Disrupted dynamic network reconfiguration of the language system in temporal lobe epilepsy

Clinical fMRI assessments of language often neglect the dynamic, interacting nature of the brain. He et al. demonstrate that dynamic network reconfigurations of the language system, selectively enhancing/diminishing regional communication, generate patterns that inform about language integrity in temporal lobe epilepsy with superior sensitivity compared to traditional activation-based static measures.

Abstract

Temporal lobe epilepsy tends to reshape the language system causing maladaptive reorganization that can be characterized by task-based functional MRI, and eventually can contribute to surgical decision making processes. However, the dynamic interacting nature of the brain as a complex system is often neglected, with many studies treating the language system as a static monolithic structure. Here, we demonstrate that as a specialized and integrated system, the language network is inherently dynamic, characterized by rich patterns of regional interactions, whose transient dynamics are disrupted in patients with temporal lobe epilepsy. Specifically, we applied tools from dynamic network neuroscience to functional MRI data collected from 50 temporal lobe epilepsy patients and 30 matched healthy controls during performance of a verbal fluency task, as well as during rest. By assigning 16 language-related regions into four subsystems (i.e. bilateral frontal and temporal), we observed regional specialization in both the probability of transient interactions and the frequency of such changes, in both healthy controls and patients during task performance but not rest. Furthermore, we found that both left and right temporal lobe epilepsy patients displayed reduced interactions within the left frontal ‘core’ subsystem compared to the healthy controls, while left temporal lobe epilepsy patients were unique in showing enhanced interactions between the left frontal ‘core’ and the right temporal subsystems. Also, both patient groups displayed reduced flexibility in the transient interactions of the left temporal and right frontal subsystems, which formed the ‘periphery’ of the language network. Importantly, such group differences were again evident only during task condition. Lastly, through random forest regression, we showed that dynamic reconfiguration of the language system tracks individual differences in verbal fluency with superior prediction accuracy compared to traditional activation-based static measures. Our results suggest dynamic network measures may be an effective biomarker for detecting the language dysfunction associated with neurological diseases such as temporal lobe epilepsy, specifying both the type of neuronal communications that are missing in these patients and those that are potentially added but maladaptive. Further advancements along these lines, transforming how we characterize and map language networks in the brain, have a high probability of altering clinical decision making in neurosurgical centres.

Introduction

Neurological diseases, such as temporal lobe epilepsy (TLE), tend to reshape the organization of cognitive functions in the brain (Tracy et al., 2014). In the case of language, the existence of an epileptogenic temporal region could cause atypical patterns of language representation (Berl et al., 2014), including but not limited to: (i) a left-to-right change in hemispheric dominance (Thivard et al., 2005; Gaillard et al., 2007); (ii) cross or bilateral dominance, with some language regions shifting to the right hemisphere while others remain left lateralized (Thivard et al., 2005; Tracy et al., 2009); and (iii) intrahemispheric reorganization, with affected language regions shifting to adjacent healthy tissue (Bell et al., 2002; Mbwana et al., 2009; for reviews see Dijkstra and Ferrier, 2013; Balter et al., 2016).

Advances in the functional anatomy of language describe ‘dual-streams’ of processing involving left frontal regions (dorsal stream) and bilateral temporal regions (ventral stream) (Hickok and Poeppel, 2007). Regardless of where language-related regions predominantly reside or lateralize, these regions must interact with one another in a time-varying fashion to enable complex linguistic functions. Taking verb generation as an example, one needs to understand the target word first (i.e. comprehension, likely recruiting the ventral stream), and then produce the appropriate action word (i.e. expression, likely recruiting the dorsal stream). During these and related cognitive processes, the brain retains both the functional flexibility necessary to meet evolving task demands, and the functional stability necessary to maintain ongoing linguistic processes. A temporal core-periphery model of functional brain organization operationalizes this notion by positing that the brain simultaneously utilizes stable functional ‘modules’ (a set of brain regions that preferentially interact with one another) and a set of dynamic, transient functional interactions to facilitate task performance (Bassett et al., 2013b). Applying this model to language processing (Fedorenko and Thompson-Schill, 2014), Chai et al. (2016) observed transient, fluctuating interactions between language-related regions during task performance: by defining ‘flexibility’ as the frequency with which a region changes its inter-regional communication over time (Bassett et al., 2011), the authors identified a dynamic organization with both ‘core’ (i.e. less flexible) and ‘periphery’ (i.e. more flexible) domains in the language system.

The core-periphery model of brain function provides an initial insight into the dynamic reconfiguration of language network architecture, as instantiated by the frequency of changes in transient interregional communication. However, the specific nature of these transient communications among regions throughout the course of linguistic processing is unknown. For example, is it not yet known whether frontal regions preferably communicate more with their immediate neighbours (e.g. other nearby ipsilateral frontal regions), or more distant areas (e.g. temporal regions or contralateral frontal regions). It is also not known whether regions in the left hemisphere preferably communicate with their ipsilateral neighbours, or with those in the right, non-dominant hemisphere. To address this gap, we categorized language-related regions into four subsystems: left frontal subsystem (covering the majority of the dorsal-stream), left temporal subsystem (covering the majority of the left ventral-stream), and their two contralateral homologues, similar to the ‘Broca’s’ and ‘Wernicke’s’ masks used in the literature (Rosenberger et al., 2009; Berl et al., 2014) (Fig. 1A). Accordingly, one can define ‘recruitment’ of a region as the probability of intra-communication with peer regions from the same subsystem, and ‘integration’ as the probability of inter-communication with regions from other subsystems (Bassett et al., 2015; Mattar et al., 2015). Taken together, the above frequency-based and probability-based statistics of network reconfiguration provide a dynamic and functionally differentiated notion of language implementation. This approach offers a unique opportunity to reshape our understanding of both adaptive and maladaptive neural representations of language, going beyond the question of ‘where it activates’ toward more complex questions regarding the integrity of dynamic mechanisms that govern inter-regional communication.

Figure 1.

Schematic overview of the approach. (A) Sixteen language-related group-constrained masks (Fedorenko et al., 2010) were used to define functional regions of interest (fROI) by intersecting individual’s activation map with each mask and selecting the top 10% of voxels. Subsystems were defined bilaterally (frontal: red; temporal: cyan). (B) Time series were extracted from functionally-defined regions of interest using both task and resting state functional MRI data, with the subsequent processing steps identical for both modalities. (C) A sliding window strategy (length/step = 40/20 s, 14 windows in total) was used to generate inter-regional coherence matrices over time. (D) Dynamic community structure was detected by maximizing a multilayer modularity quality function (Mucha et al., 2010). (E) Community identities were sorted for each functionally-defined region of interest over time (Module Allegiance) (Bassett et al., 2015; Chai et al., 2016). (F) Dynamic properties were estimated (Bassett et al., 2011, 2015; Mattar et al., 2015; Papadopoulos et al., 2016). RsfMRI = resting state functional MRI.

Here, we exercise this approach in both healthy, language intact individuals, and TLE patients with impaired, potentially non-normative, language representations. We explore the dynamic nature of language networks to answer three specific questions. First, do different language-related regions display different preferences for intra- versus inter-subsystem communication during both phasic (task) and tonic (resting state) conditions? Second, in the face of pressures to reorganize language networks due to pathologies like TLE, do patients demonstrate abnormal dynamic reconfigurations, and are any such abnormal dynamics evident during both task and resting state conditions? Third, are these dynamic reconfigurations of the language system more sensitive to language deficits compared to traditional static activation-based measures commonly used in clinical settings?

We acquired functional MRI data from unilateral TLE patients and matched healthy controls during both a verbal generation task and a resting state condition of identical length and scanning parameters. From the task data, we used a group-constrained subject-specific method to functionally define 16 regions of interest within the language network at individual level (Fedorenko et al., 2010; Chai et al., 2016). Next, using a sliding-window strategy, we generated cross-region coherence matrices over time for both conditions. We then applied dynamic network analysis methods to detect community structures over time (Mucha et al., 2010), and quantify the aforementioned network reconfigurations for both conditions (Fig. 1). We hypothesized that different subsystems will demonstrate different levels of intra- versus inter-subsystem communication during task performance due to their distinct roles, while expecting no such distinction at rest due to the absence of goal-directed cognition. We also predicted that TLE pathology will perturb language network dynamics primarily during task condition, and that the dynamic network measures will be more sensitive to language deficits than static activation-based methods.

Materials and methods

Participants

Fifty patients with refractory unilateral TLE (25 left-sided, 25 right-sided) were recruited from the Thomas Jefferson Comprehensive Epilepsy Center. All patients were surgical candidates for either a standard anterior temporal lobectomy or a thermal ablation of ictal mesial temporal lobe, determined by a multimodal evaluation including neurological history and examination, scalp video-EEG, MRI, PET, and neuropsychological testing (Sperling et al., 1996). All patients met the following criteria: unilateral temporal lobe seizure onset through surface video and EEG recordings; normal MRI or MRI evidence of pathology located within the epileptogenic mesial temporal lobe; and concordant PET finding of hypometabolism in the ictal temporal lobe. Patients were excluded from the study for any of the following reasons: previous brain surgery; medical illness with CNS impact other than epilepsy; lateral-temporal, extra-temporal or multifocal epilepsy; contraindications to MRI; or hospitalization for any Axis I disorder listed in the Diagnostic and Statistical Manual of Mental Disorders, V. Depressive disorders were allowed given the high comorbidity of depression and epilepsy (Tracy et al., 2007).

Specifically, in this study we only enrolled right-handed patients with adequate overall verbal intellectual functions by excluding patients with a verbal IQ < 85 [below 1 standard deviation (SD) of the norm], so that all participants had the cognitive capability to follow instructions and perform the functional MRI task. Given that the functional profile of the language system is associated with handedness (Knecht et al., 2000), participants with left handedness [defined with Edinburg Handedness Inventory (Oldfield, 1971)] were excluded to insure comparability. The demographic and clinical characteristics of the patient groups are presented in Table 1, along with the demographic information of 30 age-, gender-, and education-matched right-handed healthy control subjects. All controls were free of psychiatric or neurological disorders based on a health screening measure. This study was approved by the Institutional Review Board for Research with Human Subjects at Thomas Jefferson University. All participants provided informed consent in writing.

Table 1.

Sample demographic and clinical characteristics

	Left TLE (25)	Right TLE (25)	Healthy controls (30)	F/t/χ2	P
Age	39.64 ± 14.73	43.40 ± 14.84	36.23 ± 11.64	1.870	0.161
Gender (M/F)	12/13	10/15	15/15	0.593	0.743
Education	15.40 ± 2.00	14.52 ± 2.38	15.57 ± 2.31	1.647	0.199
Edinburgh Handedness	91.24 ± 20.17	90.20 ± 25.02	95.63 ± 11.37	0.629	0.536
Phonemic Fluencya	35.64 ± 10.32	36.92 ± 12.59	46.37 ± 10.60	7.708	0.001
Semantic Fluencyb	18.36 ± 5.20	18.29 ± 5.79	22.47 ± 4.69	5.890	0.004
Verbal IQ	101.68 ± 11.42	101.28 ± 11.49	N.A.	0.123	0.902
Performance IQ	99.32 ± 14.26	95.72 ± 13.93	N.A.	0.903	0.371
Full Scale IQ	100.24 ± 11.27	98.32 ± 11.07	N.A.	0.608	0.546
Age at epilepsy onset	22.76 ± 13.77	23.24 ± 13.95	N.A.	−0.122	0.903
Duration of epilepsy	16.92 ± 15.32	20.16 ± 15.85	N.A.	−0.735	0.466
Seizure focality (with/without GS or 2nd GS)	8/17	10/15	N.A.	0.347	0.556
Interictal spike (ipsilateral/bilateral)	21/4	19/6	N.A.	0.500	0.480
Preoperative intracranial EEG recording (Y/N)	5/20	6/19	N.A.	0.116	0.733
Temporal pathology (NB/HS/T/D/E)	10/12/1/1/1	14/9/1/1/0	N.A.	1.074	0.300c
Seizure type			N.A.
CPS	8	7
CPS/SPS	0	3
CPS+2nd GS	5	3
CPS/SPS+2nd GS	5	2
CPS+GS	5	5
CPS/SPS+GS	2	5
Anti-epileptic drugs			N.A.
VGNC	24	23
GABAa agonist	4	6
SV2a receptor mediated	9	8
CRMP2 receptor mediated	10	7
Multi-action	2	5
VGCC	1	2

Continuous variables are presented in mean ± SD.

Temporal pathology was diagnosed by neuroradiologists based on presurgical MRI scans: NB = normal brain; HS = hippocampal sclerosis; T = tumour; D = dysplasia; E = encephalocele. Seizure type: SPS = simple partial seizure; CPS = complex partial seizure; 2nd GS = secondary generalized tonic-clonic seizure; GS = generalized tonic-clonic seizure. Anti-epileptic drugs: VGNC = voltage-gated Na⁺ channel blockage, e.g. phenytoin, carbamazepine, oxcarbazepine, lamotrigine (plus T Type Ca²⁺ channel blockage); GABAa agonist, e.g. diazepam, clonazepam, clobazam, lorazepam, traxene, phenobarbital; SV2a receptor mediated, e.g. levetiracetam; CRMP2 receptor mediated, e.g. lacosamide (plus VGNC blockage); Multi-action: e.g. Na⁺ valproate (VGNC + GABAa agonist), topiramate (VGNC + GABAa agonist + AMPA/kainate receptor blockage + carbonic anhydrase inhibitor); VGCC = voltage-gated Ca²⁺ channel blockage, e.g. pregabalin, gabapentin.

For continuous variables, independent sample t-tests or one-way ANOVAs were carried out, as appropriate. For categorical variables, χ² tests were carried out. Significant differences are highlighted in bold.

^aMeasured by Controlled Oral Word Association and one right TLE patient did not have valid data.

^bMeasured by Animal Naming and one right TLE patient did not have valid data.

^c χ ² test only performed with normal brain and hippocampal sclerosis, since the expected frequencies of other categorical cells were all <5.

Neuropsychological testing

We specifically assessed the verbal fluency capacity for every participant to match the functional MRI task, using a phonemic [Controlled Oral Word Association (Gladsjo et al., 1999)] and a semantic [Animal Naming (Benton et al., 1994)] fluency test. In addition, for patients with TLE, verbal, performance, and full scale IQ, were available as part of their presurgical evaluation.

Imaging acquisition and preprocessing

Two 5-min functional MRI scans, one for the verb generation task and one for the resting state condition, were collected from all participants. During the verb generation task, participants were instructed to covertly generate an action word in response to a viewed noun presented on a screen. Details on the task design, acquisition parameters, data preprocessing, and task activation modelling are described in the Supplementary material. Of note, the three experimental groups did not differ in head micromovement measured by frame-wise displacements (Jenkinson et al., 2002) during either condition (Supplementary Table 1).

Group-constrained, subject-specific functionally-defined regions of interest

To obtain regions most representative of individual language systems, we adopted a method (Blank et al., 2014; Chai et al., 2016) utilizing a set of spatially-constrained masks to define functionally-defined regions of interest for each subject in both hemispheres. These spatially-constrained masks were predefined functionally through a language-localizer functional MRI paradigm in a group of healthy participants with a data-driven method (Fedorenko et al., 2010), providing better functional specificity compared to anatomically defined masks. More specifically, we obtained eight left hemisphere masks from Fedorenko et al. (2010), including the orbital inferior frontal gyrus, inferior frontal gyrus, middle frontal gyrus, anterior and middle-anterior temporal lobe, middle-posterior and posterior temporal lobe, and angular gyrus. These masks were mirror projected onto the right hemisphere to create eight homologous masks, resulting in a total of 16 group-constrained masks (Fig. 1A). For every participant, their individual activation map (T-statistics map for the verb generation > control contrast) was intersected with each mask, and the 10% of voxels with the highest contrast T-values falling within the mask were defined as a functionally-defined region of interest. This procedure ensured that the functionally-defined regions of interest were constant in size across participants and across hemispheres. Compared to traditional structurally defined or group activation defined regions of interest, these subject-specific functionally-defined regions of interest can provide additional sensitivity and functional resolution (Nieto-Castanon and Fedorenko, 2012).

Traditional measures of task activation and laterality

We quantified verb generation task activation at both voxelwise (whole brain search space) and region of interest level through laterality index. Individual laterality indices were computed from the T-statistic maps using the laterality index toolbox with a bootstrap method (Wilke and Schmithorst, 2006). Three inclusive masks were defined using the group-constrained masks by combining bilateral orbital inferior frontal gyrus, inferior frontal gyrus, and middle frontal gyrus (frontal mask, Fig. 1A, red), combining bilateral anterior temporal lobe, middle-anterior temporal lobe, middle-posterior temporal lobe, posterior temporal lobe, and angular gyrus (temporal mask, Fig. 1A, cyan), or combining all of them (whole mask). In addition, we also estimated laterality indices using the average T-values within the subject-specific functionally-defined regions of interest (mean T-values method). Similarly, we grouped all functionally-defined regions of interest into the same frontal, temporal and whole categories, then calculated the laterality index (LI) as: $LI=\frac{T_l-T_r}{T_l+T_r}$, where $T_l$ and $T_r$ are the mean T-values of the left and right masks, respectively.

Network construction

Full details are provided in the Supplementary material. Briefly, we regressed out the major head motion influence from the preprocessed functional MRI data using a 24-parameter model (Friston et al., 1996) for both conditions. The mean time series were then estimated using the 16 functionally-defined regions of interest from each condition. We further applied wavelet decomposition to extract information in the frequency interval of 0.05 ∼ 0.1 Hz (scale 2) (Percival and Walden, 2006). A sliding-window approach (length/step = 40/20 s, 14 windows in total) was applied to parse the decomposed time-series for each condition. We then used wavelet coherence to estimate the adjacency matrix for each window and coupled all 14 windows into a multilayer network (Bassett et al., 2011; Braun et al., 2015).

Dynamic community detection

For each participant, both their verb generation task and resting state multilayer networks were partitioned using a multilayer community detection algorithm to extract groups of brain regions (i.e. communities) that were functionally connected with one another at each layer (Mucha et al., 2010). A description of this algorithm is provided in the Supplementary material. Robustness against the choice of resolution parameters used in the algorithm is provided in the Supplementary material. Note, the quality of multilayer community detection was equivalent across groups for both conditions (Supplementary material and Supplementary Table 1).

Dynamic network statistics

For each dynamic community structure detected from each multilayer network during each condition, we estimated the following dynamic network statistics to characterize the functional interactions among functionally-defined regions of interest over time (see details in the Supplementary material).

Module allegiance

We used this measure to summarize the consistency with which functionally-defined regions of interest are assigned to communities over time (Bassett et al., 2015; Chai et al., 2016).

Recruitment and integration

After categorizing the 16 functionally-defined regions of interest into the following four subsystems: bilateral frontal subsystems (orbital inferior frontal gyrus, inferior frontal gyrus, middle frontal gyrus, Fig. 1A, red), and bilateral temporal subsystems (anterior temporal lobe, middle-anterior temporal lobe, middle-posterior temporal lobe, posterior temporal lobe, angular gyrus, Fig. 1A, cyan), we were able to quantify for each functionally-defined region of interest the probability with which it is assigned to the same community with functionally-defined regions of interest from the same subsystem (recruitment), or with functionally-defined regions of interest from other subsystems (integration) over time (Bassett et al., 2015; Mattar et al., 2015). We verified that this imbalanced subsystem size (three versus five) did not bias our results (Supplementary material).

Flexibility and promiscuity

We used flexibility to describe the frequency with which a functionally-defined region of interest changes its assigned community over time (Bassett et al., 2011) and promiscuity to describe the fraction of all communities in the network in which a functionally-defined region of interest participated at least once (Papadopoulos et al., 2016).

Statistical null models

To quantify the dynamic modular organization of the language system in both patients and healthy controls, we used three different random network null models as benchmarks against which to compare the verb generation task multilayer network: static, nodal, and connectional null models (Bassett et al., 2011, 2013a). Schematics of these null models are displayed in Fig. 2A, with descriptions provided in the Supplementary material.

Figure 2.

Comparison between the real functional MRI data and corresponding null models in all three experimental groups. (A) Schematics for the three null models. Columns of real data of (B) left TLE, (C) right TLE and (D) healthy controls compared to the three null models. Top: Real data compared to the static null model. The real functional MRI data from all three groups presented with distributed module allegiance values, unlike the static null model where values were either 0’s or 1’s. Middle: Real data compared to the nodal null model. The real functional MRI data from all three groups presented with hemispheric specializations in module allegiance, unlike the nodal null model where module allegiance was equal within or between hemispheres. Bottom: A ‘core-periphery’ organization emerged in all groups when the regional flexibility values were compared against the 95% confidence intervals of the connectional null model distributions. AG = angular gyrus; AT = anterior temporal lobe; BH = between hemisphere; HC = healthy control; IFG = inferior frontal gyrus; IFGorb = orbital inferior frontal gyrus; LH = left hemisphere; LTLE = left TLE; MAT = middle-anterior temporal lobe; MFG = middle frontal gyrus; MPT = middle-posterior temporal lobe; PT = posterior temporal lobe; RH = right hemisphere; RTLE = right TLE.

Non-parametric multivariate machine learning

We used random forest regression to explore potential non-linear associations between our functional MRI measures and behaviour performance (i.e. verbal fluency) with the whole sample. We first explored whether the information embedded in the predictors could explain the variance in the verbal fluency. To ensure robustness, we applied a leave-one-out cross validation to predict the response and tested its association with actual scores using a right-tailed Pearson correlation. Next, we ranked the predictors using variable importance to identify the most relevant using a permutation-based method (Altmann et al., 2010). Briefly, we estimated the true variable importance for each predictor for 500 repetitions. Next, we established a null distribution of variable importances for every predictor by estimating their variable importance from models trained with randomly permuted responses for 1000 times. Accordingly, for the true variable importance of each predictor from each repetition, the probability can be estimated based on its corresponding normal cumulative null distribution and considered as significant if larger than 95%. Lastly, we ranked all predictors based on the frequency with which their variable importances were found significant over the 500 repetitions (details for both steps can be found in the Supplementary material).

Statistical analysis

Statistical analyses were conducted using IBM® SPSS® v23, with alpha level set at P < 0.05 for all tests with appropriate correction for multiple comparisons.

Results

Demographical, behavioural, and clinical comparisons

The three experimental groups did not differ in age, gender, education, or handedness (Table 1). However, significant differences in both phonemic and semantic fluency were found [F’s(2,76) > 5.890, P’s < 0.004], with scores of the healthy controls significantly higher than both the left and right TLE groups (P_Bonferroni’s < 0.014, Table 1). Such group differences remained significant even after controlling for age, gender, and education [F’s(2,73) > 4.036, P’s < 0.022]. The two patient groups did not differ in verbal IQ, performance IQ, full scale IQ, age at epilepsy onset, duration of epilepsy illness, seizure focality, interictal spike type, and type of temporal pathology evidenced by presurgical MRI scans (Table 1).

Dynamic organization of the language system

We confirmed that a dynamic organization was indeed present in the language system before assessing its reconfiguration, by comparing the verb generation task data against three appropriate statistical null models (Bassett et al., 2011, 2013a) separately in each group. First, we tested the real data against a static null model, in which an identical adjacency matrix was assigned to each layer. We found in all groups, module allegiance distributions in the real data were more widespread compared to static null models (Kolmogorov-Smirnov tests, P’s = 0, k’s > 0.453), indicating the language network was indeed dynamically varying (Fig. 2, top panel). Second, we tested the real data against a nodal null model, in which we scrambled region identities by linking a node in a layer to a randomly chosen node in consecutive layers. We found the average module allegiance was significantly higher in the real data than the nodal null models within both the left and right hemispheres (t’s > 5.079, P’s < 0.001), but lower between hemispheres (t’s < −4.524, P’s < 0.001), in all groups (data × hemisphere interactions, F’s > 35.074, P’s < 0.001), suggesting that the dynamic structure of the language system depended upon the hemisphere-related identities of regions (Fig. 2, middle). Lastly, we compared the regional flexibility of the real data against the flexibility distribution of a connectional null model, in which we scrambled inter-regional connections within each layer. In all groups, the left frontal regions showed the least flexibility (forming a ‘temporal core’), while some left temporal and right frontal regions showed maximum flexibility (forming a ‘temporal periphery’) (Fig. 2, bottom). These results indicated that during an expressive language task, the language system of these patients still retained a dynamic structure, as was observed in healthy controls.

Regional communication preferences

Having demonstrated that a dynamic network structure was present during the task in all groups, we addressed our first research question, regional communication preferences, by describing the transient intra- versus inter-subsystem communication utilizing recruitment and integration. As the first study to examine these two properties in the language system, we begin by describing relevant data from a normative population (healthy controls) using repeated measures ANOVAs, with subsystems, hemispheres and conditions (task versus resting state) as within-subject factors.

Our test for recruitment revealed a significant main effect of condition [F(1,29) = 151.063, P < 0.001], with recruitment during task condition significantly higher than for resting state in all four subsystems [t(29)’s > 3.713, P’s < 0.001]. When considering each condition separately, we found a significant hemisphere × subsystem interaction only during task [F(1,29) = 53.902, P < 0.001]. Namely, the left frontal subsystem showed significantly higher recruitment compared to the right frontal subsystem [t(29) = 9.395, P < 0.001], while the left temporal subsystem showed significantly lower recruitment compared to the right temporal subsystem [t(29) = −2.086, P = 0.046, Fig. 3A, yellow bars]. Similarly, the test for integration also revealed a significant main effect of condition [F(1,29) = 2.550, P = 0.001]. Namely, integration during task condition was significantly higher than resting state for left frontal, temporal and right frontal subsystems [t(29)’s > 2.649, P’s < 0.013], but not the right temporal subsystem [t(29) = −0.443, P = 0.661]. When considering each condition separately, we observed a significant hemisphere × subsystem interaction only during task [F(1,29) = 32.343, P < 0.001]: the left frontal subsystem showed significantly lower integration compared to the right frontal subsystem [t(29) = −2.543, P = 0.017], while the left temporal subsystem showed significantly higher integration compared to the right temporal subsystem [t(29) = 4.157, P < 0.001, Fig. 3B, yellow bars]. These results indicated that distinct regional preferences for intra- versus inter-subsystem communications were evident during task performance but not at rest.

Figure 3.

Regional allegiance preference. Recruitment (A) and integration (B) estimated during task (white background) and resting conditions (grey background). Subsystem-to-subsystem integration during task condition: (C) left frontal to all three subsystems, (D) right temporal to all three subsystems. Asterisk indicates pairwise group differences, *P < 0.05, **P < 0.01, all Bonferroni corrected. Error bars reflect standard error (SE). LTLE = left TLE; RTLE = right TLE; HC = healthy control.

Altered dynamic network reconfiguration in temporal lobe epilepsy

We next tested for group differences in dynamic network reconfigurations during both task and resting state conditions (research question two), highlighting two major aspects: (i) to whom a region frequently aligns; and (ii) how frequently it changes that allegiance.

First, we explored group-level differences in regional allegiance preference through measures of recruitment and integration using repeated measures ANOVAs with subsystems, hemispheres, and conditions as within-subject factors and group as a between-subject factor. For recruitment, we found a significant condition × hemisphere × subsystem × group interaction [F(2,77) = 8.061, P = 0.001]. Specifically, significant group differences were found in the left frontal subsystem [F(2,77) = 6.298, P_FDR = 0.012] only during task: both left and right TLE patients showed significantly lower recruitment than the healthy controls (P_Bonferroni’s < 0.011, Fig. 3A). For integration, we found a similar condition × hemisphere × subsystem × group interaction [F(2,77) = 3.382, P = 0.039]. Again, significant group differences were found in the left frontal [F(2,77) = 5.395, P_FDR = 0.026] and right temporal subsystems [F(2,77) = 4.248, P_FDR = 0.036], only during task: left TLE patients showed higher integration than the healthy controls (left frontal: P_Bonferroni = 0.005; right temporal: P_Bonferroni = 0.109) and right TLE patients (left frontal: P_Bonferroni = 0.143; right temporal: P_Bonferroni = 0.019, Fig. 3B).

For the two subsystems producing significant group differences in integration during task condition, we broke down their integration values into the probability of alignment between each pair of specific subsystems. For the left frontal subsystem, group differences were only found between itself and the right temporal subsystem [F(2,77) = 5.411, P_FDR = 0.019], with the left TLE showing significantly higher integration than the healthy controls (P_Bonferroni = 0.005, Fig. 3C). For the right temporal subsystem, the exact same group difference was again found, involving itself and the left frontal subsystem (Fig. 3D), suggesting the integration differences in the left frontal and right temporal subsystems reported above are driven by their integration with one other.

Next, we explored group-level differences in the frequency of allegiance shifts through measures of flexibility and promiscuity using repeated measures ANOVAs with the same variables in the model. We found a significant condition × hemisphere × subsystem × group interaction for flexibility [F(2,77) = 3.187, P = 0.047]. Specifically, significant differences were found in the left temporal [F(2,77) = 9.482, P_FDR = 0.001] and right frontal subsystems [F(2,77) = 6.246, P_FDR = 0.006] only during task, whereas both left and right TLE patients showed significantly lower flexibility than the healthy controls (P_Bonferroni’s < 0.021, Fig. 4A). A similar condition × hemisphere × subsystem × group interaction was found to be significant for promiscuity [F(2, 77) = 3.187, P = 0.010]. Specifically, significant differences were found in the left temporal [F(2,77) = 5.685, P_FDR = 0.012] and right frontal subsystems [F(2,77) = 5.519, P_FDR = 0.012] only during task, whereas both left and right TLE patients showed significantly lower promiscuity than the healthy controls (P_Bonferroni’s < 0.027, Fig. 4B).

Figure 4.

Frequency of allegiance shifts. Flexibility (A) and promiscuity (B) estimated during task (white background) and resting conditions (grey background). Asterisk indicates pairwise group differences, *P < 0.05, **P < 0.01, ***P < 0.001, all Bonferroni corrected. Error bars reflect SE. HC = healthy control.

These findings imply both a disruption of regional allegiance preference and an abnormal transition of brain states during the task condition in TLE patients. Similar differences were found, after controlling for age, gender, and education, run as separate covariates [condition × hemisphere × subsystem × group interaction: recruitment, F’s(2,76) > 7.587, P’s < 0.001; integration, F’s(2,76) > 3.279, P’s < 0.043; flexibility, F’s(2,76) > 3.231, P’s < 0.045; promiscuity, F’s(2,76) > 4.849, P’s < 0.010]. These results were also robust against the potential residual influence of the head micromovement, selection of resolution parameters, and sliding-window strategy (Supplementary material).

Traditional static measures of verb generation task activation and laterality

We carried out group-level comparisons on the verb generation task activation at both voxelwise and laterality index levels to test their sensitivity to language deficits (research question three). We found similar activation patterns during the verb generation task for all three groups (Fig. 5A), whereas qualitatively compared to healthy controls, left TLE showed restricted while right TLE showed more extensive activation. However, after multiple comparison correction, the voxelwise approach showed no significant differences between any pair of groups (see results at lower threshold in the Supplementary material). We further tested the laterality indices of the verb generation task but found no significant difference with either mask we used, when laterality index was calculated using either a bootstrap method [F’s(2,77) < 0.325, P’s > 0.724, Fig. 5B] or mean T-values method [F’s(2,77) < 0.828, P’s > 0.441, Fig. 5C]. Lastly we checked the prevalence of ‘atypical’ language patterns based on the regional laterality indices (i.e. frontal, temporal) following Berl et al. (2014), and only identified one case per group (Supplementary material). In summary, these analyses of static imaging markers failed to capture any group differences at appropriately stringent levels of statistical testing.

Figure 5.

Summary of results from activation-based analyses. (A) Group activations of the verb generation task in left TLE (left), right TLE (middle), and healthy controls (right). (B) laterality index calculated with the bootstrap method using the frontal subsystem, temporal subsystem, and whole language system masks. (C) Laterality index (LI) calculated with mean t-values using the frontal subsystem, temporal subsystem, and whole language system masks. Laterality index values ranged between −1 and +1, where +1 indicated left-sided lateralization. No group differences were found for all three comparisons. fROI = functionally-defined region of interest; HC = healthy control.

Behaviour relevance

To extend our third question, we examined the association of dynamic (averaged across the language system) and static measures with verbal behaviour. Since neuropsychological measures of phonemic and semantic fluency were highly correlated (R₇₈ = 0.652, P = 9.924 × 10⁻¹¹), we averaged these to generate a composite verbal fluency score with greater construct validity.

During the task condition, we found integration, not recruitment (Fig. 6A), displayed a negative correlation with verbal fluency (R₇₈ = −0.330, P_FDR = 0.013, Fig. 6B). We also found significant positive correlations involving flexibility (Fig. 6C) and promiscuity (Fig. 6D) with verbal fluency (flexibility, R₇₈ = 0.355, P_FDR = 0.011; promiscuity, R₇₈ = 0.292, P_FDR = 0.025). These results stood robust after controlling for age, gender, education, or group (integration, R₇₅’s < −0.257, P’s < 0.024; flexibility, R₇₅’s > 0.280, P’s < 0.014; promiscuity, R₇₅’s > 0.245, P’s < 0.032). Interestingly, during the resting state condition, no significant associations between the dynamic properties with verbal fluency were found (|R₇₈|’s < 0.220, P_FDR’s > 0.107), indicating the association between dynamic properties and verbal fluency were specific to the task condition.

Figure 6.

Pearson correlations between the dynamic properties and verbal fluency. (A) Average recruitment did not show a significant correlation with verbal fluency. (B) Average integration showed a significant negative correlation with verbal fluency. (C) Average flexibility showed a significant positive correlation with verbal fluency. (D) Average promiscuity showed a significant positive correlation with verbal fluency. HC = healthy control.

As a comparison, we also tested the associations between laterality indices (frontal, temporal, whole masks) and verbal fluency. We found no significant correlation between verbal fluency and laterality indices calculated with either the bootstrap (|R₇₈|’s < 0.116, P > 0.312) or mean T-values method (|R₇₈|’s < 0.073, P > 0.524). Multiple regression analysis also did not show any significant association between voxelwise activation and verbal fluency (Supplementary material).

Furthermore, we implemented random forest regression to evaluate the non-linear associations between verbal fluency and the dynamic properties or laterality indices. Specifically, we used verbal fluency scores as the response variable, and investigated three sets of predictor variables: (i) the dynamic properties of each subsystem (16 in total) during the task condition; (ii) the laterality indices calculated with both bootstrap and mean T-values methods on frontal, temporal and whole masks (six in total); and (iii) the variable sets of (i) and (ii) combined. We found that verbal fluency was significantly correlated with the predictions made with dynamic properties (R₇₉ = 0.253, P = 0.013, Fig. 7A), but not those involving the laterality indices (R₇₉ = −0.174, P = 0.937, Fig. 7B). Combining two sets together did not enhance prediction (R₇₉ = 0.228, P = 0.023, Fig. 7C). Lastly, we combined the two sets of predictors together, and estimated their variable importance with a permutation method (Altmann et al., 2010) to determine which among these predictors was significantly relevant (Supplementary material). We found that integration of the left frontal subsystem was consistently identified as a significant predictor (100% level). In addition, the recruitment of the left frontal, and both the flexibility and promiscuity of the left temporal subsystems were occasionally identified as significant predictors (20 ∼ 40%). In contrast, almost all the laterality indices were never found to be significantly important (Fig. 7D). These results provided more definitive evidence of our earlier findings by showing that dynamic properties are more relevant to verbal fluency than laterality indices.

Figure 7.

Associations between the random forest predicted values and verbal fluency. (A) Predictions made with dynamic properties. (B) Predictions made with the laterality indices (LIs). (C) Predictions made with both sets of predictors. (D) Functional MRI measures were sorted by the proportion selected as a significant predictor based on 500 repetitions of the random forest model. Measures falling in the grey area were never selected as significant. HC = healthy control; R = recruitment; I = integration; F = flexibility; P = promiscuity; LF, LT, RF, RT = left frontal, left temporal, right frontal and right temporal subsystems; F-boot, T-boot, W-boot = laterality index calculated with the bootstrap method with frontal, temporal and whole system masks; F-mT, T-mT, W-mT = laterality index calculated with mean T-values with frontal, temporal and whole system masks. Prt = right-tailed P-values.

Clinical relevance

Lastly, in the two patient groups, we examined the potential associations between the dynamic properties of the language system and the aforementioned clinical characteristics presented in Table 1. Strictly, no significant correlations after FDR correction for multiple comparisons were observed. Nevertheless, trend associations are reported in the Supplementary material.

Discussion

Utilizing the emerging capabilities of dynamic network analysis tools (Bassett and Sporns, 2017), the current study provides a new perspective on the dynamic reconfiguration of the language system in both a normative and language-deficient population (i.e. TLE). In line with previous literature (Bassett et al., 2013b; Chai et al., 2016), we captured a core-periphery organization in the language system during an expressive language task. Specifically, left frontal regions centred in Broca’s area were identified as ‘cores’. Other regions, especially the left temporal and right frontal regions were identified as ‘peripheries’. More importantly, our data demonstrate that both left and right TLE patients possess reduced flexibility in these peripheral subsystems, indicating that pathologies such as TLE have a selective effect, diminishing the functionality of peripheral more than core regions, thereby clarifying the nature of the verbal fluency deficits observed in these patients. This may imply that aspects of key expressive language functions are still preserved in these patients; however, their language systems are less flexible in support regions, diminishing their ability to adapt dynamically to changing task demands.

With this core-periphery organization in mind, we adopted the concepts of ‘recruitment’ and ‘integration’ (Bassett et al., 2015), and describe the dynamic change in allegiance that occurs between specific regions in the language system for the first time. From healthy controls we learned that during task performance, regions within the left frontal ‘core’ subsystem display a higher preference for intra-subsystem communication, but a relatively low preference for inter-subsystem communication. In contrast, the regions within the left temporal and right frontal ‘periphery’ subsystems display a higher preference for inter-subsystem communication, but a relatively low preference for intra-subsystem communication. Based on the patterns of flexibility and promiscuity observed, we conclude that left frontal regions remain stable throughout the task, communicating mainly with other regions within the same subsystem. On the other hand, left temporal and right frontal regions actively shifted their allegiances, communicating with members from other subsystems.

Our patient/control comparison revealed that during the task, both the left and right TLE patients had a reduction in intra-subsystem communication within the left frontal subsystem, indicating a disturbance in the stability of this ‘core’ area. As the left middle frontal gyrus has been implicated in verbal working memory (Fedorenko et al., 2011), this reduced recruitment in TLE may also reflect some insufficiency in verbal working memory. Specifically, in left TLE instead of maintaining communication with other regions in the same subsystem, left frontal regions showed an increase in inter-subsystem communication with regions of the right temporal subsystem, an area otherwise less recruited during this task. This finding is in line with the notion that left TLE could lead to additional activations (Thivard et al., 2005) and connections (Powell et al., 2007) involving the healthier right hemisphere. The important contrast here is that our traditional activation-based measures (voxelwise level or regional laterality index) did not capture this extra right temporal involvement, highlighting that atypical regional communication manifests periodically not consistently in this process, and, therefore, are ‘invisible’ and missed by traditional ‘boxcar’ designs, which capture only constant activations.

As noted, our findings also provide a direct comparison of the dynamic reconfiguration of the language system during both task and resting state conditions. It is well known that the human language system retains functional organization during both task (Fedorenko et al., 2011; Blank et al., 2014; Doucet et al., 2017) and resting state conditions (Tomasi and Volkow, 2012; Muller and Meyer, 2014; Doucet et al., 2017). A previous study reported that the language system is more flexible at rest (Chai et al., 2016), a feature observed across all three groups in our study. The reduction of flexibility, as well as the increase in both intra- and inter-subsystem communication we observed during the task condition may reflect functional regulation of inter-regional communications to regions most capable to meet the current cognitive goal. In contrast, such evidence was mainly absent in the resting state condition, as were reliable group differences, indicating that pathology is more likely to interfere with language system dynamics during phasic linguistic operations.

Such task specificity was also evident in the analyses with behavioural variables such as verbal fluency. Flexibility in brain organization can bring behavioural advantages to motor learning (Bassett et al., 2011) and executive function (Braun et al., 2015). Our results show that flexibility in the language system is positively associated with verbal fluency output and skill. In addition, our probability-based measures of communication preference provide insight into this association, showing that participants with lower verbal fluency may engage in unusual amounts of inter-subsystem communication (i.e. higher integration). Evidence that a well-trained brain relies less on cross-subsystem integration during task performance (Bassett et al., 2015) suggests the increased integration seen in left TLE reflects either an incipient but flawed process of compensation that has not yet come to benefit performance, or, a transient, deleterious interaction that diminishes performance (see within-patient correlational analysis in the Supplementary material). Specifically, the integration of the left frontal subsystem was particularly important in the prediction model. Consistent with this subsystem’s role as a core to verbal fluency, when epilepsy patients connected other regions to their language network (i.e. right temporal regions), those regions were more likely to interact with the left frontal subsystem. In contrast, as a measure of intra-subsystem communication, recruitment is less sensitive to the involvement of additional regions (or lack thereof), thereby explaining its lack of association with verbal fluency performance. Moreover, we showed that both frequency-based and probability-based dynamic properties outperformed laterality indices in the prediction of the verbal fluency. These results raise the possibility that dynamic measures of language organization may be an advanced alternative to the traditional activation-based measures commonly used in clinical settings.

We acknowledge that both left and right TLE patients presented with similar levels of verbal fluency deficit compared to healthy controls. Though counterintuitive, right TLE actually compromises verbal fluency quite commonly (McDonald et al., 2008; Metternich et al., 2014). For example, Metternich et al. (2014) reported that both left and right TLE patients were impaired on both semantic and phonemic verbal fluency (Metternich et al., 2014). In line with the shared behavioural deficits, both our right and left TLE patients showed similar reductions in flexibility/promiscuity in the left temporal and right frontal ‘periphery’ subsystems, along with reduced recruitment in the left frontal ‘core’ subsystem during task performance. Thus, our data implies a common mechanism may underlie the verbal fluency deficits in these two distinct patient populations. For example, antiepileptic drugs (AEDs) could reduce cognitive speed (Ortinski and Meador, 2004), in turn diminishing rapid verbal fluency in both groups (Mula et al., 2003; Blum et al., 2006). Alternatively, the right-sided pathology may compromise right hemispheric language processes (Chiarello et al., 2006; Hickok and Poeppel, 2007), and the spread of seizure activity from right-to-left temporal (or the presence of bilateral interictal discharges) could compromise left hemispheric language processes as well (Badawy et al., 2012). Interestingly, Metternich et al. (2014) did find that left compared to right TLE patients were slightly more impaired, a finding not observed in the present study. This could be attributed to our exclusion of patients with low verbal IQ, which likely reduced the presence of more severe language deficits, particularly in our left TLE group. More importantly, our data indicates that the language deficits in the two patient groups were not essentially equivalent as suggested by part of our data, since left TLE patients have, in particular, a higher level of integration, a feature not seen in right TLE.

Several methodological considerations are pertinent to this study. First, although identical parameters were used in both task and resting state scans, the length of 5 min was relatively short. In this light, it is important to note that the dynamic organization of the language system was originally discovered with tasks lasting from only 4 to 6 min (Chai et al., 2016). Second, the selection of window length and sliding steps could still potentially influence our measures of network dynamics (Telesford et al., 2016). Therefore, we retested our main analyses with (i) more windows; and (ii) a larger window length. Importantly, these changes to the analytic pipeline produced matched findings (Supplementary material). Third, as a covert speech task, no individual responses to the verb generation task were recorded during performance. This task design limited our ability to link the dynamics of the language system with real-time performance profiles and even effort or strategy differences. Fourth, all IQ measures were collected for patients but not available for healthy controls. Hence, potential IQ differences between patients and healthy controls could well have existed. With such limitations in mind, we tested the associations between the IQ measures and the dynamic properties within the patient groups, and found no associations relevant to the major task-related effects we reported. Fifth, the group-constrained mask used in this study generally provided good coverage of regions potentially involved in both ipsilateral and contralateral patterns of reorganization reported in TLE (Bell et al., 2002; Thivard et al., 2005; Gaillard et al., 2007; Mbwana et al., 2009; Tracy et al., 2009). Though rare, we acknowledge some patients may have reorganized elsewhere. Accordingly, a whole brain analysis may be warranted to account for all patterns of reorganization. Lastly, we should note that AEDs can influence the blood oxygen level-dependent signal (Jansen et al., 2006; Haneef et al., 2015; Wandschneider et al., 2017), potentially influencing patient/control differences. Unfortunately, AED regimen heterogeneity (type, dosage, number of AEDs) prevented further testing of these effects. Given the mixed distribution of medications across the patient groups (Supplementary Table 5), it seems unlikely that any left/right TLE difference could be traced to a specific medication. Moreover, it is important to keep in mind that the influence of medication would likely cancel out in any task versus rest comparison.

In summary, we provide evidence that a dynamic organization of the language system exists during linguistic operations, going beyond the dynamics present at rest. We demonstrate how abnormal interactions in both core and periphery subsystems reveal a lack of adaptive transient communications in TLE patients during an expressive language task, thereby providing a novel explanation of the language deficits induced by such types of neurological diseases. Our results introduce a new perspective and characterization of language network dynamics, increasing our understanding of language dysfunction and maladaptive seizure-driven plasticity in epilepsy.

Glossary

Abbreviation

TLE

temporal lobe epilepsy