to yield subgroups on the next lower level. Three levels of partitioning are defined, with various numbers of groups on each level: 2 “systems,” 9 “blocks,” and 33 “clusters.” (C) Matrix of Fisher’s Z-transformed correlation coefficients averaged over all subjects. Nodes are ordered to keep clusters, blocks, and systems contiguous. Colored bars and dendrogram along the sides indicate group membership of nodes at all three levels and correspond to those from (B). (D) Categorization of link type depends on the lowest level at which the two associated nodes are classified in the same group; i.e. same cluster (“cluster links,” brown), same block but different clusters (“block links,” tan), same system but different blocks (“system links,” beige), or between systems (“cross links,” white). Colored bars illustrate the partition of nodes into different groups at the three levels, similar to (B) and (C), but the specific groupings in (D) are conceptual only and do not represent real data.

Figure 3. Brain maps of hierarchical partitioning.

(A) Task-positive and task-negative regions defined through univariate GLM methods have high correspondence with (B) the 2 systems found through modularity-based partitioning, with the exception of a few subcortical,fronto-parietal, and temporal areas (highlighted by yellow circles). Brain maps of (C) the 9 blocks and (D) the 33 clusters show that functional groups tend to be spatially localized and confined to only a few anatomical regions.

Figure 4. Functional groups have high concurrency with anatomical regions.

Colors correspond to number of nodes belonging to the group (system, block, or cluster) as well as the anatomical region (AAL template).

Figure 5. Links spanning higher levels of the hierarchy are more correlated with task accuracy.

Plots of average link weight (Fisher’s Z transformed) versus task accuracy for (A) cluster links, (B) block links, (C) system links, and (D) cross links. Plots of the ratio of the average weight of (E) system links to cluster links, and (F) block links to cluster links.

Figure 6. Performance modulation network quantifies modulation of correlations by task accuracy and reveals key hub regions.

(A) We investigated the hub-like nature of the 33 clusters by averaging together link weights of all links within clusters as well as links between pairs of clusters to create cluster-level connectivity matrices for all subjects. Performance modulation for every link was then found by calculating the correlation of link weight with task accuracy across participants to create the performance modulation matrix. (B) We classify as hubs the clusters with the highest eigenvector centrality values calculated on the performance modulation matrix from (A). The hubs found (large spheres) encompass key reading areas such as fusiform gyrus and left temporal gyrus, as well as dorsal striatum, thalamus, and left hippocampus. Views depicted are lateral (top), posterior (center), and medial (bottom).