From “rest” to language task: Task activation selects and prunes from broader resting‐state network

Abstract

Resting‐state networks (RSNs) show spatial patterns generally consistent with networks revealed during cognitive tasks. However, the exact degree of overlap between these networks has not been clearly quantified. Such an investigation shows promise for decoding altered functional connectivity (FC) related to abnormal language functioning in clinical populations such as temporal lobe epilepsy (TLE). In this context, we investigated the network configurations during a language task and during resting state using FC. Twenty‐four healthy controls, 24 right and 24 left TLE patients completed a verb generation (VG) task and a resting‐state fMRI scan. We compared the language network revealed by the VG task with three FC‐based networks (seeding the left inferior frontal cortex (IFC)/Broca): two from the task (ON, OFF blocks) and one from the resting state. We found that, for both left TLE patients and controls, the RSN recruited regions bilaterally, whereas both VG‐on and VG‐off conditions produced more left‐lateralized FC networks, matching more closely with the activated language network. TLE brings with it variability in both task‐dependent and task‐independent networks, reflective of atypical language organization. Overall, our findings suggest that our RSN captured bilateral activity, reflecting a set of prepotent language regions. We propose that this relationship can be best understood by the notion of pruning or winnowing down of the larger language‐ready RSN to carry out specific task demands. Our data suggest that multiple types of network analyses may be needed to decode the association between language deficits and the underlying functional mechanisms altered by disease. Hum Brain Mapp 38:2540–2552, 2017. © 2017 Wiley Periodicals, Inc.

INTRODUCTION

Resting‐state networks (RSNs) have been a major center of interest in the last decade as they show spatial patterns generally consistent with networks revealed during a relevant cognitive task. This observation led to the suggestion that cognitive networks remain dynamically active at rest (Doucet et al., 2011; Fox et al., 2005; Smith et al., 2009). To reveal these relevant RSNs, functional connectivity (FC) analyses have been conducted using either a seed‐based (Fox et al., 2005) or an independent component analysis (ICA) (Damoiseaux et al., 2006) approach. For instance, Smith et al. (2009) demonstrated a strong spatial similarity between intrinsic RSNs and the networks revealed through a variety of typical functional magnetic resonance imaging (fMRI) activation paradigms. This finding suggested that a given cognitive network remains intrinsically and functionally active during a diverse set of mental and cognitive processes. To our knowledge, however, only a limited number of studies have examined whether the brain networks involved in fMRI task activation and a relevant RSN are identical when compared in the same sample (Alnaes et al., 2015; Breckel et al., 2013; Buckner et al., 2013; Hasson et al., 2009). Without such a direct comparison, the correspondence between them cannot be assumed, and the cognitive state or function in which they each play a role cannot be considered identical.

Recent evidence has shown that the brain representation of a cognitive network, as defined by resting‐state activity, is more complex than previously thought. Buckner et al. (2013) argued that RSNs reflect a combination of stable anatomically constrained and state‐dependent signal components. In line with this theory, they revealed the spatial variability of the network associated with a semantic classification task by describing a more left‐lateralized network during task implementation than at rest. Thus, this highlights network configuration change in association with the two different conditions. Their study suggests that while a RSN and a cognitive task network display some systematic relationships or associations, even subtle changes in cognitive activity can modulate the patterns of activity observed, begging the question as to exactly how these two depictions of cognitive network activity are related. For instance, Alnaes et al. (2015) revealed that dorsal attention network connectivity was modulated by attentional load levels, describing increased connectivity between the network and the visual cortex, and decreased connectivity with the sensory‐motor cortex. Their results supported the existence of dynamic network reconfigurations based on attentional effort. While their study was limited to healthy participants, the authors highlighted how such an investigation shows promise for decoding and characterizing altered FC related to abnormal cognitive functioning in clinical populations.

In the context of clinical populations, it is well‐known that patients with unilateral temporal lobe epilepsy (TLE) suffer from language deficits, which may be partially caused by abnormal language reorganization (Pravata et al., 2011; Waites et al., 2006). However, these studies only focused on one state (resting‐state or language task), preventing any interpretation about whether these language deficits were related to functional abnormalities present in one or both state(s), or caused by an abnormal network transition from one state to the other one.

In this study, we investigated changes in network spatial configuration during an expressive language task (i.e., verb generation (VG)) and a conscious resting‐state condition using a seed‐based FC approach. More specifically, we tested whether a similar, spatially constant language network could be obtained during multiple conditions: a VG period, a resting‐state condition set between the VG active periods, and a continuous resting‐state condition. A total of 72 volunteers, including 24 healthy participants and 48 patients with TLE, were studied. To derive the language network from each condition, we utilized a group‐based seed in the left IFC, capturing the major expressive language center (Broca's area). We tested whether spatially identical networks could be extracted from each condition, focusing in particular on the spatial similarity of an FC‐based language network present during rest, as opposed to the more standard task‐activation approach, within each group. We hypothesized that the continuous resting‐state condition would demonstrate a broader, less lateralized and less task‐specific network, potentially indicative of the wider set of regions available for implementing expressive language activity. Moreover, we hypothesized that the spatial changes would vary as a function of the experimental group, with TLE altering responses to the VG task, in addition to altering both the temporally proximal and distant resting‐state conditions. In short, the purpose of this study was to determine whether the expressive language network implementing a covert‐speech VG task is spatially similar to the intrinsically (functionally) connected language network. Last, we examined whether objectively measured verbal fluency scores were more closely associated with the task‐driven or the intrinsic networks.

MATERIALS AND METHODS

Participants

A total of 48 patients with refractory unilateral TLE (24 left‐sided (LTLE) and 24 right‐sided (RTLE)) matched on age, handedness, and gender were recruited from the Thomas Jefferson University Comprehensive Epilepsy Center. A combination of EEG, MRI, PET, and neuropsychological testing was used to lateralize the side of seizure focus (Sperling et al., 1996). All patients met the following criteria: unilateral temporal lobe seizure onset through surface video/EEG recordings; normal MRI or MRI evidence of mesial temporal sclerosis (MTS) in the epileptogenic temporal lobe; concordant PET finding of hypometabolism in the ictal temporal lobe. TLE patients were excluded from the study for any of the following reasons: previous brain surgery; extratemporal or multifocal epilepsy; medical illness with central nervous system impact other than epilepsy; contraindications to MRI; right‐hemisphere dominance for language (based on an fMRI language task (see below for the description of the task)); psychiatric diagnosis other than an Axis‐I Depressive Disorder; or hospitalization for any Axis I disorder listed in the Diagnostic and Statistical Manual of Mental Disorders, IV. Depressive disorders were allowed given the high comorbidity of depression and epilepsy (Tracy et al., 2007).

Twenty‐four healthy controls were recruited to match the patient participants in age, gender, and handedness. All controls were free of psychiatric or neurological disorders based on health screening measures. This study was approved by the Institutional Review Board for Research with Human Subjects at Thomas Jefferson University.

All participants have provided a written informed consent. All participants were native English speakers.

MRI Data Acquisition

All participants underwent MRI on a 3 T X‐series Philips Achieva clinical MRI scanner (Amsterdam, The Netherlands) using an 8‐channel head coil. A total of 5 min of a resting‐state condition was collected, as well as a VG task to provide an expressive language network. Anatomical and functional acquisitions were similar for all participants. Regarding the resting‐state condition, participants were instructed to remain still, keep their eyes closed but not fall asleep throughout the scan. Single shot echoplanar gradient echo imaging sequence acquiring T2* signal was used with the following parameters: 120 volumes, 34 axial slices acquired parallel to the AC–PC line, TR = 2.5 s, TE = 35 ms, FOV = 256 mm, flip angle = 90°, 128 × 128 data matrix voxels, gap = 0 mm). The in‐plane resolution was 2 mm² and the slice thickness was 4 mm. Regarding the VG task, participants were instructed to covertly generate an action word in response to a viewed noun (common objects) presented on a screen. Each word was presented for 2 s, within a 30‐s block (ON condition). These blocks were alternated with passive viewing of a central stimulus (#####) in epochs of 30 s (OFF condition) for a total of 5 min. The scanning parameters were similar to those described for the resting‐state condition, except for the number of axial slices of 36. A training session was conducted before entering the scanner to ensure that the instructions were understood. After the task, all participants reported that they were able to complete the task as instructed. Prior to the collection of the functional images, T1‐weighted images were collected using an MPRage sequence (180 slices, 256 × 256 voxels; TR = 640 ms, TE = 3.2 ms, flip angle = 8°, FOV = 256 mm) in positions identical to the functional scans to provide an anatomical reference. The in‐plane resolution for each T1 slice was 1 mm². Each EPI imaging series started with three discarded scans to allow for T1 signal stabilization.

Preprocessing Analyses

VG and resting‐state fMRI data were preprocessed in the same way using SPM8. Slice timing correction was used to adjust for variable acquisition time over slices in a volume, with the middle slice used as reference. Next, a six‐parameter variance cost function rigid body affine registration was used to realign all images within a session to the first volume. Motion regressors were computed and later used as regressors of no interest. To maximize mutual information, coregistration between functional scans and the structural image (MPRage) was carried out using six iterations and resampled with a 7th‐Degree B‐Spline interpolation. Functional images were then normalized and warped into standard space (MNI152) to allow for signal averaging across subjects. We utilized the standard normalization method in SPM8. All normalized images were smoothed by convolution with a Gaussian kernel, with a full‐width at half maximum of 8 mm in all directions. For the VG task, these preprocessed data were entered in the first‐ and second‐level analyses to reveal the activation network.

For the FC analyses only, we added additional preprocessing steps, identical for both resting‐state and VG data. In detail, sources of spurious variance were removed through linear regression: six parameters obtained by rigid body correction of head motion, the cerebrospinal fluid, and white matter signals. Finally, both the VG and the resting‐state data were temporally filtered in the band [0.008–0.1] Hz (Cordes et al., 2001).

VG Analysis

Unthresholded individual networks emerging from the VG task were entered into a second‐level random‐effects analysis (one‐way ANOVA) to determine the network at the group level as well as the differences between the experimental groups (LTLE, RTLE, Normal Controls, NCs). A height threshold was set at P < 0.05 (FWE‐corrected) at the whole‐brain level, with a cluster size K > 26 voxels, for the group network (Supporting Information, Table S1). A height threshold was set at P < 0.001 (uncorrected) at the whole‐brain level, and P < 0.05 (FWE‐corrected) at the cluster‐level (i.e., cluster size K > ∼100 voxels), for the comparison between the groups. The voxel associated with the maximal peak activation (e.g., highest T‐value within the most significant cluster, see Supporting Information, Table S1) in the left IFC was extracted as a seed for the FC analysis, separately for each group (LTLE, RTLE, and NC). This network, based on the full VG task (On minus Off epochs), will be referred to as the “VG‐Full Network.” The group analyses were done in SPM8.

Definition of the Seed Region

The seed region was defined based on the maximal activation peak within a region‐of‐interest mask located in the left IFC pars opercularis (IFC‐oper, Supporting Information, Table S1), defined at the group level. The pars opercularis region was specifically selected because it was the only region including a maximal peak of activation in the three groups (relative to pars orbitalis and pars triangularis). A 6‐mm‐radius sphere was created, separately for each group (Fig. 1). A 6‐mm‐radius sphere was chosen to (1) identify the strongest language activation (i.e., the major language center), specific to each group; (2) have a complete overlap between the seed and the activation map, regardless of the group. In addition, the seed only included voxels with an activation level of T value ≥ 6, confirming the high involvement of this region in language processing, regardless of the group.

Figure 1.

Point of maximal activation located in the inferior frontal cortex pars opercularis, during the verb generation task, for each group. Seed used for the functional connectivity analyses, for each group (controls: red (center: −52, 14, 4); right TLE: green (center: −52, 10, 2); left TLE: blue (center: −48, 12, 24)). [Color figure can be viewed at http://wileyonlinelibrary.com]

FC Computation

For each individual, a correlation map was produced by extracting the time course from the seed region and computing the correlation between that time course and the time course from all other brain voxels. These maps were then submitted to a Fisher r‐to‐Z transformation. All further analyses were conducted on these transformed data. Last, three networks were extracted for each individual: one from the ON condition of the VG task (to be referred to as the “VG‐ON Network,” VG‐onN), one from the OFF condition of the VG task (“VG‐OFF Network,” VG‐offN), and one from the continuous resting‐state (RSN). These network computations and extractions were done using in‐home Matlab scripts.

For validation purposes, we conducted additional analyses to test the effect of hemodynamic shift on our VG‐onN and VG‐offN findings. To do so, we shifted the time series associated with the ON and OFF conditions, respectively, by 5 s (i.e., 2 volumes) to account for the delay in the hemodynamic response (Hasson et al., 2009). Following this shift, we concatenated the functional acquisitions associated with the ON and OFF conditions, respectively. Owing to the 5‐s shift of the time‐series, this step resulted in a time‐series of 62 volumes for the VG‐off condition and 58 volumes for the VG‐on condition. Using these new time‐series, we then recomputed the two networks (VG‐onN and VG‐offN), as described in the above paragraph. Correlational data capturing the strong relationship between the networks yielded by the two approaches (no shift vs hemodynamic shift) is presented in Supporting Information, Table S2.

Computation of Laterality Index

For each network (e.g., VG‐Full network, RSN, VG‐onN, and VG‐offN), individual laterality indices (LIs) were computed using the LI toolbox available in SPM8, which relies on a bootstrap method (Wilke and Schmithorst, 2006). We used an inclusive mask containing the frontal, parietal, and temporal lobes, excluding the interhemispheric space (10 mm). LI values ranged between −1 and +1, where −1 indicated left‐sided lateralization.

To compare the LI values, an ANOVA with Network (VG‐Full, VG‐onN, VG‐offN, and RSN) as a within‐subject factor and Group as a between‐subject factor (LTLE, RTLE, and NCs) was computed.

Spatial Comparison Between the FC‐Related Networks

Because the seed regions within the IFC‐oper were different for the three groups, all analyses comparing the three networks described above were carried out and reported only within each of the experimental groups.

For each group, the main analyses involved paired t‐tests computed to compare VG‐onN versus RSN, VG‐offN versus RSN, and VG‐onN versus VG‐offN. For each network comparison, a height threshold was set at P < 0.001 (uncorrected) at the whole‐brain level, and P < 0.05 (FWE corrected) at the cluster‐level (e.g., K > approximately 100 voxels).

Computation of Goodness‐of‐Fit

To compare the networks between the groups, we used a goodness‐of‐fit (GOF) approach (Greicius et al., 2004). In other words, a linear template‐matching procedure was used, which takes the average z‐score of voxels falling within the template minus the average z‐score of voxels falling outside the template. Higher GOF reflects the degree to which the network matches the template. We tested two templates: first, we selected the VG‐full network specific to each group as the template, and computed a GOF for each individual and each FC‐based network (VG‐onN, VG‐offN, and RSN; e.g., a control individuals' network matching the NC VG‐Full Network, a left TLE individuals' network matching the LTLE VG‐Full Network, etc.). This template matching procedure allowed us to compare the spatial similarity between the language network extracted from the full VG task and each FC‐based network. Second, we chose the group‐specific RSN as the template, and computed a GOF for the two other FC‐based networks (VG‐onN and VG‐offN), for each individual. Note, while the paired t‐tests described above provided a local measure of similarity, this procedure provided a global measure of similarity between the FC‐based networks.

We also tested whether the GOFs differed between the groups and/or the networks. To do so, an ANOVA with Network (VG‐Full Network, VG‐onN, VG‐offN, and RSN) as a within‐subject factor and Group as a between‐subject factor (LTLE, RTLE, and NC) was computed. If significant (p < 0.05), posthoc analyses were computed and reported at p < 0.05 (Bonferroni corrected).

Relation Between Network Laterality Index, Goodness‐of‐Fit, and Language Performance

To test for an association between the language networks and objective language performance, we obtained a measure of verbal semantic fluency, utilizing the total number of words generated for a semantic category (Animals, from Controlled Oral Word Association Test (Benton et al., 1994)).

Within each experimental group, we computed Spearman's correlations between the LIs (or the GOFs) and the verbal fluency measure. To allow potentially meaningful effects to be considered in future study designs, the threshold for statistical significance for this analysis was set at P < 0.05, uncorrected.

Statistical Analyses of the Behavioral Data

The clinical and behavioral data were compared between the three experimental groups using independent sample t‐test, one‐way ANOVA, or chi‐square tests as appropriate.

RESULTS

Behavioral Data

The three experimental groups did not differ by age or gender (Table 1). The right and left TLE groups did not differ by illness duration or full‐scale IQ. However, the groups did differ in their neuropsychological language scores (Table 1, P = 0.001). In detail, the left TLE patients displayed significantly lower semantic scores than the healthy controls (P < 0.001) and the right TLE patients (P = 0.037). The right TLE patients did not differ from the healthy controls (P = 0.08).

Table 1.

Clinical and behavioral description of the experimental groups

	Controls	Left TLE	Right TLE
N (females)	24 (12)	24 (9)	24 (12)
Age (SD)	37.2 (11.3)	41.6 (13.6)	40.0 (13.7)
TLE duration (years)	‐	18.4 (17.3)	14.8 (13.4)
Presence of MTS (N)	0	16	10
Type of seizures	‐	SPS: 4 CPS: 24 GTCS: 5 2nd GTCS: 10	SPS: 9 CPS: 22 GTCS: 3 2nd GTCS: 7
FSIQ	‐	95 (15)	99 (12)
Verbal semantic score	23.0 (6.5)	16.3 (4.9)	20.0 (6.5)

Abbreviations: CPS: complex partial seizures; FSIQ: full‐scale intelligence quotient; GTCS: generalized tonic clonic seizures with CPS still the primary seizure type; 2^nd GTCS: secondary generalized tonic clonic seizures with CPS still the primary seizure type; MTS: mesial temporal sclerosis; SPS: simple partial seizures only.

VG Language Activation Network

All the participants were left‐hemisphere dominant for language based on the full VG task data (LI = −0.74; Table 2). The fundamental regional location and spatial extent of the VG‐Full Task activation network did not differ between the groups (Fig. 2, top raw).

Table 2.

Network LIs for each group

	VG‐full network (activation)	VG‐OnN	VG‐OffN	RSN
Controls	−0.79 (0.12)	−0.72 (0.17)	−0.65 (0.19)	−0.61 (0.17)
Right TLE	−0.72 (0.19)	−0.63 (0.2)	−0.62 (0.2)	−0.65 (0.13)
Left TLE	−0.72 (0.16)	−0.64 (0.19)	−0.57 (0.22)	−0.7 (0.14)

Mean LI (standard‐deviation) for each group and network.

Figure 2.

Language network extracted in various conditions, for each group. (A) Network extracted from the VG activation task (Off minus On epochs). (B) Functional connectivity‐based network extracted from the VG‐on condition. (C) Functional connectivity based network extracted from the VG‐off condition. (D) Functional connectivity‐based network extracted from a continuous resting‐state condition. [Color figure can be viewed at http://wileyonlinelibrary.com]

The peaks of maximal activation (T‐value > 10) within the left IFC for the right TLE and controls (distance < 5 mm) were close by, while the left TLE group displayed an activation maxima at a more distant and superior voxel (distance > 20mm with both right TLE and control groups) (Fig. 1).

Across all groups, the major cluster was located in the left IFC (Supporting Information, Table S1 and Figure 2 row A). This activation covered the subregions of the IFC, namely, the pars orbitalis, pars triangularis, and pars opercularis, extending to the precentral gyrus. To a lesser degree, a small contralateral cluster was observed in the right IFC/insula, for each group. Activation was also revealed in the supplementary motor area (SMA) as well as in the left inferior parietal gyrus. Finally, a subcortical cluster was evident in the left putamen for both the controls and left TLE patients.

FC‐Related Networks Comparison

FC‐related networks were generated using the seed of the activation maxima, extracting, and defining this network separately and uniquely for the different conditions (rest, VG‐onN, and VG‐offN), within each experimental group. Figure 2 displays the seed‐based results for the VG‐onN (B), VG‐offN (C), and RSN (D) analyses.

Controls

VG‐onN and RSN comparison

The RSN was larger and more bilateral than the VG‐onN, which was more left‐lateralized (Figure 2 and Table 2). In detail, the RSN showed additional recruitment of right‐sided regions, including a large cluster including the supramarginal gyrus (T = 7.9; P < 0.001, FWE corrected at the cluster level), and smaller clusters in the cuneus (T = 6.0; P < 0.001, FWE corrected at the cluster level), caudate (T = 6.0; P = 0.02, FWE corrected at the cluster level) and the orbitofrontal cortex (T = 5.7; P < 0.001, FWE corrected at the cluster level) (Supporting Information, Table S3). In contrast, the VG‐onN showed stronger FC between the seed and almost exclusively left‐sided regions, localized in the thalamus (T = 6.8; P = 0.03, FWE corrected at the cluster level), mesial temporal lobe (T = 6.3; P = 0.002, FWE corrected at the cluster level), and IFC (T = 6.0; P = 0.003, FWE corrected at the cluster level). Also, it revealed stronger connectivity with a right cluster in the cerebellum (T = 5.1; P < 0.001, FWE corrected at the cluster level).

VG‐offN and RSN comparison

When comparing the VG‐offN and the RSN, results were spatially similar to the comparison between the VG‐onN and RSN, although to a weaker extent (Supporting Information, Table S3 and Fig. 2). In other words, the RSN mostly recruited a large cluster in the right frontal cortex (T = 7.0; P < 0.001, FWE corrected at the cluster level), the supramarginal gyrus (T = 6.4; P < 0.001, FWE corrected at the cluster level), and the superior temporal gyrus (T = 5.8; P = 0.02, FWE corrected at the cluster level) while the VG‐offN recruited clusters in the left IFC (T = 6.7; P = 0.01, FWE corrected at the cluster level), left thalamus (T = 4.9; P = 0.04, FWE corrected at the cluster level), and right cerebellum (T = 5.0; P = 0.05, FWE corrected at the cluster level).

VG‐onN and VG‐offN comparison

The direct comparison between VG‐onN and VG‐offN did not reveal significant differences.

Right TLE

VG‐onN and RSN comparison

The RSN was larger and more bilateral than the VG‐onN (Fig. 2). In detail, the RSN showed additional recruitment of regions located in the right middle temporal cortex (T = 5.6; P = 0.02, FWE corrected at the cluster level). In contrast, the VG‐onN showed stronger FC between the seed and clusters in the left inferior (T = 8.7; P = 0.001, FWE corrected at the cluster level), superior (T = 6.9; P = 0.02, FWE corrected at the cluster level), and middle (T = 4.7; P = 0.05, FWE corrected at the cluster level) frontal cortex, as well as the left thalamus (T = 5.8; P = 0.02, FWE corrected at the cluster level), and the right SMA (T = 5.04; P = 0.005, FWE corrected at the cluster level) (Supporting Information, Table S3).

VG‐offN and RSN comparison

The RSN mostly recruited a large cluster in the left superior temporal cortex (T = 5.4; P = 0.04, FWE corrected at the cluster level), while the VG‐offN recruited clusters in the left dorsolateral frontal cortex (T = 5.8; P = 0.05, FWE corrected at the cluster level) and the right anterior cingulate cortex/SMA (T = 5.1; P < 0.001, FWE corrected at the cluster level; Supporting Information, Table S3).

VG‐onN and VG‐offN comparison

The VG‐onN had increased FC between the seed and one cluster located in the left IFC pars triangularis (T = 7.9; P = 0.04, FWE corrected at the cluster level), relative to the VG‐offN.

Left TLE

VG‐onN and RSN comparison

The left TLE's RSN showed larger recruitment, relative to the VG‐onN. In detail, the RSN recruited additional regions in the precuneus (T = 5.9; P = 0.001, FWE corrected at the cluster level), bilaterally, and the left IFC (T = 5.5; P < 0.001, FWE corrected at the cluster level). The VG‐onN recruited exclusively left‐lateralized regions located in the superior frontal cortex (T = 6.7; P = 0.003, FWE corrected at the cluster level), pallidum (T = 4.2; P = 0.004, FWE corrected at the cluster level), cerebellum (T = 5.1; P = 0.002, FWE corrected at the cluster level), and orbitofrontal cortex (T = 4.6; P = 0.001, FWE corrected at the cluster level) (Supporting Information, Table S3).

VG‐offN and RSN comparison

When comparing the VG‐offN and RSN, the results were spatially similar to the results noted above, although FC differences appeared smaller in spatial extent. In detail, compared to the VG‐offN, the precuneus (T = 5.1; P < 0.001, FWE corrected at the cluster level) and the left middle frontal cortex (T = 4.3; P = 0.001, FWE corrected at the cluster level) were more notably recruited for the RSN; while a cluster in the left orbitofrontal cortex (T = 4.8; P = 0.045, FWE corrected at the cluster level) was more connected to the seed for VG‐offN (Supporting Information, Table S3).

VG‐onN and VG‐offN comparison

The VG‐onN did not show increased FC, relative to the VG‐offN. In contrast, the VG‐offN had increased FC in the right IFC pars triangularis (T = 4.8; P = 0.03, FWE corrected at the cluster level) and the right inferior parietal gyrus (T = 5.0; P < 0.001, FWE corrected at the cluster level), in comparison to the VG‐onN.

Summary of the Spatial Differences Between the FC‐Based Networks

Overall, we found that the major spatial differences were revealed between the VG‐related networks and the RSN, regardless of the group. In particular, the RSN appeared strongly bilateral, recruiting multiple regions contralateral to the seed. This was also the case for the VG‐off condition, though to a smaller degree. Only very limited differences were observed between the VG‐onN and VG‐offN, with the VG‐offN and the VG‐onN exclusively recruiting right‐ and left‐sided regions to the seed, for the left and right TLE group, respectively. This strong similarity might have been influenced by hemodynamic shift or drift across the On and Off blocks of the VG task. To test for this effect, we recomputed the networks with a 5‐sec shift and found a consistently high spatial correlation between the networks identified with and without the 5‐sec shift (r ≥ 0.9, regardless of the groups; see Supporting Information, Table S2). This suggests that our data were not substantially influenced by hemodynamic shift or drift.

Goodness‐of‐Fit Results

Template: RSN

To directly compare the groups and the FC‐based networks, we computed a GOF measure for both VG‐onN and VG‐offN, relative to the RSN, with the latter used as a template. Note, a higher GOF indicates a better spatial match between the network and the template. We computed an ANOVA with Networks (VG‐onN and VG‐offN) as a within‐subject factor and Group as a between‐subject factor (LTLE, RTLE, and NC). The results revealed a significant effect for Group (F = 3.5, P = 0.036). Post‐hoc analyses revealed a trend such that the NC group had a lower GOF than the LTLE group (P = 0.059, Bonferroni corrected; Cohen's d effect size = 0.68). No significant interaction or network effect was revealed. This suggests that the VG‐based networks strongly overlap and are subsumed or included in the RSN, regardless of experimental group.

Template: VG‐full network

The ANOVA revealed a significant main effect for Network (F = 39; P < 0.001), indicating that the RSN differed the most from the VG‐Full network (P < 0.001; Fig. 3). We also revealed a significant effect for Group (F = 6.3, P = 0.003). Post‐hoc analyses showed that both left TLE and NC groups differed from the right TLE group (P = 0.013 and P = 0.007, Bonferroni corrected; Cohen's d effect size = 0.57 and 0.63, respectively), with the latter showing higher GOF values. The interaction between Group and Network was also significant (F = 4.7, P = 0.012), indicating that the right TLE group had a higher GOF than both the LTLEs and NCs for both the VG‐onN (Cohen's effect size = 0.82 and 0.83, respectively) and VG‐offN (Cohen's effect size = 0.66 and 0.64, respectively), but not for the RSN.

Figure 3.

Results of the ANOVA on GOF for each FC‐based network compared to the VG‐Full network template. (A) shows the main effect of network. (B) shows the interaction between network and group. [Color figure can be viewed at http://wileyonlinelibrary.com]

Laterality Index Measure

An ANOVA on the LI with Network (VG‐Full Network, VG‐onN, VG‐offN, and RSN) as a within‐subject factor and Group as a between‐subject factor (LTLE, RTLE, and NC) revealed a significant main effect for Network (F = 16.5, P < 0.001) and a significant Group by Network interaction (F = 3.5, P = 0.034) (Fig. 4 and Table 2). Post‐hoc tests (paired t‐tests) on the Network effect showed that the VG‐Full network was more left lateralized than the three FC‐based networks (P < 0.001, for all networks), indicating that regional recruitment is more left‐hemisphere lateralized when driven by task activation than when generated by FC measures. Within the FC‐based networks, the VG‐onN was more left‐lateralized than the VG‐offN (t = −2.9, P = 0.004). The post‐hoc analyses for the interaction revealed that only in the control group was the RSN significantly less lateralized than the VG‐Full (P = 0.001, Cohen's d = 0.83) and VG‐on (P = 0.01, Cohen's d = 0.57) networks, suggesting that brain networks associated with a key language center become less lateralized in the setting of undirected (nontask) mental activity in healthy individuals. Importantly, this result is in line with the regional results we report above (depicted in Fig. 2). In contrast, for the right TLE group, the LI did not differ significantly between any of the networks. For both the NC and left TLE groups, the LI of the VG‐offN was higher (i.e., the network less lateralized) than both the VG‐Full (Cohen's d = 0.63 and 0.54, respectively) and the VG‐on networks (Cohen's d = 0.52 and 0.50, respectively), but not the RSN. These contrasting group patterns are likely driving the significant interaction.

Figure 4.

Results of the ANOVA on the LI of each network. (A) shows the main effect of network. (B) shows the interactions between network and group. [Color figure can be viewed at http://wileyonlinelibrary.com]

Summary of the LI/GOF Results

Overall, regardless of group, the networks that differ the most from each other were the RSN and the VG‐Full, with the VG‐onN and VG‐offN showing intermediate changes. Importantly, for the healthy controls only, our data revealed a more bilateral RSN relative to the task‐related networks (VG‐Full and VG‐onN). The RSN in normals appears to recruit not just more left‐sided regions, but also more right‐sided regions not present in the task‐related VG network. When combining this observation with our GOF results, the data suggest that the VG‐related networks are actually subsets of the RSN, with the RSN completely overlapping and subsuming the VG networks. In contrast, the TLE patients showed different LI patterns, with the RSN more lateralized than expected, and VG‐on and VG‐off networks significantly different from the RSN. This suggests that pathology alters both intrinsic and language FC‐based networks, though likely for different reasons in left and right TLE. The alterations in FC‐networks point to atypical language representations, and suggest a reorganization of language functions has occurred in our TLE groups.

Association With Verbal Fluency Performance

Verbal fluency, as indexed by a semantic fluency measure, was mildly associated with both our LI and GOF measures in the LTLE group. This group's verbal fluency scores are impaired at approximately one SD below their same age peers (Table 1). Within the LTLE group, better verbal fluency was associated with a more left lateralized task activation network (r = −0.47, P = 0.02) and a better fit to the RSN (GOF for VG‐onN: r = 0.42, P = 0.041; GOF for VG‐offN: r = 0.47, P = 0.02). No significant correlations were observed among the LI, GOF, and verbal fluency measures in either the RTLE or NC groups.

DISCUSSION

This study is among the first to provide a direct comparison between a network instantiated by a task‐driven cognition and an intrinsic FC‐based network, with both networks resulting from the same cognitive processing domain (i.e., expressive language). In the last decade, most resting‐state fMRI studies have interpreted FC‐based networks as cognitive networks (Smith et al., 2009). Our data provide a more specific delineation of their relationship to task‐driven cognition. We show that while a language‐related RSN does not relate to objectively measured language skills, it configures and makes resonant a broad, spatially expansive network from which a task‐based network is selected to meet the demands of a specific task.

Regarding the language (VG‐Full) network, while we did not find statistical differences between the groups, qualitative differences can be observed (Fig. 2A and Supporting Information, Table S1). In particular, the major differences for the Full Task network were located in the left inferior parietal gyrus and other posterior temporal regions, with the controls showing larger clusters than the LTLE and RTLE. These areas have been associated with various aspects of reading and lexical responses (Schwartz et al., 1999). Also, subcortical activity (e.g., putamen) was not present in the RTLE group. Evidence of abnormal language networks in TLE has been consistently described and interpreted in association with language deficits (see reviews by Balter et al., 2016 and Hamberger and Cole, 2011).

Regarding the FC‐based networks, we found that for the controls and, to a lesser degree, the left TLE patients as well, the RSN recruited regions bilaterally, whereas both VG‐on and VG‐off conditions produced more left lateralized FC networks that matched more closely the expressive language network revealed through the activation paradigm. The regions picked up by the RSN involved right‐sided regions located in IFC, precentral, and prefrontal cortices, though clearly the spatial extent of the recruitment was broader and more robust in the controls. This difference does suggest that at the level of the RSN, atypical language organization was present in TLE. Note, while the right TLE group also showed qualitative recruitment of right hemisphere regions for the RSN, relative to the other networks (Fig. 2), our quantitative analysis indicated a relative consistency in LI between the four networks (Table 2 and Fig. 4). In no case did the right‐sided areas of the VG‐on or VG‐off task show unique FC not picked up by the RSN. Similarly, the left‐sided FC evident in the RSN map largely subsumed and overlapped the connectivity evident in both the VG‐on and VG‐off maps, with the RSN areas again broader in terms of spatial extent. Note the VG‐on and VG‐off networks were remarkably similar, in all the groups, with the RSN always displaying a larger spatial extent.

A key activation versus FC difference involved the more limited and selective networks present for the two VG conditions (on and off), a difference quite striking for the right hemisphere, but still present in the left hemisphere. This form of task constraint and lateralization is similar to the RSN/task differences reported by Buckner et al. (2013) for a semantic language task. Our findings suggest that our expressive language‐related RSN captures a wide set of bilateral regions which did not all appear necessary for implementing VG. Instead, the RSN may reflect a set of intrinsic, language‐ready regions, likely supported by anatomical/structural connectivity present between homotopic regions (Joliot et al., 2015; Pizoli et al., 2011). This finding suggests that the integrity (or the left‐lateralization) of the language RSN (LI = −0.65) may not be sufficient for the successful deployment of demanding language skills. However, given the presence of covert speech in our VG task‐related condition and the evidence regarding spontaneous covert speech at rest (Delamillieure et al., 2010), this left‐lateralized feature may be a function of covert speech processes.

Our data indicate that while doing forced covert speech and accomplishing the lexical retrieval demands of the VG task, a narrower set of cells are activated within the wider net of regions connected, resonant, and available within the RSN. The most dramatic example of the winnowing down of this wide RSN net can be found in the right hemisphere, the nondominant hemisphere in our sample. In the setting of our VG task, the absence of this right‐sided activity likely indicates language capable “nondominant” hemisphere regions were not necessary, pointing to the strongly left‐hemisphere dominant nature of our task. This may also indicate that our VG task did not require recruitment of a broader set of executive function, working memory, attentional, or other resource‐demanding processes that might have called upon “nondominant” hemisphere functionality in the setting of other tasks. Therefore, our results demonstrate differences between the RSN and task‐related networks in terms of representing brain connectivity for cognition, and point to the reconfiguring of networks as one moves from “rest” to the demands of a specific expressive language task. Our observation of more left‐lateralized FC‐based networks during the VG on and off conditions suggests that these task‐specific networks are more dictated by phasic functional demands, and represent only a subset of language‐related FC of the RSN. We believe that this latter RSN connectivity may be more reflective of underlying anatomic connectivity, though this possibility would need to be tested in future studies. Out of concern that the strong similarity between the VG‐onN and VG‐offN might have been influenced by hemodynamic shift or drift across the On and Off blocks of the task, we recomputed the networks with a 5‐s shift. These data demonstrated a consistently high spatial correlation between the networks identified with and without a 5‐s shift (r ≥ 0.9, regardless of experimental group; see Supporting Information, Table S2), strongly suggesting this shift effect can be disregarded as an explanation for our result.

Our data show that resting‐state and task‐network configurations differ, with our study offering one of the first quantitative demonstrations of this difference, during an expressive language processing task. Such a modulation between resting‐state FC and a task‐based network has been previously described during an attentional task (Alnaes et al., 2015). Alnaes et al. (2015) have described FC networks as containing information that is highly predictive of the mental state, potentially even able to distinguish cognitive states. Our data, however, suggests that one must use caution when opining on the cognitive interpretation of RSNs. We propose a specific depiction of the relationship between RSNs for a given domain and the specific task instantiations that might occur within that domain. Namely, we argue that this relationship can be best understood by this notion of the pruning or winnowing down of the RSN. This “pruning” hypothesis does not suggest a dissociation between task‐related and RSNs, but clarifies that RSNs instantiate a broad region of neural capabilities (connections) which specific‐task requirements will then come to select or ignore. Missing from our study is the consideration of structural connectivity and the degree to which it more closely resembles and reflects this wide net of RSNs or the “pruned” task‐related connectivity we have observed.

With regard to interpreting RSNs, our data highlight another key point. Our results indicate that pathology can alter intrinsic language FC‐based networks in selective ways, which in turn points to the presence of atypical language representation. For instance, our TLE groups showed different LI patterns, though for both groups, the RSN was more lateralized than it was in normals. We suspect that the reason for this profile likely differs for left versus right TLE pathology. In RTLE, the right‐sided (nondominant) pathology may work to drive down local connectivity with the left hemisphere areas the specific areas important to implementing language. It is worth noting that the RTLE group, which showed relatively normal verbal fluency performance, had a higher GOF (VG‐Full Network Template) with the RSN, VG‐onN, and VG‐offN compared to the LTLE and NC groups. This may suggest that at the global level, a closer match to the intrinsic language RSN helps support language. Our findings may be particularly relevant to the understanding of individual variability in the cognitive capacity of patients. In patients with refractory TLE who are recommended for surgery, the ability to predict language outcomes remains quite limited and is commonly based on an invasive method such as the WADA test (Chelune, 1995). Our findings may thus lead to the development of multiple normative connectivity maps which can be utilized to evaluate the integrity of a cognitive domain such as language (e.g., resting state, task‐on, and task‐off connectivity). Certainly, more investigations are needed to (1) explore other cognitive domains such as attention, episodic memory, and executive function that are particularly at risk in TLE to determine whether a similar mechanism is at work (i.e., pruning down of the resting‐state to form a task‐specific network), (2) explore other language tasks to verify that the pruning mechanism is present and understand its different manifestations, and, last, (3) determine whether task‐based FC can reliably predict cognitive outcome after brain surgery.

In contrast, in the setting of LTLE, a better fit to the normative FC‐based RSN and a more lateralized language network may reflect reorganization to the healthier regions ipsilateral of the ictal hemisphere. In fact, the associations we observed between better verbal fluency and a better fit to the normative FC‐based RSN (n.b., also a more lateralized activation pattern) are consistent with adaptive reorganization, albeit flawed because the absolute level of verbal memory performance is impaired relative to normative levels.

In terms of study limitations, it must be noted that the task‐based networks and RSN differed with respect to whether the participants' eyes were open (task) or closed (resting‐state condition). This may have influenced the regional and spatial extent of the findings, though it would not explain the group or within‐task differences. Additionally, while we decided to use a seed‐based approach similar to other reports (Buckner et al., 2013), other techniques are available, such as ICA (Alnaes et al., 2015; Hellyer et al., 2014) and graph‐theory analyses (Breckel et al., 2013). It is not yet clear which technique is more relevant to determine the relationship and potential modulations that occur between intrinsic RSNs and more specific cognitive activities. In our context, we believe that a seed‐based approach was the best choice to reveal the expressive language network, as many ICA studies failed to reveal the language network at rest (for instance, Damoiseaux et al., 2006; Esposito et al., 2006), and the graph‐theory approach does not focus on one specific network but rather on the whole brain. Our seed‐based approach also had the advantage of capturing the region unique and specific to the task performance of each experimental group. As noted, however, the seeds used did differ slightly within the IFC (n.b., the left TLE group had the most distant seed). While there is literature from classic aphasiology (Benton and Ardila, 1996) and task‐related fMRI (Binder, 1997) showing the importance of the broader IFC for language processing, we cannot exclude the possibility that some of our results involving the FC‐based networks are caused by this spatial difference. We tried to reduce this bias by not comparing the network maps between the groups. Ultimately, we believe that in forcing each group's networks to conform to the same seed, we would be causing an immediate shift away from each group's ground truth, introducing a level of inaccuracy and “abnormality” distinct from what is normative and characteristic of each group. More investigations need to be done to improve our understanding of the relationship between task‐specific instantiations of networks and the more general intrinsic RSNs which they may draw upon. Our proper interpretation of FC‐based networks depends on such understanding. Also, while the duration of both fMRI scans was 5 min each, making the two conditions more comparable, it would be important to investigate the effect of longer scans on the spatial configuration of the FC‐based networks. It is important to keep in mind that, while for a unique continuous resting‐state condition, it is easy to collect more data to get more reliable results for FC, the same is not true when collecting a cognitive task. This concern is heightened when considering clinical populations that may not be able to stay in the scanner for more time (particularly if they have to complete a cognitively challenging task). The VG task (and its duration) has been specifically designed to optimize activation in the shortest possible time frame, and, therefore, is only 5 min long. More investigations are clearly needed using longer cognitive tasks to explore the spatial configuration of cognitive networks using an FC approach. Last, in regards to a recent paper (Eklund et al., 2016), some of our results might be compromised by false‐positive clusters. However, as we used a consistent statistical approach within each group, we believe that this bias may be constant and therefore will not affect our major interpretations. More investigations may be needed using other software and statistical corrections to reproduce our findings.

We showed in this study that depictions of (expressive) language networks are particularly sensitive to the method of extraction (seed‐based FC vs activation) and specific to particular task demands. Future studies should investigate other cognitive behavioral domains (including other language tasks, such as comprehensive language) to determine if our “pruning” hypothesis regarding RSN and task activation represents a more general principle. Moreover, the role of structural connectivity plays in the modulatory relationships that we observed between our language RSN and task‐related network needs to be studied and defined. While most of the RSNs (e.g., such as the default mode network, the motor network, or working memory networks) have been shown to be robustly related to structural connectivity (Greicius et al., 2009; van den Heuvel et al., 2009), the functional–structural relation is less clear for the language network (Morgan et al., 2009). Also, a more complete assessment of the links between cognitive or behavioral performance and the network renderings we examined here (resting‐state FC versus task‐based activation) needs to be undertaken. Last, our data on GOF suggest that, in the setting of TLE pathology, capturing the degree of spatial matching between task‐cognitive networks (e.g., VG‐Full condition) and FC‐based networks may be more sensitive to the atypical forms of language organization that distinguish patients with left or right TLE from healthy individuals.

CONCLUSION

In conclusion, our study is one of the first to investigate the spatial constancy of an expressive language network in healthy participants and patients with focal epilepsy. We compared task‐related and FC‐based networks and found that each network has unique attributes which may be reflective of their task‐dependent (phasic) or task‐independent (tonic) status. We argue that the dynamic FC changes associated with a language task involve a “pruning” of the broader RSN FC to obtain a more precise and spatially efficient configuration of the connectivity needed to successfully perform a task. We also argue that while the RSNs may be sensitive to the effects of pathology, they reflect a different and broader view of network functionality for a given cognitive domain, a view that is less informative about the more localized regions called upon to perform a given task. In this sense, our results clearly indicate that the functional interpretation of RSNs needs to be carefully considered, as it may not be as consistent with task‐induced cognitive networks as it was previously thought. From a clinical perspective, our data suggest that TLE brings with it variability to both task‐dependent and task‐independent networks, with additional indications of atypical language organization for both types of networks. This implies that in the setting of TLE pathology, capturing the language network through FC, and not through clusters of task activation, may be more sensitive to atypical forms of language organization. At the very least, our data show that multiple levels of network analyses (e.g., task‐based activation, task‐related FC, resting FC) may be needed to decode the association between language deficits and the underlying functional mechanisms altered by disease. Identifying such unique network features may move us closer to a functional explanation of the verbal deficits reported in specific diseased populations such as TLE.

Supporting information

Supporting Information