The organization of functional neurocognitive networks in focal epilepsy correlates with domain‐specific cognitive performance

Understanding and diagnosing cognitive impairment in epilepsy remains a prominent challenge. New etiological models suggest that cognitive difficulties might not be directly linked to seizure activity, but are rather a manifestation of a broader brain pathology. Consequently, treating seizures is not sufficient to alleviate cognitive symptoms, highlighting the need for novel diagnostic tools. Here, we investigated whether the organization of three intrinsic, resting‐state functional connectivity networks was correlated with domain‐specific cognitive test performance. Using individualized EEG source reconstruction and graph theory, we examined the association between network small worldness and cognitive test performance in 23 patients with focal epilepsy and 17 healthy controls, who underwent a series of standardized pencil‐and‐paper and digital cognitive tests. We observed that the specific networks robustly correlated with test performance in distinct cognitive domains. Specifically, correlations were evident between the default mode network and memory in patients, the central‐executive network and executive functioning in controls, and the salience network and social cognition in both groups. Interestingly, the correlations were evident in both groups, but in different domains, suggesting an alteration in these functional neurocognitive networks in focal epilepsy. The present findings highlight the potential clinical relevance of functional brain network dysfunction in cognitive impairment.


Significance
Although epilepsy is most commonly associated with seizures, the disease is more often than not accompanied by non-seizure symptoms, such as cognitive deficits. Interestingly, even in focal epilepsies, the cognitive challenges often extend beyond the seizure onset brain area's classical cognitive expression, suggesting a widereaching perturbation of peripheral networks. In the present study, we demonstrate, using source-space electroencephalography, functional connectivity, and graph theory concepts, that the organization of neurocognitive brain networks is associated with domain-specific neuropsychological test performance. These findings might hold clinical relevance with regard to the identification, monitoring, and treatment of cognitive symptoms in epilepsy.

| INTRODUC TI ON
Fundamental physiological processes in the human brain, such as those supporting control of movement (King et al., 2018), autonomic functions (Fan et al., 2012), and cognition (Shine et al., 2019), are sustained and regulated by complex interactions both within and between functional networks. Thus, network disruptions are likely to amount to an important role in mediating physiological dysfunction and associated behavioral manifestations, such as cognitive impairment (Stam, 2014;Uhlhaas & Singer, 2006). Importantly, in contrast to the traditional modular paradigm, the hypothesis that cognition arises from the activity of large-scale networks (Bressler & Menon, 2010;Mesulam, 1990) can account for the presence of cognitive dysfunction beyond local structural pathology (Tailby et al., 2018). Thus, the impact of functional brain networks on cognitive dysfunction has attained increased focus, both in nonclinical (Douw et al., 2011;Langer et al., 2012) and clinical groups (Hassan et al., 2017;Vlooswijk et al., 2011).
Epilepsy is considered the epitome of brain network dysfunction (Kramer & Cash, 2012), putatively caused by hypersynchronous neuronal network activity (Engel et al., 2013). During the past decade, the functional connectivity (Bastos & Schoffelen, 2015) and graph theory (Behrens & Sporns, 2012) frameworks have successfully been applied to delineate network alterations in focal epilepsy (FE) using electro-and magnetoencephalographic (EEG/ MEG) methods (Horstmann et al., 2010;Niso et al., 2015;Vecchio et al., 2015). Importantly, concurrent with functional network changes, patients with epilepsy are often burdened with cognitive impairment (Henning et al., 2019;Lin et al., 2012). Intriguingly, such impairment is increasingly understood as a distinct manifestation of underlying brain network pathology, rather than being directly linked to seizure activity . This notion lines up with evidence that FE often manifests with widespread cognitive impairment extending beyond the seizure onset area's classical cognitive expression (Oyegbile et al., 2004). Furthermore, network disruptions, manifested as interictal epileptiform discharges in EEG peripheral to the seizure onset zone, are associated with impaired short-term memory encoding (Ung et al., 2017). Also, cognitive dysfunction in FE patients may change over time, possibly reflecting the dynamic nature of brain networks and their disturbance, and may even persist after successful seizure remission | 2671 HATLESTAD-HALL ET AL. non-invasive highly specialized experimental technology, such as functional magnetic resonance imaging (fMRI), research has begun delineating distinct intrinsic neurocognitive brain networks (Deco et al., 2011;Uddin et al., 2019). Consistently reported among such functionally and spatially distinct networks are the default mode network (DMN), the central executive network (CEN), and the salience network (SN). The DMN is anchored in the medial prefrontal and posterior cingulate cortices, and the posterior extent of the inferior parietal lobule (Andrews-Hanna et al., 2010;Buckner et al., 2008), whereas the CEN relies on the lateral prefrontal cortex and the anterior inferior parietal lobule reaching into the intraparietal sulcus, and the SN on anterior insular and anterior middle cingulate cortices (Seeley et al., 2007;Uddin, 2015;Yeo et al., 2011). Together, these have been suggested to comprise a triple network model of human cognition (Menon, 2011;Seeley et al., 2007;Uddin, 2015), in which the DMN and CEN maintain antagonistic activity mediated by the SN (Chand et al., 2017;Goulden et al., 2014). These network constellations may sustain a broad spectrum of cognitive functions (Uddin et al., 2019), where the DMN is most closely associated with internally directed and self-generated cognitive processes (Andrews-Hanna et al., 2014), the CEN with higher order executive functioning, and the SN with an identification of salient information with regard to current goal states (Seeley et al., 2007). Cognitive deterioration is accompanied by altered connectivity within this large-scale system, both in pathological (Agosta et al., 2012;Damoiseaux et al., 2012) and normal aging (Chand et al., 2017). In FE, such network reorganization could reflect a compensatory mechanism to sustain an adequate level of cognitive functioning, possibly at the expense of brain health and longevity Sen et al., 2018).
The investigation of brain networks is methodologically challenging, albeit fMRI has made significant contributions in characterizing spatial properties of functional brain networks (Uddin et al., 2019).
However, to expand on the temporal dynamic nature of brain networks (Fox & Raichle, 2007), the millisecond timescales provided by EEG and MEG are necessary for insights into oscillatory activity in specific frequency bands associated with differential neuronal processing of information (Rossini et al., 2019). Neuronal oscillations in different frequency bands have consistently been shown to perform differential roles in cognition (for a review, see Lopes da Silva, 2013).

| Participants
Twenty-three patients diagnosed with uni-or bilateral FE (15 females, age 54.7 ± 5.9 years) were recruited to the study from neurological outpatient clinics in the Oslo region, Norway, in connection with routine follow-up visits. In addition, 17 age-matched healthy control (HC) subjects (12 females, age 55.9 ± 6.7 years) participated. The HC participants were recruited from the FE patients' social networks, providing controls with similar socioeconomic background to the patients.
All FE subjects had used the same ASM for at least 6 months before they participated in the study. The patients were all considered to be in a chronic phase of epilepsy (duration of epilepsy 24.4 ± 13.7 years).

| Cognitive testing
All participants underwent a standardized cognitive assessment constituent of both paper-and-pencil and digital tests.

| EEG acquisition and electrode localization
EEG was recorded with a BioSemi 128-channel system with a sampling rate of 2,048 Hz during a task-free (i.e., resting-state) paradigm.
The participant was comfortably seated in a chair, resting with his or her eyes closed, but awake, for 4 min. To minimize between-subject variation regarding task comprehension, instructions were given in written form on a computer monitor, as recommended by van

| Network analysis
The network analysis was conducted with functions available from the Brain Connectivity Toolbox (BCT; ver. 2019-03-03, RRID:SCR_004841; Rubinov & Sporns, 2010) and with in-house MATLAB code (available upon request). Network nodes were defined as AAL regions and network edges as the PLV estimate between regions. The three neurocognitive networks were composed of distinct sets of nodes, defined in approximate accordance with current anatomical definitions of these networks (Uddin et al., 2019).
The selection of networks for analysis was a priori defined, based on the relatively strong consensus regarding their anatomical definitions and their consistently described relevance to cognition (for a review, see Uddin et al., 2019). The mapping of the networks to AAL regions is presented in Figure 1 and Table 2.
In terms of network analysis, the functional connectivity matrices were weighted (PLV) and undirected. The matrices were analyzed under a fixed density threshold of 65%, constrained by the requirement of the matrices to remain fully connected. As no basis exists on which to define an ecologically valid threshold level (Fornito et al., 2012;van Wijk et al., 2010), the 65% threshold was selected as it was previously demonstrated to maximize the group difference in graph metrics in FE (Hatlestad-Hall et al., 2021). To implement this, we first computed the minimum spanning tree (MST; Stam et al., 2014;Tewarie et al., 2015) of the inverse functional connectivity matrix, and then added edges to the MST backbone incrementally by descending weight order until the density threshold was reached. For each functional connectivity matrix, 100 rewired null models with preserved weight, degree, and strength distributions were generated from the dense matrix (BCT: null_model_und_sign; Rubinov & Sporns, 2011). These null models were processed identical to the empirical network, and the mean global network metrics calculated from them were used to normalize the corresponding metric for the empirical network.
We calculated the small world index, which was defined as the ratio between the network's normalized clustering coefficient and normalized characteristic path length (Humphries & Gurney, 2008;Watts & Strogatz, 1998): where CC is the clustering coefficient and CPL is the characteristic path length, and the subscripts obs and rand denote the metric of the observed (empirical) network and the mean metric calculated from the randomly rewired networks, respectively. The terms SW index and SW-ness are used interchangeably.
Differences in the overall level of functional connectivity between subjects or groups can generate spurious differences in network SW = CC obs ∕ CC rand CPL obs ∕CPL rand metrics (van den Heuvel et al., 2017). In this context, overall functional connectivity was defined as the mean PLV in the individual connectivity matrix, discarding intraregional estimates, and calculated separately for each frequency band. We explored possible differences in overall functional connectivity for each of the three neurocognitive networks in a post hoc group comparison. As these values did not differ between groups, the potential effect of overall functional connectivity differences between groups was disregarded.

| Cognitive test performance and self-reported deficits
The complete list of group-level performance on the cognitive tests is presented in Among the cognitive test scores, only the social cognition task (ERT number of hits) displayed a trend-significant negative correlation with the duration of epilepsy (r = −0.479, p = 0.021, q = 0.378), reflecting lower task performance with increasing time since the debut of epilepsy.

| Neurocognitive network small worldness
Interestingly, we observed subtle group differences in SW indices in all three neurocognitive networks across several frequency bands

| Correlation of network small worldness with cognitive test performance
In both groups, positive and negative robust correlations between the levels of the SW index and cognitive test performance were observed in all neurocognitive networks and among all frequency bands.

| Default mode network
In the DMN (Figure 4)

| Central-executive network
In the CEN (Figure 5), robust partial correlations were evident between the SW indices and cognitive test scores in healthy controls, but not in patients. In particular, the theta band SW index displayed the correlations of moderate magnitude with multiple test scores in the domains of working memory and executive functioning, specifically in the DS sequencing condition (r = 0.730, p = 0.001, q = 0.005) and the CWIT inhibition/switching task (r = −0.641, p = 0.010, q = 0.020). Thus, better performance on tests of working memory and executive functioning was associated with higher levels of SW in the theta frequency range of the CEN, but only in the healthy controls. The association between the CWIT inhibition/switching and the SW index was also evident in

Overall cognitive performance
Sum of quartile points 44.0 ± 9.9 46.9 ± 9.5 −0.30 Note: Group means and standard deviations. In the column "Effect size," blue font color indicates better performance by the healthy controls compared to the patient group, and vice versa for red font color.

F I G U R E 2
Network-wise SW index comparison between groups. An asterisk marks trend-significant group differences (p < 0.05); however, none of the observed group differences survived FDR adjustment. The violins represent a kernel density estimate. Box plot components: White dot = median; vertical thin gray bar = upper/lower adjacent value; vertical thicker gray bar = first/third quartile; colored dots = individual observations

| Salience network
In the SN (Figure 6), associations between test scores and SW indices were observed for both patients and controls, and were most prominent in the alpha band. In the social cognition task, the patients' scores were positively correlated with task performance (r = 0.456, p = 0.038, q = 0.075), whereas the controls' scores were negatively

| D ISCUSS I ON
In recent years, EEG-based functional network abnormalities have been consistently reported in epilepsy (Horstmann et al., 2010;Quraan et al., 2013;Vecchio et al., 2015), and such network-level dysfunctions are increasingly considered crucial mechanisms in sustaining non-seizure symptoms, such as cognitive impairment (Kramer & Cash, 2012;Uhlhaas & Singer, 2006). However, to our F I G U R E 3 Correlation between the SW index and time living with epilepsy. A single asterisk marks trend-significant correlations (p < 0.05). A double asterisk marks significant correlations after FDR adjustment (q < 0.1) knowledge, their association with distinct measures of cognitive test performance in epilepsy has not been highlighted before. Here, we investigated whether functional network organization, as reflected in source-level EEG and quantified as the SW index, was associated with performance on a range of cognitive tests. Specifically, we targeted three well-established neurocognitive networks, that is, the DMN, CEN, and SN, and examined to what degree their SW-ness would predict domain-specific cognitive performance in middleaged patients with chronic FE and in healthy controls. Interestingly, although we only observed minor group differences in cognitive test performance and SW-ness separately, we were able to demonstrate that associations between them depended on the examined neurocognitive network, the EEG frequency band, and the presence or absence of epilepsy. Our analyses revealed that SW-ness of the DMN was associated with memory performance in patients, whereas the CEN organization was closely related to working memory and executive functioning in controls. For both groups, a significant correlation between the SW index of the SN and performance on a social cognition task was evident. Interestingly, for the latter network specifically, SW-ness across several frequency bands varied in relation to the patients' time living with epilepsy. In the following, we discuss how these findings relate to the existing literature, and the clinical relevance of the relationship between neurocognitive network organization and cognitive functioning in epilepsy.

| Neurocognitive networks in epilepsy
The DMN supports a range of cognitive functions, including memory (for a review, see Uddin et al., 2019). Altered functional connectivity in the DMN or its constituent regions has been shown to predict memory deficits in neurological disease such as amnestic mild cognitive impairment (Dunn et al., 2014;Wang et al., 2013). Importantly, DMN connectivity aberrancies, as measured by fMRI, have been implicated in the sustenance of epilepsy symptoms with regard to seizure propagation (Ofer et al., 2019),

F I G U R E 4
Partial correlation coefficients between SW-ness of the default mode network and cognitive test performance. Significant coefficients are marked with an asterisk. Significant coefficients which survived FDR correction are framed memory , and executive dysfunction (Zhou et al., 2019). In keeping with this, we demonstrate that increased SW-ness of the DMN in the theta band is associated with poorer memory performance in patients. The fact that this finding is specific to the theta band is interesting, as theta oscillations are closely related to memory processes (Axmacher et al., 2008;Raghavachari et al., 2006;Sauseng et al., 2009). Furthermore, several studies suggest the global network alterations in FE to be relatively prominent in the theta band (Horstmann et al., 2010;Quraan et al., 2013). This is further supported by the observation in the current study that, among the investigated frequency bands, theta showed the most robust group difference in DMN SW-ness (Figure 2, top-left), with patients displaying the larger SW indices. Interestingly, the association between memory and the DMN was not found in the healthy controls, suggesting that the SW-ness of the network does not index global memory integrity, but rather the degree of deterioration, a notion further reinforced by a significant interaction effect between SW-ness and group.
Consistent with the established role of the CEN as a neural substrate for executive functioning (Markett et al., 2014;Seeley et al., 2007), here the network's SW-ness in the theta band robustly predicted performance on tasks relying on mental flexibility and working memory.
However, in contrast to the DMN, the CEN-specific associations were only evident for healthy controls, and not for patients. One potential interpretation of this discrepancy may be that the CEN is adversely affected in epilepsy, leading to dysfunctional regulation of the network in response to current goal states. Indeed, in a study of newly diagnosed FE patients, network functional connectivity alterations were found selectively in the CEN (Alonazi et al., 2019), suggesting that this network might be particularly susceptible to disruption by epileptic neuronal activity. However, from our data, this potential disruption does not appear to be reflected in the network's SW-ness, which was similar across the groups. Conversely, this similarity may suggest the presence of compensatory mechanisms in the patients, reducing the discrepancy of the network's SW-ness and cognitive performance relative to the healthy controls. In this scenario, such compensatory F I G U R E 5 Partial correlation coefficients between SW-ness of the central-executive network and cognitive test performance. Significant coefficients are marked with an asterisk. Significant coefficients which survived FDR correction are framed measures may suppress the direct relationship between network SWness and executive functioning observed in the controls. Behaviorally, executive dysfunction is prevalent in several syndromes of epilepsy, such as FE with frontal or temporal origin (Hermann et al., 1997;Oyegbile et al., 2004). Moreover, in a study employing graph theory on cognitive test scores in a large cohort of temporal lobe epilepsy patients, a test of mental flexibility and rapid set switching (Trail Making Test B) comprises a central hub in the network of tests (Kellermann et al., 2016), implicating that other test results are relatively likely to be correlated with it. Following this notion, it is possible that prevalent cognitive issues, such as memory impairment (Taylor et al., 2010), are caused either directly by disruption of the CEN, or indirectly via behavioral executive dysfunction as a mediator. We encourage future research to investigate additional characteristics of the CEN in order to delineate its role in cognitive dysfunction in epilepsy.
The SN is attributed with the task of identifying the most relevant among the continuous stream of internal and external inputs to the brain, including emotional information (Kennedy & Adolphs, 2012;Menon, 2011;Seeley et al., 2007). In line with this, we observed in our data a robust association between the patients' SW-ness of the SN in the alpha and beta bands, and their ability to identify emotions from rapid presentation of human faces (the ERT). The higher the SW-ness, the more correct responses were made. Moreover, in both these frequency bands, patients displayed a trend toward lower SW-ness of the SN relative to the controls. These observations, in light of the firmly established evidence that patients with epilepsy, predominantly with temporal origin, suffer impairment of social cognition and facial recognition (Bora & Meletti, 2016;Edwards et al., 2017;Monti & Meletti, 2015), suggest that an alteration in the SN's functional organization might amount to a crucial mechanism in sustaining these deficits in epilepsy. Furthermore, the regulation of alpha and beta oscillations has been found to be involved in face perception (Popov et al., 2012(Popov et al., , 2013. However, in our data, the alpha band SW-ness of the SN was also predictive of processing speed, inhibitory control, and overall cognitive performance in the patients, suggesting that SN organization might be implied in a broader range F I G U R E 6 Partial correlation coefficients between SW-ness of the salience network and cognitive test performance. Significant coefficients are marked with an asterisk. Significant coefficients which survived FDR correction are framed of cognitive domains. Indeed, considering the hypothesis of accelerated cognitive aging, which postulates that cognitive decline in epilepsy follows a normal trajectory, but at an increased rate (Breuer et al., 2016(Breuer et al., , 2017, the SN might be among several intrinsic brain networks which, in a compensatory effort, alters their intra-and internetwork organization in response to pathological processes (Bernas et al., 2020). Our finding of an association between SW-ness of the SN and the patients' number of years living with epilepsy might lend further support to this notion.

| Clinical relevance
Cognitive impairment remains a prominent personal and diagnostic challenge in epilepsy, with approximately 70% of patients reporting subjective memory problems, independent of age, medication, and seizure status (Henning et al., 2019). In objective measures, about half of newly diagnosed patients over the age of 15 present with abnormal performance in the memory and psychomotor speed domains, regardless of seizure type and epileptogenic brain area (Taylor et al., 2010). Currently, seizure control with ASM is the prioritized aim in epilepsy treatment (Schmidt & Schachter, 2014;Sørensen & Kokaia, 2013). However, new etiological models increasingly consider the possibility that cognitive dysfunction arises as an independent manifestation of a complex underlying pathology that also provokes seizures (Bjørke et al., 2021;Hermann et al., 2008;Lin et al., 2012). Thus, in many clinical cases, achieving freedom from seizures will not be sufficient for treating cognitive dysfunction  changes to the brain's hub organization (Haneef et al., 2015), and that topiramate affects network connectivity to a larger extent compared to other medications (van Veenendaal et al., 2017).
Furthermore, the theoretical framework for the clinical application of graph methodology remains sparse. Graph theory has primarily developed describing theoretical and synthetic networks, and the translation of its concepts to clinically significant characterizations of complex biological systems, such as the human brain, is largely rudimentary (Douw et al., 2019;Papo et al., 2014). For example, the concept of increased SW-ness is, as a rule, associated with higher network efficiency. It reflects the system's ability to achieve integrated functioning through facilitation of long-range signal transmission between functionally specialized modules of short communication paths (Bassett & Bullmore, 2017;Bertolero et al., 2015). However, in the context of functional brain connectivity and clinical relevance, this notion is one of controversy (Papo et al., 2016). Provided the complexity of the human brain, it is perhaps more likely that SWness constitutes one relevant parameter of network organization, and that additional metrics are required for a more comprehensive characterization of brain networks (Hallquist & Hillary, 2019).

| Limitations
For the present study, some limitations should be outlined. First, compared to fMRI and MEG, EEG has inherent shortcomings regarding spatial resolution. Here, this limitation was mitigated by employing high-quality source reconstruction, based on individual anatomy.
Nevertheless, the investigated resting-state networks were necessarily approximated by predefined composites of AAL regions. The selection of network regions was guided by an updated consensus definition of the networks (Uddin et al., 2019). Furthermore, electrophysiologically derived functional connectivity is subject to the effect of source leakage. If the leakage varies across groups, differences can arise in the connectivity matrices, and by extension, the graph metrics. While some functional connectivity measures eliminate the effects of leakage by removing zero-lag synchronization (Bruña et al., 2018;Ewald et al., 2012), these show low testretest reliability, in contrast to PLV, which was used in the present study (Colclough et al., 2016;Garcés et al., 2016). Here, the effect of source leakage might have impacted the direct group comparisons of the SW index; however, its effect on the observed SW cognition correlations is presumed minimal. Finally, the current data are based on relatively small samples of patients with FE and healthy controls.
The results should therefore see replication in larger samples. Also, due to the low number of participants, the patients were not stratified with regard to clinical factors such as epilepsy focus, ASM type and dosage, and seizure frequency, which have previously been reported to affect network metrics (Haneef et al., 2015;van Dellen et al., 2009). It is important to note that the present patient sample is relatively homogeneously composed, reflecting people in middle adulthood and in a chronic phase of epilepsy.

| Conclusion
Here, we have demonstrated that graph characteristics of functional neurocognitive networks derived from source-level EEG display frequency-and domain-specific correlations with performance on cognitive tests. These findings highlight the role of functional brain network dysfunction in cognitive impairment, an important issue reiterated by revised etiological models suggesting that impaired cognition and epilepsy may arise independently from a shared pathology.

DECL AR ATION OF TR ANS PAREN C Y
The authors, reviewers and editors affirm that in accordance to the policies set by the Journal of Neuroscience Research, this manuscript presents an accurate and transparent account of the study being reported and that all critical details describing the methods and results are present.

ACK N OWLED G M ENTS
The authors thank Helle Stangeland for her contributions in the early phases of this project, and the participants of the study. Gratitude is also extended to the AI-Mind task force for providing a creative and future-oriented research environment.

PA RTI CI PA NT I N FO R M E D CO N S E NT
Following the Declaration of Helsinki, all participants provided informed written consent.

E TH I C A L A PPROVA L
Ethical approval for the study was granted by the Regional Committees for Medical Research Ethics -South-East Norway.

CO M PE TI N G I NTER E S TS
The authors report no competing interests.

AUTH O R CO NTR I B UTI O N S
All the authors had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

PE E R R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1002/jnr.24896.