Metastable neural dynamics underlies cognitive performance across multiple behavioural paradigms

Abstract

Despite resting state networks being associated with a variety of cognitive abilities, it remains unclear how these local areas act in concert to express particular cognitive operations. Theoretical and empirical accounts indicate that large-scale resting state networks reconcile dual tendencies towards integration and segregation by operating in a metastable regime of their coordination dynamics. Metastability may confer important behavioural qualities by binding distributed local areas into large-scale neurocognitive networks. We tested this hypothesis by analysing fMRI data in a large cohort of healthy individuals (N = 566) and comparing the metastability of the brain's large-scale resting network architecture at rest and during the performance of several tasks. Metastability was estimated using a well-defined collective variable capturing the level of 'phase-locking' between large-scale networks over time. Task-based reasoning was principally characterised by high metastability in cognitive control networks and low metastability in sensory processing areas. Although metastability between resting state networks increased during task performance, cognitive ability was more closely linked to spontaneous activity. High metastability in the intrinsic connectivity of cognitive control networks was linked to novel problem solving or fluid intelligence, but was less important in tasks relying on previous experience or crystallised intelligence. Crucially, subjects with resting architectures similar or 'pre-configured' to a task-general arrangement demonstrated superior cognitive performance. Taken together, our findings support a key linkage between the spontaneous metastability of large-scale networks in the cerebral cortex and cognition.

Keywords: cognition; coordination dynamics; deep learning; fMRI; metastability; metastable neural dynamics; neurocognitive networks; resting state.

FIGURE 1

Schematic overview of the modified convolutional neural network architecture used to classify the eight different network configurations (seven tasks and one resting state; Kawahara et al., 2017). Each block represents the input/output of a numbered filter layer. The third dimension (m) represents the result of convolving the input with m different filters (feature maps). First, an interaction matrix composed of the interactions of 14 networks (based on synchrony or metastability) is entered as input. This is convolved with an edge‐to‐edge (1, E2E) filter which weights the edges associated with adjacent brain networks in topological space. The output from this layer is then convolved with an edge‐to‐node (2, E2N) filter which assigns each network a weighted sum of its edges. Next, a node‐to‐graph (3, N2G) layer outputs a single response based on all the weighted nodes. Finally, the number of features is reduced to eight output classifications through a series of fully connected (4/5/6, FC) layers

FIGURE 2

Empirical global metastability of fMRI BOLD signal in the resting state (in green) and during several cognitively demanding tasks (in grey). Bars display mean, 95% CI and one SD with individual subjects indicated. Tasks arranged in ascending order of mean metastability. One‐way ANOVA revealed significantly higher metastability during task execution relative to resting state (p < .01)

FIGURE 3

Statistically significant (p < .01; corrected) increases in BOLD signal metastability between empirical resting state networks during task as compared to resting state. Circular graphs show largest connected sub‐graph of increased metastability identified by the network‐based statistic at a fixed threshold (16). Nodes are scaled to reflect the relative importance of their interactions (the sum of their effect sizes). Overall, the connectivity of the dorsal attention (green) and fronto‐parietal networks (yellow) are the most metastable

FIGURE 4

Increases in metastability (blue) are associated with a greater number of network connections than equivalent increases in synchrony (red). Figure shows size of sub‐graph identified by the network‐based statistic (see Figure 3) rank ordered by metastability

FIGURE 5

Convolutional neural network performance in terms of classification accuracy where each row represents the true class and each column represents the classification made by the neural network. Diagonal elements report the percentage of instances correctly classified. Off‐diagonal elements report the percentage of instances that are incorrectly classified. Inputs are classified as belonging to one of eight different network states (seven tasks plus one resting state condition) where each row/column corresponds to the interaction of one network with 13 others (in terms of either synchrony or metastability). (a) Classification accuracy in terms of the synchrony between networks (average accuracy = 76%; chance level 12.5%). (b) Classification accuracy in terms of the metastability between networks (average accuracy = 46%). (c) Classification accuracy in terms of occluded network synchrony (average accuracy = 2%). Here, classification accuracy was reduced by masking out (setting to zero) a small subset of network interactions (see Figure 6)

FIGURE 6

Each network state (one rest and seven tasks) is defined by a small number of task‐evoked changes in synchrony between resting state networks. Here, network connectivity important for correct classification in more than 90% of individuals (as determined by guided backpropagation) is masked out (black). Such 'occluded' inputs are associated with exceptionally poor classification accuracy (Figure 5c). The width of each column/row is scaled to reflect the relative number of regions in each network

FIGURE 7

Principal component analysis reveals a task‐general network architecture. Task‐based reasoning is principally characterised by high metastability in regions associated with cognitive control (top; red) and low metastability in regions associated with sensory processing (blue; middle). On average the first principal component accounts for 78% of the variance. Loadings are distributed equally between the seven tasks. Regions are colour coded by the sum of their ingoing/outgoing connectivity. Nodes are colour coded according to the Gordon atlas (bottom). Node diameter is proportional to the sum of ingoing/outgoing connectivity. Recurrent connections correspond to activity within a network

FIGURE 8

Resting state metastability of cognitive control networks is predictive of task performance. Each edge represents a statistically significant positive correlation between the intrinsic metastability of a connection and cognition/behaviour (p < .01; FDR corrected) shaded to reflect standardised effect sizes (Pearson's r or correlation coefficients). Nodal diameter is scaled to reflect the sum of their ingoing/outgoing connectivity. Cognitive measures were obtained outside the scanner (top); measures of behavioural accuracy were acquired inside the scanner (bottom). Metastability in the connectivity of cognitive control networks is linked to task performance including the fronto‐parietal (adaptive task control), cingulo‐opercular (sustained tonic attention) and dorsal attention networks (attending to visuospatial stimuli). Note the markedly different profiles presented by fluid and crystallised intelligence

FIGURE 9

The efficiency of the transformation between resting and task‐based network architecture is conditioned on high metastability in the couplings of cognitive control networks and low metastability in the couplings of sensory networks. (a) Slope (coefficients) of the regression line (negative or positive) relating metastability of network connectivity to update efficiency. (b) Statistically significant correlations between metastability of network connectivity and update efficiency (p < .01; FDR corrected). (c) Statistically significant correlations between metastability and update efficiency in low (blue; left) and high (red; right) metastability subnetworks (p < .01; FDR corrected) where low metastability is associated with unimodal sensory networks (auditory, motor and visual) and high metastability is related to cognitive control (dorsal attention, fronto‐parietal, cingulo‐opercular and default mode networks). Some networks such as the salience, medial parietal, parieto‐occipital and thalamus were sites of convergence for both high and low metastability connections (yellow; centre). Nodal diameter is scaled to reflect the sum of their ingoing/outgoing edges