Probabilistic mapping of language networks from high frequency activity induced by direct electrical stimulation

Abstract Direct electrical stimulation (DES) at 50 Hz is used as a gold standard to map cognitive functions but little is known about its ability to map large‐scale networks and specific subnetwork. In the present study, we aim to propose a new methodological approach to evaluate the specific hypothesis suggesting that language errors/dysfunction induced by DES are the result of large‐scale network modification rather than of a single cortical region, which explains that similar language symptoms may be observed after stimulation of different cortical regions belonging to this network. We retrospectively examined 29 patients suffering from focal drug‐resistant epilepsy who benefitted from stereo‐electroencephalographic (SEEG) exploration and exhibited language symptoms during a naming task following 50 Hz DES. We assessed the large‐scale language network correlated with behavioral DES‐induced responses (naming errors) by quantifying DES‐induced changes in high frequency activity (HFA, 70–150 Hz) outside the stimulated cortical region. We developed a probabilistic approach to report the spatial pattern of HFA modulations during DES‐induced language errors. Similarly, we mapped the pattern of after‐discharges (3–35 Hz) occurring after DES. HFA modulations concurrent to language symptoms revealed a brain network similar to our current knowledge of language gathered from standard brain mapping. In addition, specific subnetworks could be identified within the global language network, related to different language processes, generally described in relation to the classical language regions. Spatial patterns of after‐discharges were similar to HFA induced during DES. Our results suggest that this new methodological DES‐HFA mapping is a relevant approach to map functional networks during SEEG explorations, which would allow to shift from “local” to “network” perspectives.


| INTRODUCTION
Direct electrical stimulation (DES) has been for a long time a routine clinical practice during intraoperative and extraoperative neurological evaluation (Penfield & Jasper, 1954). Extraoperative evaluation is specifically performed during presurgical stereo-electroencephalography (SEEG) evaluation in patients with severe drug-resistant epilepsy, candidates to curative resective surgery. DES is performed in these patients for two main objectives (Kahane et al., 1993;Kahane & Dubeau, 2014;Penfield & Jasper, 1954): (a) spatial definition of the "epileptogenic zone" (Rosenow & Lüders, 2001) and surgical boundaries by attempting to reproduce the patient's usual type of seizure, and (b) spatial identification of "functional zones" (i.e., essential cortex) to be spared during surgery in relation to cognitive functions such as language. DES is currently considered as a "gold-standard" method for individual functional mapping for both intraoperative and extraoperative evaluations (Desmurget, Song, Mottolese, & Sirigu, 2013;Mandonnet, Winkler, & Duffau, 2010). In this context, when a small electrical current is delivered in one cortical site and a transient functional disruption is observed (e.g., speech arrest [SA]), this region is considered to be an eloquent area. Thus, given the causality between electrical stimulation and behavior deficit or dysfunction, DES is used to predict potential deficits during cortical resection and the decision to perform the resective surgery is based on its results (for recent reviews of DES, see Borchers, Himmelbach, Logothetis, & Karnath, 2011;Desmurget et al., 2013;Selimbeyoglu & Parvizi, 2010).
It has been shown that a cortical site stimulation could induce different behavioral responses and a similar behavioral response (e.g., SA) can result from the stimulation of various sites (Borchers et al., 2011;Desmurget et al., 2013). This is in line with the current view considering that cognitive functions brain representation depends on large brain networks (Power et al., 2011;Yeo et al., 2011), that is, distributed groups of interconnected and synchronized neurons, rather than isolated functional cortical areas (Duffau, Gatignol, Mandonnet, Capelle, & Taillandier, 2008;Duffau, Moritz-Gasser, & Mandonnet, 2014). Thus, during DES, an specific language dysfunction could be observed by the stimulation of different cortical site of a specific subnetwork (e.g., phonological or semantic paraphasia [SP]) (Mandonnet et al., 2010;Mandonnet et al., 2016). This is also in agreement with results indicating that over 50% of patients show postoperative language deficits (naming decline) despite DES for language mapping prior to surgery (Davies, Risse, & Gates, 2005). Therefore, functional deficit induced by 50 Hz DES of a cortical region may reflect dysfunction of a large-scale network (Mandonnet et al., 2010;Mandonnet et al., 2016). For instance, in a case study, DES on basal temporal cortex (BTC) produced aphasic symptoms without language deficits after BTC resection (Ishitobi et al., 2000). This was explained by the association during the DES on BTC with intrastimulus remote discharges in posterior superior temporal cortex and thus suggested that distant effect outside the stimulating current field is exerted by DES (Ishitobi et al., 2000). Thus, intrastimulus remote discharges as well as after-discharges can result in false localization of functional cortex as they may occur remotely from the stimulate site and consequently induce behavioral manifestation unrepresentative of the stimulate cortical site (Blume, Jones, & Pathak, 2004;Karakis et al., 2015). In the present study, we hypothesize that DES-induced language interference (naming errors) could be considered as input gates into a larger language network (Mandonnet et al., 2010).
For the present study, we developed a new approach to map the networks showing high frequency activities (HFAs) induced by 50 Hz DES during language disturbances. The method is largely inspired by the one already developed by our group to map ictal HFAs (David et al., 2011). Specifically, this methodology aims to identify brain areas whose HFA is significantly greater than baseline HFA using standard neuroimaging time series analysis. This involves a simple categorical comparison (using t tests) between mean activity at baseline and mean activity over short windows (e.g., 3 s) at various times (e.g., 0-20 s) after DES. The significance of these differences was evaluated in relation to the variability of fluctuations within each time segment. The data features we compared were the fluctuations in HFA (70-150 Hz), shown to be a potential specific biological support of cognitive function (Lachaux, Axmacher, Mormann, Halgren, & Crone, 2012;Muller et al., 2018;Perrone-Bertolotti et al., 2020). We applied this methodology to assess large-scale language networks by evaluated DES-induced language errors and measured HFA modification during these errors in nonstimulated brain regions. According to our working hypothesis, each cortical site in which DES induces language errors may be part of a larger language network, in which the activity is also modulated by DES (Mandonnet et al., 2010). Therefore, language errors induced by the DES may reflect activity perturbation of a large-scale language network instead of only the stimulated site and would explain why similar language symptoms are associated with electrical disturbances not only of local but also of remote cerebral regions (Mulak, Kahane, Hoffmann, Minotti, & Bonaz, 2008). Consequently, this methodological approach can be able to identify language network and subnetworks reflected by the HFAs cooccurrence, in either adjacent or distant regions of the stimulation site, and in particular in perisylvian and extrasylvian language areas.
To address our objective, we retrospectively analyzed data from extraoperative DES performed in 29 patients with drug-resistant epilepsy. For each patient, we identified the language errors induced by the DES during a picture-naming task and reported the corresponding stimulation cortical sites. This task, classically used during clinical DES language mapping, involves a large fronto-temporo-parietal network (Corina et al., 2010;Haglund, Berger, Shamseldin, Lettich, & Ojemann, 1994;Indefrey & Levelt, 2004;Ojemann, 1983;Rofes et al., 2018).
First, we mapped the HFA elicited in nonstimulated recorded sites during each observed type of language error. Second, we put in relation the HFA modifications with five classical core language regions in order to identify language subnetworks related to linguistic processes underlying language errors induced by the DES during naming.

| Patients
From the archives of Grenoble University SEEG laboratory from 2009 to 2013, we identified 29 patients (13 females) who carried out standard presurgical evaluations and experienced language symptoms elicited by DES (Table 1). Patients were fully informed and gave their consent to undergo SEEG recordings as part of the presurgical evaluation of their drug-resistant epilepsy, in addition to noninvasive exams (high-resolution structural magnetic resonance imaging [MRI], video-EEG monitoring, neuropsychological evaluation). DES and SEEG recordings were performed according to the routine procedure used at Grenoble University Hospital (Kahane et al., 1993;Kahane & Dubeau, 2014).

| Electrode implantation and positioning
A total of 11-15 semirigid, multilead electrodes were stereotactically implanted in each patient. Each electrode had a diameter of 0.8 mm and, depending on the target structure, consisted of 8-18 contact leads 2 mm wide and 1.5 mm apart (DIXI Medical, Besançon, France).
Electrodes implantation was strictly related with individual clinical hypothesis. A preoperative MRI and postoperative MRI or CT scan were co-registered to assess the locations of the electrode contacts for each patient using a coordinate system in relation to the anterior commissure/posterior commissure plane. Electrode contact positions were finally expressed in the Montreal Neurological Institute (MNI) coordinate system to allow group analyses after brain spatial normalization using Statistical Parametric Mapping 12 software (SPM12, Wellcome Department of Imaging Neuroscience, University College London, www.fil.ion.ucl.ac.uk/spm). Visual inspection of the contact locations was also used to check whether each electrode contact was located in gray or white matter. Furthermore, to be able to perform group analysis, we used regions of interest (ROIs) in the automated anatomical labeling (AAL) parcellation system (Tzourio-Mazoyer et al., 2002), including 43 cortical regions per hemisphere being potentially implantable with SEEG electrodes.

| SEEG recordings
SEEG recordings were performed using a video-EEG monitoring system (Micromed, Treviso, Italy) that allowed to simultaneously record up to 256 monopolar contacts, so that a large range of mesial and cortical areas, as well as fissural cortices, was sampled for each patient.
Sampling rate was 512 Hz, with an acquisition band-pass filter between 0.1 and 200 Hz. Data were acquired using a referential montage with reference electrode chosen in the white matter. All other recording sites were chosen in the gray matter. For data analysis, we used a bipolar montage between adjacent contacts of the same electrode to improve sensitivity to local current generators. Coordinates of virtual bipolar contacts that were used to construct images (see below) were chosen to be at an equal distance of two successive contacts.

| Direct electrical stimulation
Stimulations at 50 Hz were applied between two contiguous contacts at different levels along the axis of each electrode, chosen to be in the gray matter according to visual inspection of spontaneous SEEG signals. A bipolar montage was used to help eliminate the signal common to adjacent electrode contacts and to avoid artifacts coming from distant sources. Bipolar stimuli were delivered using a repetitive biphasic square wave electric currents of 1-3 ms pulse width designed for a safe diagnostic stimulation of the human brain (Micromed), according to parameters proved to produce no structural damage. The intensity used was usually 3 mA or less (down to 1 mA), for a maximum duration of 5 s, or less depending on the type of the induced clinical response.

| Picture naming
Patients were required, in others, to overtly perform a picture-naming task in black and white drawing images representing objects from several semantic categories (Snodgrass & Vanderwart, 1980). Prior to DES, patients were trained to name each picture (control trials) in order to ensure task feasibility and picture familiarity (behavioral baseline level). During DES, pictures were presented randomly by the neurologist and the stimulation was concomitantly applied. Naming errors were considered only when the behavioral performance was disrupted during DES compared to the reference condition (control trials without DES) performances. It is important to note that patients also performed other tasks (such as reading) for the clinical purpose of functional explorations, not presented here. However, the reading errors induced by stimulations were used to control for the specialization of language areas in naming (see Section 3.2).

| Types of naming errors
Naming errors were retrospectively analyzed for each patient by a speech therapist according to other previous DES studies using picture naming (Corina et al., 2010). Three types of naming errors have been identified: (a) SA: with total or partial lack of naming, including anomia, speed reduction, and hypophonia (soft speech); (b) SP including word substitution with a related or associated word (i.e., the word "clock" when the picture presented represented a "watch" or the word "animal" for the target word "dog"); (c) phonological paraphasia (PP) including words errors related to phonological retrieval and phoneme selection (e.g., phonemic omission, substitutions, elision, transposition or inversion) in which a nonreal word with a phonological resemblance to the target word was generated (e.g., for the target word "pencil" the word "bencil" is generated).
2.7 | Data processing and statistical analysis 2.7.1 | Review of SEEG/video events Each DES trial having induced a language interference was carefully reviewed in order to: (a) classify the language error (see above), allowing several errors to occur in the same trial and (b) add events to each SEEG files that will be used in further processing of SEEG data (see below). The events aimed at quoting (Figure 1): • The start and end of the DES based on the stimulation artifact.
• The start and end of the language symptom based on the video.
• The end of the after-discharge or fast oscillations induced by the stimulation, as such electrophysiological signature was visually identified in all instances.

| Processing of single SEEG trials
The processing of single SEEG trials was derived from the procedure we developed to map epileptogenicity, that is, fast oscillatory activity at seizure onset (David et al., 2011). , start and end of the language symptom (symptom ON; symptom OFF, respectively), and end of after-discharge (after-discharge OFF). Note that in this case, the symptom was reading speech arrest. The time-frequency decomposition of power of each channel (units: z-score according to baseline levels) clearly shows the artifact of stimulation with vertical bars at the start and the end of stimulation and horizontal patterns at the frequency of stimulation and its harmonics. The left calcarine sulcus did not respond to the stimulation as no significant change of SEEG could be detected. In the left middle temporal gyrus, one can notice changes of SEEG power above 120 Hz in correlation with the clinical symptom (start between 2 and 3 s after stimulation onset). After the stimulation, the after-discharge showed activity below 40 Hz thus ignored when computing SEEG power time series at every time bin by averaging values along the frequency dimension. In addition, as a conservative measure, SEEG power matrix elements with a z-value above 10 in at least 5% of channels were also removed from all channels, that is even in channels where the value did not reach the threshold. The same processing steps were applied to the baseline data, even if there were no artifact to avoid any analysis bias in the final statistical comparison.
For mapping after-discharges, similar analyses as for DES-induced HFA were performed but the frequency band was set to 3-45 Hz, and the period of interest was set from the end of the stimulation period (t = 5 s) up to the termination of the after-discharge.
Knowing SEEG electrode positions in a standardized coordinate system (MNI coordinates) and SEEG power values for each electrode, it is possible to create images of SEEG power by local spatial interpolation (David et al., 2011). We used an interpolation procedure that constrains voxels to the neocortical mantle, the hippocampus, and the amygdala, which are the structures the most frequently explored for epilepsy surgery in temporal lobe epilepsy patients. It works as follows (David, 2019): (a) for each electrode contact, the mesh vertices at a distance less than 1 cm are detected and (ii) to each vertex detected in the vicinity of at least one electrode, the assigned value is the average of the SEEG power values of the close electrodes, weighted by the inverse of the distance between the vertex and the electrodes (stronger weight is given to the closest electrodes). Once created, the images of SEEG log power were smoothed using an isotropic Gaussian kernel with a width of 3 mm (equivalent to the distance between successive SEEG electrodes) in order to control family-wise error in the context of spatially correlated imaging data using the theory of Gaussian random fields (Worsley, Taylor, Tomaiuolo, & Lerch, 2004). Statistical significance of the difference between baseline and symptom periods (two sample t-test) was directly obtained by the family-wise errorcorrected associated p-values. By applying a threshold on this p-value (.05 corrected for multiple comparison), it was possible to determine the regions showing significant fast discharges concomitantly to the language symptoms during DES, for each trial.

| Group analysis
Group analysis of SEEG responses was performed for deriving maps of the probability of recording significant responses for each type of language error. We pooled together patients and events in conjunction analysis with the following steps: 1. For each trial, a binary map showing voxels having a significant SEEG power during the language error was obtained by thresholding the statistical maps of SEEG log power to p < .05.
2. Each surviving voxel was assigned to a specific ROI using a predefined parcellation scheme. Every parcel with at least one activated voxel was considered as active, and thus a binary map of active (1) and nonactive (0) parcels was produced for each event. For this study, we used the AAL scheme (Tzourio-Mazoyer et al., 2002).
3. The group probability map of symptom-related SEEG power was obtained by averaging across trials the binary maps at ROI level.
For each type of symptoms, we considered only ROI recorded at least 10 times to ensure a certain level of reproducibility in the results.
The values of the group probability map ranged between 1 (ROI systematically showing significant error-related increase of HFA power when recorded) and 0 (ROI found to never show any significant symptom-related increase of HFA power when recorded). We called those maps "HFA maps." A similar methodology was applied in order to produce maps of the probability to induce a particular symptom with DES for each ROI (number of times a symptom was induced when stimulating ROI n divided by the number of times ROI n was stimulated). We call those maps "DES maps." For display purposes, those HFA and DES maps are shown in figures on a canonical inflated brain.
Finally, we used HFA maps to evaluate probabilistic functional cooccurrence of HFA elicited during naming symptoms on a language network composed of five ROIs. These five ROIs were significantly stimulated and recorded in our dataset: inferior frontal gyrus at the pars triangularis and opercularis; insula; superior and middle temporal gyrus. For that purpose, we simply computed the group probability using conjunction analysis, which gave a value between 1 (recorded ROIs systematically showing increased HFA power when one specific language-related ROI presented a symptom during DES, independently of the symptom type) and 0 (recorded ROIs found to never show any HFA power modification when one specific languagerelated ROI presented a language symptom during DES, independently to the symptom type). This ROI-based analysis was duplicated to map the propagation of after-discharges.

| Visual analysis of SEEG-video recordings
Visual analysis of DES responses as illustrated in Figure 1 indicated a median symptom onset of 1 s, with a median duration of 6.26 s. The median value of the termination of the observed postdischarge was 11.7 s. Temporal precision of behavioral measurements was, however, limited because of the difficulty to assess the dynamics of language symptoms from SEEG-video recordings. One should also note that because of the huge stimulation artifact, it was not possible to visually measure the onset of postdischarge from SEEG recordings. According to the raw data of naming errors (Table 2), the most frequently observed errors were naming SA (79.81%), naming PP (10.58%) followed by naming SP (9.62%). Naming errors were more frequently observed in the left dominant (LH, 87.50%) than in the right nondominant hemisphere (RH, 12.50%). Importantly, only NSA errors were observed in the right hemisphere.

| DES and probabilistic naming error analysis
To quantify the involvement of an ROI during naming errors we first reported the probability to observe a naming error, when stimulating the ROI (Table 3 and Figure 3). This probability was defined as the number of errors observed in one ROI over the total amount of stimulation with language interference performed in that same region. Only ROIs with at least three induced language interference were considered, which represented 11 ROIs of the left hemisphere (Table 3).

| SEEG HFA and naming errors
We present now the SEEG HFA power modification observed during naming errors. We evaluated the significant HFA difference (t test) between: before and during induced DES naming errors (see Section 2). Figure 4 and Supplementary   Table S1 for details).

| SEEG and naming subnetworks
In order to evaluate brain language functional subnetworks related to naming, we evaluated HFA induced by DES in core language regions.
Specifically, we designated five ROIs: the left inferior frontal gyrus at the pars opercularis and pars triangularis, the left insula, the left middle temporal and superior temporal cortex. This ROI analysis allowed us to evaluate the probability to observe an HFA change in a near or in a far region from these core language regions stimulation and reveal functional subnetworks ( Figure 5 and Supplementary Table S2 as well as supplementary Figure S2 and Table S7 for results in other frequency bands).
We showed that stimulation of the left inferior frontal gyrus at the pars opercularis ROI was related with HFA modification in a large left fronto-parieto-temporal network (see supplementary Table S2   Interestingly, the ROI analysis of HFA responses and of afterdischarges showed very similar patterns, although the probability of HFA responses was found slightly higher ( Figure 6 and supplementary Table S7 for probability values). It demonstrates that the areas connected to the stimulated site can be inferred from both types of features.

| DISCUSSION
In the present study, we bring new evidence supporting the view that language is the result of distributed populations of interconnected and synchronized neurons, including subnetworks rather than to isolate functional areas, related to different language processes (Duffau et al., 2008;Duffau et al., 2014). Our hypothesis was that cortical sites related to language disturbances (naming errors) during DES, may be considered as input gates into a larger language network (Mandonnet et al., 2010). We identified and put in relation both cerebral regions  (Table 3) during naming errors. It is important to note that only one patient included in the present study presented a right hemispheric predominance for language as assessed by fMRI, therefore, its SEEG implantation was in the left nonspecialized hemisphere (Table 1). Furthermore, the seven patients with right hemisphere SEEG implantation presented left hemispheric specialization for language. Thus, the right hemisphere sampling in the present study is representative of a right nonspecialized hemisphere.
Raw data also indicated that SA was the most prevalent naming error type followed by phonemic paraphasia (PP), and the less prevalent was SP (Table 2). In line with these results, DES probability group analysis showed that SA was induced by DES performed in a larger perisylvian network than PP and then SP. Important PP and SP were shown in DES performed in similar brain regions than SA, and only PP errors were observed on a supplementary region, outside the common network, the middle frontal cortex. This result suggests a common language network for the three types of naming errors evaluated in the present study (Table 3). It is important to note that in the present study, we were unable to differentiate the SA symptom induced by motor difficulties or by lexico-semantic access (i.e., anomia). Indeed,  Table S7 Mandonnet (2017) suggested that to be able to made this difference during picture naming, the patient should always be asked to say aloud automatic sentences "This is a…." Unfortunately, this methodology was not used in the present study, which is a retrospective analysis of data recorded in a clinical routine setup.
The common network observed in the present study during DES for the three types of language symptoms could be explained by the low rate of PP and SP types of errors (9 and 10% of the total number of observed errors, respectively), similarly to previous DES findings (Corina et al., 2010;Duffau et al., 2014). It is possible that SA errors observed during perisylvian areas stimulation induces a complete inhibition of language output and preventing to verify whether PP or SP processing has been distributed. In line with this assumption, in the present study, all the DES cortical regions inducing SA covert the DES cortical regions inducing PP and SP.
Specifically, PP were observed during DES performed on lateral middle frontal, inferior frontal (pars opercularis) and lateral temporal cortices; SP errors were observed during DES performed on insula and lateral temporal cortices (Table 3). PP and SP cortical distribution can be related to the dual stream organization for naming , including a ventral (fronto-temporal) and a dorsal (frontotemporo-parietal) pathway. The ventral stream supports processing of information from visual analysis to meaning, reflected in our study by SP and SA errors induced by the DES on temporal cortices. The dorsal stream is rather involved in processing of information from visual analysis to overt articulation; these processes are reflected in our study by PP and SA errors induced by DES applied on fronto-temporal and insular cortices (Table 3 and Figure 3).
The dual stream organization during naming was supported by the SEEG HFA analysis (Table S1 and Figure 4) which highlighted the massive coactivation observed at intrahemispheric and interhemispheric levels. The DES-induced SA errors were associated with high probability of HFA changes in an extended coactivated bilateral fronto-parieto-temporal network including the left occipital gyrus (related to visual processing of pictures, see Table S1). The probability of HFA changes induced by DES associated with PP and SP errors, concerned a left restricted and less extended network (Table S1).
We were also able, by the analysis of ROI related to HFA coactivation or to after-discharges, to identify different language subnetworks. We put in relation these subnetworks with literature results on naming. Specifically, at least four linguistic processes underlying picture naming could be proposed: visual perception, semantic retrieval, phonological representation, and speech articulation Indefrey & Levelt, 2004;Lau et al., 2015;Rofes et al., 2018). Our ROI analysis showed that the inferior frontal gyrus (pars opercularis) is related with an extended left lateralized network including cortical regions identified as related with the three of the four linguistic processing involved on naming. Specifically, coactivated regions including medial temporal cortices and the temporal pole (semantic retrieval), phonological representations (fronto-parietal regions) and motor, premotor and inferior frontal regions (motor planning and speech articulation). A recent study revisiting the brain subnetwork involved during naming process identified the inferior frontal gyrus (pars opercularis) as core region on the phonological, syntactic and cognitive control hub supported by the extended anatomical connectivity with frontal, temporal and prefrontal regions   Farah, 1997). In particular, during naming the production of the target word implies a selection from a competing set of other words, this selection mechanism implied regulatory cognitive processing by biasing competitive interaction among incompatible representations (Schnur et al., 2009). We also showed a more extended number of regions with less probability results, including parietal and medial temporal regions.  (Binder, Desai, Graves, & Conant, 2009), and in relation with supramodal integration and concept retrieval as proposed in previous studies focused on the left mid temporal cortex.
Only the insula and the superior temporal ROIs showed bilateral HFA network representation. Indeed, the left insula showed the strongest coactivation with left and right frontal regions, as well as right insula. We interpret this subnetwork as reflecting articulatory processes and executive control required by language production.
Indeed, insula is involved in articulatory planning (Dronkers, 1996;Price, 2012), articulatory planning and control processes (Ackermann & Riecker, 2010;Lau et al., 2015) as well as in decision-making (Droutman, Bechara, & Read, 2015). Finally, results for left superior temporal ROI showed extended bilateral fronto-parieto-temporal network including several occipital regions and possibly related to semantic and phonological integration during visual processing. According to its anatomical location, this ROI may ensure the interplay between the dorsal and ventral streams Moritz-Gasser, Herbet, & Duffau, 2013;Vigneau et al., 2006). Indeed, DES findings showed that the superior temporal gyrus has a multimodal organization with neurons involved in both visual recognition and naming (Roux et al., 2015). Furthermore, the superior temporal cortex was involved in word retrieval during a verbal self-monitoring task (Hocking, McMahon, & de Zubicaray, 2009;Price, 2010), which uses similar processes as picture naming performed by our patients.
In addition, we also explored medial temporal regions such as the hippocampus. Although the DES applied to this region did not induce language symptoms, the HFA SEEG recordings revealed significant HFA responses in the left hippocampus and bilateral parahippocampus during SA and SP errors. Interestingly, ROI analysis revealed that the hippocampus and the parahippocampal gyri were involved in three of five explored subnetworks (absent for the pars triangularis and pars opercularis ROI analysis, see Table S2). Medial temporal regions are generally less considered by the language models, although they can have a role during word production (Price, 2012), as suggested by neuropsychological data in patients with hippocampal sclerosis (Bonelli et al., 2011;Hamberger et al., 2007;Lau et al., 2015). Hamberger (2015) highlighted that the role of the hippocampus during language production and more specifically during naming. Recently, the hypothesis that the hippocampus plays an essential role on the association between the identity (picture) and the corresponding label (name) retrieval during naming was evaluated (Hamamé, Alario, Llorens, Liégeois-Chauvel, & Trébuchon-Da Fonseca, 2014). These authors recorded HFA (50-150 Hz) in SEEG during a naming task and showed that hippocampus activity latency predicted the naming behavioral latency. Most interestingly, they showed that the absence of the hippocampus activity was related to difficult naming (e.g., "tip-ofthe-tongue"). Our results support the idea that the hippocampus is involved during language production and more generally in a global process of incoming visual information and linguistic output retrieval association.
One major limitation of the present study is the limited number of patients included (N = 29). Although this cohort is large for an SEEG study, SEEG anatomical implantation is different between patients, as they depend on clinical assumptions. This implies that a larger number of patients would have been needed to be able to have a good anatomical coverage of the full-language network. The limited number of patients included in the study prompted us to conduct an analysis in terms of ROIs. This ROI analysis is also a major limitation of our study because we decreased the inherent anatomical precision of SEEG.
Another limitation of the present study is related to the more important left hemisphere sampling, including 1,481 sites on the left and 599 sites in the right hemisphere. This suggests that even if the majority of the included patients (28/29) presented a left hemisphere specialization for language, the results related to the hemispheric involvement during naming should be taken with caution due to the sampling bias in favor of the left hemisphere. Finally, to have better insight into the specificity of induced fast oscillations for language function, it would have been extremely relevant to contrast the responses to the same stimulated sites with versus without induced symptoms. However, asymptomatic electrophysiological DES data was not available in our clinical setup and we could not proceed to this type of analysis.
In conclusion, our study supports the idea that DES-induced lan-