Between 2017 and 2020, we recruited 20 children with FLE (n = 9) or TLE (n = 11) and 29 healthy controls. All participants were French speakers, aged between 6 and 18 years, born at term, and had no history of previous neurosurgery, metabolic disorder, traumatic brain injury or intellectual disability. Medical files of patients with epilepsy followed at the Neurology Division of the Mother and Child University Hospital Center Sainte-Justine (CHUSJ, Montreal, QC, Canada) were screened. For children with epilepsy, exclusion criteria were: multifocal epilepsy, significant structural abnormalities with mass effect, genetic syndrome known to significantly affect neurodevelopment, autism spectrum disorder, multiple ( ≥ 3) comorbidities (e.g., attention deficit hyperactivity disorder, dyslexia, dyspraxia, etc.). Patients and their caregivers were approached by their neurologist (BO, AL, PD, PM, ER) and subsequently contacted by the research coordinator (AMH). The study’s protocol (2014-664, 3,876) was approved by the CHUSJ’s research ethics board. All parents and children provided written informed consent. Descriptive statistics of the sample’s sociodemographic and clinical characteristics are presented in Table 1.

Participation in the study included (1) a combined functional near-infrared spectroscopy (fNIRS) and electroencephalography (EEG) data acquisition at rest and during a receptive language task, and (2) a neuropsychological assessment. Both are described in the following sections.

The raw EEG signal was filtered (bandpass range: 0.3–35 Hz, notch: 60 Hz) and screened for potential neurophysiological abnormalities by a clinical epileptologist (BO) (BrainVision Analyzer, Version 2.2, Brain Products GmbH, Gilching, Germany). This aimed to identify sub-clinical seizures in the participants with epilepsy and assure the absence of pathological findings within the control group. fNIRS data recorded simultaneously to a subclinical seizure identified on the EEG were excluded from further analyses to avoid contamination of the epileptogenic networks when mapping the language brain networks.

Raw fNIRS data was imported into the LIONirs toolbox (Tremblay et al., 2022), which is embedded in SPM12 (Statistical Parametric Mapping) and operates in MATLAB (The MathWorks, Inc., 2019b, Natick, MA, United States). Signal coherence among channels for the peak frequency commonly related to the cardiac beat (0.8–2.3 Hz) allowed to exclude channels with poor optode-to-scalp coupling (Yücel et al., 2021; Fourdain et al., 2023). Intervals with aberrant signal variation based on moving average criteria were identified using an automatic algorithm. An artifact was identified if (a) changes in light intensity within a time interval of 6 s exceeded three standard deviations, (b) the noise lasted for more than 3 s, and (c) at least 5% of channels were affected. Temporally correlated ( r > 0.8 ) channels within the same interval were considered as artifacts and intervals less than 2 s apart were considered as one event. A visual inspection was also performed, and minor manual adjustments were applied. Subsequently, parallel factor analysis (PARAFAC), i.e., a multidimensional decomposition method (Harshman, 1970; Bro, 1997) which has recently been validated for fNIRS data (Harshman, 1970; Bro, 1997; Hüsser et al., 2022), was used to correct the signal during relatively isolated artifact intervals. Noisy periods that lasted for a relatively long interval and where no clear signature could be extracted were excluded. Optical density of the signal between uncorrected noisy intervals, was transformed into delta optical density to normalize the signal. A fourth-order zero-phase Butterworth band-pass filter with cut-off frequencies 0.001–0.5 was applied to reduce confounds of systemic physiology related to respiratory or cardiac pulsation, while keeping the low frequencies presumably related to the hemodynamic response (Pinti et al., 2019; Santosa et al., 2020). Optical density was then transformed into relative oxy−/deoxyhemoglobin (HbO/HbR) concentration changes using the modified Beer–Lambert law (Kocsis et al., 2006) with age-appropriate differential pathlength factors (DPF; Scholkmann and Wolf, 2013). The averaged signal of all channels was regressed from every channel, a procedure which has been shown to successfully remove remaining physiologic confounds appearing in frequency bands that are in the range of the hemodynamic signal and thus hard to filter (Saager and Berger, 2008). See Figure 2 for an overview of the fNIRS preprocessing pipeline.

FC and network analyses were conducted on HbO matrices only, which is considered a sensitive indicator of regional oxygenation differences in response to cerebral activation (Homae et al., 2007) and has a good concordance with the fMRI BOLD signal (Sato et al., 2013; Tung et al., 2021). Circular bootstrapping was used to extract 200 random 60-s samples from each data set (resting-state and task) of each participant (Bellec et al., 2010). For each valid sample, we computed cross-correlations, i.e., pairwise Pearson’s correlations ( ρ ) between the time course of all measurement channels ( y i , y j = 1 , … 50 ). The symmetric matrix of each subject ( A n ) included the averaged ρ i j of all possible pairs (k = 1,225) over all samples. Correlation coefficients underwent Fisher’s transformation to obtain normally distributed values ( z i j ). Developmental effects were removed by estimating and regressing the contribution of age from A n with a general linear model (GLM): A n = β 0 + β 1 age + ϵ , resulting in B n . The hemodynamic signal related to receptive language processing was isolated by subtracting the FC matrix computed using the resting-state data from the FC matrix computed using data gathered during the story listening task ( B n task − B n resting − state ). The subtraction allowed to control for inter-ictal epileptic or default mode network activity and yielded in a residual matrix ( R n ) specific to language comprehension.

The brain networks were characterized with common graph theoretical measures of local and global processing efficiency. Segregation is the network’s tendency to build local clusters and refers to local information processing. It is measured with: (1) the clustering coefficient ( γ ), which is the level a node’s neighbors are directly connected to each other and describes the clustering density around each node; and (2) the local efficiency (E_local) index, which considers direct and indirect paths to describe a node’s integration among its neighbors (Watts and Strogatz, 1998; Latora and Marchiori, 2001, 2003; Newman, 2003; Achard and Bullmore, 2007; Fornito et al., 2016; Sporns, 2018). Integration describes a network’s ability to efficiently integrate information from distinct parts and refers to global information processing. It is commonly assessed using: (1) the characteristic’s path length (CPL, λ ), which quantifies the averaged shortest path between all possible nodes - a lower λ implies more rapid and efficient information processing; and (2) the global efficiency (E_global) index, which also refers to the network’s global information exchange and is the reciprocal of the λ ( 1 λ ) (Latora and Marchiori, 2001; Fornito et al., 2016). Additionally, the small-world index ( σ ), which is a network characteristic that considers the normalized balance of a network’s segregation and integration σ = γ / λ (Watts and Strogatz, 1998), was considered. A network with small-world properties is characterized by a good balance of global and local processing efficiency, which is inherent for a healthy brain network architecture since early infancy and throughout adulthood (Fransson et al., 2011; Fornito et al., 2016; Liao et al., 2017).

Network metrics ( γ , E_local, λ , E_global, σ ) were computed with in-house Matlab scripts incorporating functions from The Brain Connectivity Toolbox (Rubinov and Sporns, 2010). In the graph theoretical framework, each measurement channel ( i , j ) of the fNIRS data represents a node and every correlation pair ( ρ i j ) is an edge. Individual connectivity matrices ( B n task ) computed from the language task signal only that were corrected for developmental effects, were transformed into undirected weighted graphs ( W n ) using a range of absolute correlation thresholds ( τ ). The ideal range is commonly achieved when a fully connected network, that is each node is at least connected with one other node (degree centrality [K]≥ 1), is accompanied by a small-world topology ( γ > 1 , λ > 1 , σ > 1 or in small networks when γ > γ _rand) (Montoya and Solé, 2002; Bassett et al., 2008; Rubinov and Sporns, 2010). These criteria were most appropriate for our set of data and have been tested in other studies (Provost et al., 2023; Roger et al., under review). In the current study, a satisfying mean minimal K of 1 was maintained from a threshold of 0.01–0.17 for the task matrices. Within that range, the average normalized γ and λ also fulfilled the criteria of 1. Small-world topology based on the criteria for small networks was also fulfilled for this range; the actual index was however only higher than 1 for thresholds above 0.12. Given the size of the cerebral network, i.e., <3,000 nodes, analyses were conducted for threshold ranges of τ 1 = 0.01 , to τ m task = 0.17 , with Δ τ = 0.01 . Correlations fulfilling the threshold criteria kept their absolute value ( ρ i j ). Correlations on the principal diagonal representing each channel’s correlation with itself and correlations below the threshold were set to 0. For every threshold, we extracted 100 random graphs, derived network metrics and calculated a group mean over all graphs and subjects ( M rand ). We then divided individual metrics, derived from the real graph, by M rand resulting in normalized metrics. Finally, we calculated the area under the curve (AUC) for the normalized network metrics over the sparsity range, which provided a valuable alternative to analyzing metrics for each threshold individually (Zhang et al., 2011; Lei et al., 2015; Suo et al., 2015; Ding et al., 2017).

The statistical analyses were conducted using the R statistical software and Matlab functions integrated in the LIONirs toolbox. Between group analysis (epilepsy vs. control group) allowed to: (1) detect potential differences regarding the sample’s demographic characteristics (age, sex and SES) and socio-affective wellbeing measure with the Beck-Y questionnaire; (2) ensure equal fNIRS signal quality (the number of good trials included for FC matrices); and (3) statistically compare graph network metrics as well as cognitive performance. Student’s t-tests, or the non-parametric Mann–Whitney-U test when normality was violated, were used for continuous variables and Fisher’s exact test were used to analyze categorical variables.

Due to the high number of connectivity links (n = 1,225), 2000 unpaired permutation tests based on student t-test statistic were computed to compare the residual FC matrices of the epilepsy group to the control group (Galán et al., 1997; Lage-Castellanos et al., 2010). Insufficient power precluded quantitative analyses of patient subgroups (FLE vs. TLE, left vs. right epilepsy). Therefore, the mean and standard deviation of the control group’s residual FC matrices were used to transform the patients’ residual FC matrices into z-scores. The z-scores of four different patient subgroups, i.e., FLE, TLE, left and right lateralized epilepsy, were averaged to qualitatively evaluate the impact of epilepsy lateralization and localization on organization of cerebral language networks.

Pearson’s correlation analysis between cognitive measures (estimated IQ, expressive and receptive language skills) and the AUC of the graph metrics were used to better understand the associations between network organization and the neuropsychological profile. Subsequently, linear regression models with the cognitive measures and the group factor (epilepsy and healthy control) as predictor variables, were tested to evaluate their relative explanation of the different network properties, i.e., the AUC of the network’s segregation ( γ , E_local) of each hemisphere, the network’s integration ( λ , E_global), and its small-world topology ( σ ). Post-hoc analyses were applied to disentangle the moderating effect of the group factor on the association between cognitive measures and the AUC of the different brain network. Explorative correlation analyses between clinical factors and the AUC of the graph metrics of epilepsy patients were computed to explore the impact of the type of epilepsy (frontal/temporal), epilepsy lateralization (left/right/bilateral), age of epilepsy onset, epilepsy duration and seizure control on the language brain network topology. Pearson’s correlation coefficients and Spearman rank correlations were used for continuous and categorical variables, respectively. Results are all interpreted with a critical alpha level of 0.05 and common effect size measures, i.e., Cohen’s d, partial eta squared ( η p 2 ), and Varga and Delaney’s A (Cohen, 1988; Vargha and Delaney, 2000).

The final sample was composed of 13 children with epilepsy and 26 healthy controls (see Table 1 for socio-demographic and clinical characteristics). Comparisons between healthy controls and epilepsy patients regarding their socio-demographic characteristics revealed no significant differences for sex (Fisher’s exact test: p = 0.728), age (t(37) = 1.30, p = 0.203, d = 0.44, small effect), socio-economic status operationalized as the family income (W = 155.5, p = 0.793, A = 0.46, very small effect) nor the mean parental education (W = 202.0, p = 0.311, A = 0.60, small effect). The BECK-Youth Inventory (BYI-II) was only completed by participants aged ≥7 years, i.e., 10 epilepsy patients and 23 healthy controls. Both groups have a socio-affective well-being within the normative range, with no statistically significant differences between groups for any of the three dimensions of the BYI-II, i.e., self-concept (t(31) = 0.78, p = 0.439, d = 0.30, small effect), depression (t(31) = −0.93, p = 0.343, d = 0.37, small effect) and anxiety (t(31) = −2.00, p = 0.056, d = 0.75, medium effect). We thus considered both groups as being comparable and did not include any of the socio-demographic nor the socio-affective factors as covariables for primary statistical analysis. The etiology of epilepsy was unknown in eight patients, potentially structural (e.g., atrophy, cyst, sclerosis) in three, though one had incomplete signs of a hippocampal sclerosis, and an auto-immune disease was suspected in one. Etiology could not be verified for one. None of the patients presented a self-limited epilepsy syndrome (Specchio et al., 2022). Nine patients received a monotherapy, and four a combination of two ASMs. Carbamazepine and levetiracetam were the most common ASMs in this study group and anyone received treatment with topiramate, which is known to affect language functions (Lee et al., 2003; Szaflarski and Allendorfer, 2012). Supplementary Table S1 for details on individual patients.

On average, we recorded 12.2 (SD: ±2.8) and 12.3 (SD: ±2.1) minutes of resting-state and task data, respectively. After the bootstrapping of 60-s trials, we extracted an average of 93.2 (SD: ±59.5) and 121.6 (SD: ±59.1) valid resting-state and 67.2 (SD: ±51.9) and 95.3 (SD: ±54.2) task segments for the epilepsy and control groups, respectively. The number of valid segments for the resting-state (t(46) = 1.63, p = 0.109, d = 0.48, small effect) nor the language paradigm (t(37) = 1.54, p = 0.131, d = 53, medium effect) statistically differed between groups. When the comprehension of the story line presented during the receptive language task was assessed, epilepsy patients gave significantly fewer correct answers (40%) as compared to the healthy controls (90%, W = 181, p = 0.002, A = 0.54, very small effect).

Permutations were conducted at a channel-level, for visualization and interpretation purposes, channels were attributed to four regions of interest (ROI), i.e., frontal, temporoparietal, temporal and temporooccipital areas of both hemispheres (Figure 3A; Koessler et al., 2009; Nguyen et al., 2018). Averaged residual connectivity matrices ( R ) of epilepsy patients and healthy control children are illustrated in Figure 3B. Student’s t-test statistics after 2000 permutations revealed numerous significant differences (p < 0.05) between residual matrices of patients with epilepsy and healthy control children (Figure 3C). Group differences indicated that, compared to their healthy peers, children with FLE or TLE overall had decreased frontotemporal FC within the left hemisphere, increased frontotemporal FC within the right hemisphere, and mixed results regarding FC alterations of inter-hemispheric FC (p < 0.05, d |Min-Max|: 0.74–2.81, medium to large effect, see Figure 3D). Precisely, the epilepsy group predominantly had reduced FC between homologous regions (e.g., fronto-frontal connections) and a few stronger long-range inter-hemispheric connections (e.g., right frontal-left temporal) compared to the control group. The averaged z-score matrices for the four patient subgroups implied that children with FLE and right lateralized pathology, as compared to healthy controls, had the most different frontotemporal FC (see Figure 4). These patients had particularly increased FC of posterior temporal areas. Strikingly, children with left lateralized epilepsy had mainly left hemispheric and interhemispheric FC alterations, while their right hemisphere seemed almost unaffected and comparable to their healthy peers. The sample size prevented further statistical analyses between the subgroups.

Connectivity strengths across all channels showed overall a similar distribution of FC. Averaged FC in the epilepsy and control group was 0.11 and 0.12, respectively (see Figure 5A). Weighted graph networks of children with FLE or TLE and healthy controls revealed a small-world topology ( γ > 1 , λ > 1 , γ > γ _rand) for the language brain network over the sparsity range of τ 1 = 0.01 , to τ m task = 0.17 . Figure 5B displays the averaged AUC for each graph network metrics. We found a significant group difference for the AUC of E_local in the right hemisphere (t(37) = −2.07, p = 0.046, d = −0.70, medium effect) and a relevant but not significant difference between groups for the clustering coefficient for the same hemisphere (t(37) = −2.07, p = 0.046, d = −0.66, medium effect), but no differences between patients and controls for any of the other network metrics (p > 0.05).

Explorative correlation analysis between some clinical factors and the AUC of the graph metrics of epilepsy patients revealed that overall epilepsy type showed the highest correlation across the AUC of all network metrics (Spearman r = |0.58–0.69|), followed by seizure control (Spearman r = |0.32–0.54|). Age of epilepsy onset (Pearson r = |0.20–0.33|) had small to moderate correlations with network metrics. Lateralization (Spearman r = |0.05–0.28|) and epilepsy duration (Pearson r = |0.00–0.10|) did not show any strong associations with the AUC of network metrics (see Table 2).

Results of the neuropsychological evaluation revealed that, compared to children of the control group, children of the patients group had a significantly lower estimated global IQ (t(35) = 5.43, p < 0.001, d = 1.91, large effect, missing n_epi/ctrl: 1/1), reduced overall verbal capacities (verbal comprehension index of the Wechsler intelligence scale, VCI: t(35) = 3.37, p = 0.002, d = 1.19, large effect, missing n_epi/ctrl: 1/1), lower receptive (t(36) = 3.64, p < 0.001, d = 1.27, large effect, missing n_epi: 1), and expressive language abilities (t(37) = 2.94, p = 0.006, d = 1.00, large effect) as well as decreased non-verbal skills (perceptual reasoning index of the Wechsler intelligence scale, PRI: t(37) = 4.78, p < 0.001, d = 1.62, large effect). Detailed results of the subtests are illustrated in Figure 6.

Correlation coefficients revealed high correlations between the different cognitive measures (r > 0.70). Supplementary Table S2 provides the correlation coefficients between all variables. Due to multicollinearity between cognitive measures and because the estimated IQ had the most relevant and consistent association with network metrics across different metrics (r = |0.18–0.29|), we opted for a linear model that included the centered estimated IQ, the group factor and their interaction term to predict the AUC of the brain network metrics. Statistical results of this model are summarized in Table 3. 14.4% of the E_local variance in the left hemisphere (F(3,33) = 3.02, p = 0.044, d = 0.08, small effect) was explained by this model. 13.0% of the E_local variance in the right hemisphere (F(3,33) = 2.79, p = 0.056, d = 0.02, small effect) was explained by the included variables, which was however not statistically significant. We did not find any main effects of the group nor the estimated IQ, but there was a significant interaction effect between the group and estimated IQ for E_local of the left hemisphere (Beta = −1.46, p = 0.014, η p 2 = 0.16, large effect) and an equally large but not significant effect for E_local of the right hemisphere (Beta = −1.29, p = 0.055, η p 2 = 0.10, medium to large effect). The F-statistics of all the other metrics did not reveal a significant model fit.

Post-hoc simple slope analyses revealed that the association between the estimated IQ and E_local of the left hemisphere differed significantly between groups (p = 0.014) and at trend level for the E_local of the right hemisphere (p = 0.055). Precisely, an increase in the estimated IQ was associated with a significant decrease of E_local in the left (−1.43, p = 0.006) and the right hemispheres (1.17, p = 0.046) in the epilepsy group. In the control group, an increase of the estimated IQ did not relate to any changes of the E_local in the left (+0.03, p = 0.915) nor the right (+0.12, p = 0.714) hemispheres (see Figure 7).

In the current study, the healthy control group presented typical cerebral language networks characterized by a joint recruitment of left temporoparietal, temporal and temporooccipital areas as well as of bilateral frontal and temporal areas (Skeide and Friederici, 2016; Hickok, 2022). Compared to healthy controls, children with FLE and TLE demonstrated weaker left hemispheric and inter-hemispheric FC as well as increased right frontotemporal connections. These group differences support the known incidence of atypical functional language networks in pediatric focal epilepsy (Berl et al., 2014; Baciu and Perrone-Bertolotti, 2015; Vannest et al., 2019; Marcelle et al., 2022). While the enhanced recruitment of right hemispheric frontotemporal regions is a frequent and well-established cerebral alteration of language processing in epilepsy patients, the impact on long-distance connections is less documented. In a recent fNIRS study in adult patients with frontotemporal epilepsy, Tung et al. (2021) investigated brain FC during a semantic and a phonemic verbal fluency task and found weaker inter-hemispheric FC in adult epilepsy patients compared to healthy controls. Our findings thus suggest that weaker inter-hemispheric FC between homologous frontotemporal regions in patients with FLE and TLE compared to controls is a characteristic of cerebral language network reorganization across ages and different language domains (Sepeta et al., 2015; Tung et al., 2021).

Graph theory analysis revealed that global network topology of language processing was comparable between both groups, but that local network properties, i.e., local processing efficiency, in the right hemisphere was increased in the patient group compared to the control group. This suggests that the altered FC patterns in our epilepsy group only marginally affected the network’s overall architecture and that the increase of right intra-hemispheric FC results in more segregation of right hemispheric frontotemporal regions. The reduction of FC within the left hemisphere does not however lead to significant changes in network architecture within this hemisphere. The typical developmental timeline of network architecture indicates that local network topology, i.e., network segregation, tends to decrease with age, while global network efficiency usually increases (Supekar et al., 2009; Gozdas et al., 2019). Hence, the above reported alteration could indicate a certain immaturity of brain language networks in children with FLE and TLE. Similarly, local network organization seems a common area of alterations in patients with epilepsy as the results of the meta-analysis by Slinger et al. (2022) underline. In contrast to our findings, they report however a lower clustering coefficient, suggesting a reduction of network segregation in adolescent and adult patients with focal epilepsy. Alterations of local network properties appear to be a key characteristic of brain network topology in patients with focal epilepsy such as FLE and TLE, and this both for large-scale brain networks as well as specific functional brain networks such as cerebral language networks. The direction of reorganization, i.e., whether fewer or more subnetworks are built, seems to depend on the patients’ age.

Exploratory analyses of cerebral language networks in patient subgroups revealed that, compared to healthy controls, children with right lateralized epilepsy had overall more altered FC in cerebral language networks than patients with left lateralized epilepsy. The latter mainly had reduced FC among ipsilateral frontotemporal regions, suggesting that their language networks demonstrate primarily local deviations from the FC of healthy controls. Although epilepsy lateralization reportedly causes distinct differences in language networks, the literature rather suggests that patients with left lateralized FLE and TLE, i.e., seizure onset in the dominant hemisphere for language processing, experience more important alterations in their language networks compared to those with right hemispheric lateralization (Hamberger and Cole, 2011; Rodríguez-Cruces et al., 2020; Tung et al., 2021). Our data and others show that right hemispheric epilepsy does however not seem to prevent reorganization of cerebral language networks, causing a distinct pattern of alteration (Smith, 2016; Tung et al., 2021). This is in line with the current understanding of the typical neurobiology of cerebral language networks that suggests that despite a left hemispheric dominance, bilateral frontotemporal areas are involved in language processing (Hickok and Poeppel, 2007; Hickok, 2009, 2022). Regarding the role of epilepsy type, our results indicate that FC in children with FLE was more strongly affected compared to those with TLE with much more differences in intra- and inter-hemispheric FC in FLE than TLE compared to healthy controls. A distinct pattern of FC for epilepsy type has recently also been identified in adult patients with FLE and TLE (Caciagli et al., 2023). Precisely, fMRI data acquired during two expressive language tasks indicate that patients with FLE compared to those with TLE demonstrate increased FC within left temporooccipital and bilateral occipital areas. In our data set, epilepsy type and to some extent seizure control, showed the strongest associations with graph metrics of network topology. Epilepsy type and lateralization seem relevant predictors for the organization of cerebral language networks, though more studies are necessary to better characterize the specific pattern of reorganization associated with each subgroup. In this context, it is also important to emphasize the heterogeneous seizure semiology. Precisely, cerebral language network organization may be affected differently when seizures are purely unilateral, or epileptic activity propagates to the other hemisphere or seizures commonly result in a secondary generalization. Future studies with larger samples size should consider more precise characteristics of seizure semiology.

The neuropsychological results indicated that children with FLE or TLE had an overall cognitive performance and receptive and expressive language capacities within the normative range, however significantly lower than their healthy peers. In our sample, FLE and TLE seem to have a general impact on the cognitive capacities rather than a specific effect on language abilities. This is in line with the current state of the literature on pediatric FLE or TLE, which shows that focal epilepsy is a network disorder and affects multiple cognitive domains (Garcia-Ramos et al., 2015; Kellermann et al., 2016; Hermann et al., 2017; Verche et al., 2018; Rodríguez-Cruces et al., 2020).

During the fNIRS-EEG recording, patients with epilepsy demonstrated reduced task performance at the passive story listening task, i.e., lower comprehension of the story line, compared to the control group. This suggests that the diffuse FC alterations and increased local processing efficiency in the right hemisphere measured during language processing seem to represent aberrant network organization. Specific regression analysis revealed however a more complex pattern. The estimated IQ alone did not predict any of the network’s characteristics. However, we found a moderation effect of the group factor (patients vs. controls) on the association between the estimated IQ and the network’s local processing efficiency in the left hemisphere and at a trend level in the right hemisphere. Precisely, the estimated IQ was negatively related to the level of local processing efficiency in the patient’s group, while no such relationship was identified for the control group. No moderation effect was found for the clustering coefficient nor any of the global network metrics. Although current findings from neurotypically developing children suggest that the direction of the association between intellectual capacities and graph theoretical measures differs between cerebral regions (Gozdas et al., 2019), network segregation generally decreases during childhood and adolescence (Supekar et al., 2009). The association of a lower network segregation with higher cognitive functioning may imply that compensatory mechanisms in the left hemisphere enabled a more mature and therefore more efficient cerebral language network.

This study has several methodological strengths that can be highlighted here. First, epileptic activity can significantly affect hemodynamic fluctuations and mislead the interpretation of cerebral language processing. The simultaneous acquisition of the electrophysiological data (EEG) with the hemodynamic signal (fNIRS) thus allowed to monitor epileptic activity and ensure a good quality of the hemodynamic signal without contamination of seizure activity. In future studies, the joint multimodal analysis of EEG and fNIRS FC would certainly also offer rich information to characterize language brain networks in individuals with epilepsy. However, a task paradigm better adapted to both signals (slow fNIRS hemodynamic and fast EEG neuronal signals) and a greater number of EEG electrodes should be used. This nevertheless constitutes interesting paths for future research. Another methodological strength of this study is the subtraction of the resting-state FC matrices from the task FC matrices allowing to maximally isolate cerebral activation related to receptive language processing. Finally, a third strength is the use of a graph theoretical approach to characterize brain language networks in children with FLE and TLE, where such literature on this specific group of patients is still scarce.

This study also has several limitations that must be considered when interpreting its results. First, the heterogeneous clinical phenotype of epilepsy and the high number of comorbidities makes the recruitment of a homogeneous sample challenging and resulted in the current study in a relatively small sample. We carefully considered the sample size while performing statistical analysis and emphasize the value of effect sizes. The impact of few clinical factors (e.g., epilepsy type and lateralization, age at onset, duration of epilepsy, seizure frequency and control) on brain functional connectivity could only be addressed in an explorative manner and the sample size precluded their consideration for advanced analysis. Due to high heterogeneity, other clinical variables (e.g., etiologies, number, and type of ASM, etc.) could not be considered for analyses. It has however been shown that clinical factors correlate with the extent and type of cognitive difficulties (Jambaqué et al., 1993; Gallagher and Lassonde, 2005; Verche et al., 2018). Therefore, their role for cerebral language processing and the neuropsychological outcome should be specifically addressed in upcoming research with larger samples. This will allow to shed light on differences in brain network architecture of epilepsy patients and enhance the understanding of the role of clinical characteristics on the association between neuropsychological profiles and cerebral network organization. Finally, the current findings are deducted based on a story listening task. Therefore, whether they represent specific cerebral characteristics of receptive language, or rather a general effect of language processing must be analyzed in a study applying different language task paradigms. This would for instance allow us to compare neuronal response between receptive and expressive language processing or other distinct linguistic processes.