Persistency and flexibility of complex brain networks underlie dual-task interference

Abstract

Previous studies on multitasking suggest that performance decline during concurrent task processing arises from interfering brain modules. Here, we used graph-theoretical network analysis to define functional brain modules and relate the modular organization of complex brain networks to behavioral dual-task costs. Based on resting-state and task fMRI we explored two organizational aspects potentially associated with behavioral interference when human subjects performed a visuospatial and speech task simultaneously: the topological overlap between persistent single-task modules, and the flexibility of single-task modules in adaptation to the dual-task condition. Participants showed a significant decline in visuospatial accuracy in the dual-task compared with single visuospatial task. Global analysis of topological similarity between modules revealed that the overlap between single-task modules significantly correlated with the decline in visuospatial accuracy. Subjects with larger overlap between single-task modules showed higher behavioral interference. Furthermore, lower flexible reconfiguration of single-task modules in adaptation to the dual-task condition significantly correlated with larger decline in visuospatial accuracy. Subjects with lower modular flexibility showed higher behavioral interference. At the regional level, higher overlap between single-task modules and less modular flexibility in the somatomotor cortex positively correlated with the decline in visuospatial accuracy. Additionally, higher modular flexibility in cingulate and frontal control areas and lower flexibility in right-lateralized nodes comprising the middle occipital and superior temporal gyri supported dual-tasking. Our results suggest that persistency and flexibility of brain modules are important determinants of dual-task costs. We conclude that efficient dual-tasking benefits from a specific balance between flexibility and rigidity of functional brain modules.

Keywords: complex networks; fMRI; flexibility; interference; modularity; multitasking.

Figure 1

Illustration of the paradigm, behavioral results and methodological framework of the study. (A) Experimental Design: Each subject was measured in four different conditions: (1) Resting‐state scan: Subjects were instructed to rest inside the fMRI scanner and close their eyes. (2) Visuospatial task: Subjects were asked to report the location of small Gabor patches briefly presented either in the left or right visual periphery using their left middle or index finger, respectively. The contrast of the Gabor patches was adjusted at two difficulty levels: one easy and one difficult (see Materials and Methods). Here, Gabor patches are magnified and the scene size is shrunk for clear illustration. (3) Speech task: Subjects were instructed to discriminate two compound CVs as either /da/ or /ga/ using their right middle or index finger. The speech sounds were presented in synchrony with a band‐limited Gaussian white‐noise, and were mixed with a third speech sound /ba/ to make the task challenging. Correspondence between the target CVs and fingers was randomized across subjects. (4) Dual‐task: Subjects performed the visuospatial and speech tasks in parallel. Subjects kept using the same hand and fingers to report their responses. The order of the single tasks was randomized across subjects. (B) Accuracy and Reaction Times During Single‐Task and Dual‐Task Conditions. Bars illustrate mean accuracy and mean of median reaction times. Error bars represent standard error of means. (C) Methodological Framework: Visualization of brain modules derived from a representative subject. Arrows illustrate the hypothesized associations between the brain network modules and dual‐task interference. High topological overlap between single‐task network modules (persistency of modules) can lead to interference effects during dual‐task processing. In contrast, higher topological reconfiguration of single‐task network modules in adaptation to the dual‐task condition (flexibility of modules) possibly leads to less behavioral dual‐task interference. Both hypotheses were tested based on the topological similarity between network modules across different task conditions. (D) Quantification of Topological Similarity Between Task‐Related Network Modules: To estimate the topological similarity between brain network modules, the nodal composition of a pair of modules (m,n) was compared across task‐related networks g ₁ and g ₂ and similarity weights ω_m,n were computed which reflect the percentage of common nodes between modules (m,n). For example, comparing the module m = 1 of the toy graph g ₁ on the left side and the module n = 1 of the toy graph g ₂ on the right side, the two modules share one common node (i.e., node 9) out of six nodes, resulting in the ${\omega }_{1,1}=\frac{1}{6}$ . Similarity weight ω_m,n ranges between zero (nonoverlapping modules) and one (completely overlapping modules). The similarity weights derived from this procedure were then summed and the result was divided by the total number of module comparisons (in our example 9) to obtain a global measure of topological similarity between modules across two networks. To visualize the task‐related brain networks in graph space, the Fruchterman–Reingold algorithm was used to find the optimal node placement for each network with respect to its modules, and the Kamada–Kawai algorithm was employed to place the nodes within each network module [Traud et al., 2009].

Figure 2

Global similarities between task‐related modules and their relations to the decline in visuospatial accuracy in the dual‐task condition. While topological overlap between single‐task modules correlated with the decline in visuospatial accuracy during dual‐task processing, the flexible reconfiguration of single‐task modules in adaptation to the dual‐task condition predicted lower behavioral dual‐task interference. Participants were grouped into a high‐interference (first column, purple) and a low‐interference (first column, green) group according to the decline in their visuospatial accuracy in the dual‐task condition. Mean similarity between modules was quantified across single visuospatial and speech (A), single visuospatial and dual‐task (B), and single speech and dual‐task networks (C) to capture persistency (A) and flexibility (B and C) of functional brain network modules in the context of dual‐tasking. First column: The mean similarities between task‐related modules for high‐ versus low‐interference subjects in visuospatial accuracy are illustrated. The mean similarity weights per subject were integrated across thresholds applied to the connectivity matrices and compared between the two groups by means of a two‐sample permutation test with 10,000 repetitions and the resulting p‐value (two‐sided) was reported. Second column: Differences in mean similarity between both groups were computed (black points) and statistically compared at each connectivity threshold separately by means of permutation tests with 10,000 repetitions. Black points represent the arithmetic difference in means; red points visualize one side of the bias‐corrected and accelerated (BCa) bootstrap confidence interval for a two‐sided test at p<0.05 [Efron, 1987], and solid points represent significant results obtained from the permutation tests at the significance level of p<0.05 (two‐sided). Third column: The Spearman's ρ correlation between mean similarities and interference scores in visuospatial accuracy across all subjects was tested at each connectivity threshold by means of permutation tests with 10,000 repetitions, and the resulting correlation coefficients were considered significant (solid points) if p<0.05 (two‐sided). Black points represent the Spearman's ρ correlation coefficients, red points visualize one side of the bias‐corrected and accelerated (BCa) bootstrap confidence interval for a two‐sided test at p<0.05 [Efron, 1987], and solid points represent significant results obtained from the permutation tests at the significance level of p<0.05 (two‐sided). Black and red curves were estimated using local polynomial regression fitting by means of LOESS functions. Fourth column: Visualization of the linear relationship between mean similarity of brain modules and interference scores in visuospatial accuracy across all subjects at the connectivity threshold where the most significant Spearman's ρ correlation was obtained. Shaded area represents the two‐sided parametric confidence interval at the significance level of p<0.05.

Figure 3

Correlations between different metrics of network modularity and the decline in visuospatial accuracy in the dual‐task condition. Here, we illustrate the magnitude of the Spearman's ρ correlations between the interference scores in visuospatial accuracy and mean topological similarity between modules (red points), modularity Q ^* (grey points), or number of modules (orange points) obtained from the task‐related networks which were compared in the analysis of persistency (comparison across single task networks (A)) or flexibility (comparison across a single task and the dual‐task network (B/C)). The correlations were tested at each connectivity threshold by means of permutation tests with 10,000 repetitions, and the resulting correlation coefficients were considered significant (solid points) if p<0.05 (two‐sided). Differences between the values obtained for a specific metric across different comparisons are due to the stochastic initialization of the modularity detection algorithm. The smooth curves were estimated using local polynomial regression fitting by means of LOESS functions.

Figure 4

Brain regions where nodal similarity between task‐related modules was significantly different between high‐ and low‐interference groups in visuospatial accuracy. (A) Persistency of single‐task modules. Red nodes: Subjects with larger decline in visuospatial accuracy in the dual‐task condition showed significantly larger overlap between single‐task modules within the right fusiform as well as pre‐ and postcentral gyri. Blue node: Low‐interference subjects showed significantly higher nodal consistency of single‐task network modules in the right inferior lingual gyrus. (B) Flexibility of visuospatial task modules in adaptation to the dual‐task condition. Red nodes: Participants with larger decline in visuospatial accuracy in the dual‐task condition showed significantly higher similarity between single visuospatial and dual‐task network modules (that is, lower flexibility of visuospatial task modules) within the posterior cingulate as well as the left middle occipital and postcentral gyri. Blue nodes: Lower flexibility of visuospatial task modules in the right middle occipital gyrus was associated with lower decline in visuospatial accuracy during the dual‐task condition. (C) Flexibility of speech task modules in adaptation to the dual‐task condition. Red nodes: Subjects with larger decline in visuospatial accuracy in the dual‐task condition showed higher similarity between speech task and dual‐task network modules (that is, lower flexibility of speech task modules) within the right superior and middle frontal gyri, anterior cingulate as well as the bilateral pre‐ and postcentral gyri. Blue nodes: Lower flexibility of speech task modules in the right middle occipital and superior temporal gyri and the right rolandic operculum was associated with lower decline in visuospatial accuracy during the dual‐task condition. Node‐specific mean similarities were compared between high‐ and low‐interference groups by means of two‐sample permutation tests with one million repetitions. To visualize the nodal effects, Cliff's delta was used as a nonparametric measure of effect size. Brain nodes are color‐coded according to their Cliff's delta value. A positive delta value (red) indicates a significantly higher similarity between modules in high‐ compared with low‐interference subjects. In contrast, a negative delta value (blue) indicates a significantly higher similarity between modules in low‐ compared with high‐interference subjects. Therefore, while higher nodal similarity of modules at red nodes was associated with larger decline in visuospatial accuracy in the dual‐task condition, higher similarity of modules at blue nodes predicted lower decline and more accurate visuospatial performance in the dual‐task condition. Yellow circle: Within the right somatomotor cortex participants with higher dual‐task costs in visuospatial accuracy showed larger modular overlap between single‐task modules (A), and lower flexibility of speech task modules in adaptation to the dual‐task condition (C). Brain slices: L and R indicate the left and right hemispheres, respectively. The brain volumetric slices represent the group‐averaged T1 image and are ordered from the bottom to the top of the brain (left to right) according to their MNI z coordinate (mm).