The inclusion criteria included: (a) extraoperative SEEG recording at the epilepsy center of Juntendo University Hospital in Tokyo, Japan between October 2021 and September 2023, (b) functional brain mapping with measurement of language task-related SEEG signal changes as described below. We performed all procedures as part of our routine presurgical evaluation, and the spatial extent and duration of SEEG recording and types of tasks were determined based on clinical demands (25, 31). The Juntendo University Institutional Review Board has approved the present study. We obtained written informed consent from the patients or their legal parents/guardians.

All patients had platinum depth electrodes implanted stereotactically using the ROSA robotic trajectory guidance system (Zimmer Biomet, Warsaw, IN, United States). An anchor bolt tightly secured each SEEG depth electrode and minimized movement-related artifacts. Each depth electrode had either six or ten contacts (length: 2.41 and 2.29 mm, respectively; center-to-center distance: 10 and 5 mm, respectively; AD-TECH, Racine, WI, United States).

Using a pre-implant T1-weighted spoiled gradient-recalled echo sequence MRI and a post-implant CT image, we created a fusion image accurately denoting SEEG depth electrodes within the brain (14, 18, 32). Based on the assessment of coronal, axial, and sagittal fusion images, a given depth contact was classified into one of the following anatomical categories as previously defined (18): (a) deep white matter (white matter deeper than the bottoms of two adjacent sulci), (b) shallow white matter (white matter more superficial than the bottoms of two adjacent sulci), (c) gray matter (i.e., cerebral cortex), (d) cerebrospinal fluid space (i.e., ventricle and subarachnoid space), and (e) outside the arachnoid surface of the brain. We used the Statistical Parametric Mapping 12¹ to spatially normalize each electrode to the standard Montreal Neurological Institute space [Figure 2; (14, 18, 32)].

We acquired SEEG data at the bedside with a sampling rate of 1,200 Hz using a g.HIamp biosignal amplifier (g.tec medical engineering GmbH, Austria). The original reference was the voltage of a white matter contact. We excluded electrode sites in the seizure onset zone (25), those located in the cerebrospinal fluid space, and those continuously affected by artifacts from further analysis.

We measured language task-related modulations of SEEG signals using a method similar to those previously reported (14, 31, 33). None of the patients had a seizure within 2 h prior to or during task performance. While awake and comfortably seated in a room with unwanted noises minimized, patients received the auditory naming task with vocalized (overt) and non-vocalized (covert) responses (i.e., a total of two tasks: Figure 1). Each task consisted of 100 trials with unduplicated stimuli. We used the Japanese auditory stimuli provided online (31). All stimuli have a 1.8-s duration. Stimuli were presented every 10 s. Patients were instructed to answer “I do not know” in Japanese when they did not know the answer or did not understand a stimulus.

We re-referenced SEEG signals using EEGLAB v2019.1 (34) and analyzed event-related responses using the three different montages. (Common average reference montage) the reference consisted of the average of SEEG signals of all channels. We computed the average reference voltage by excluding all of the SEEG signals at the seizure onset zone, irritative zone, artifactual channels (25), and those outside the brain (18, 35). (Bipolar montage) Each electrode was referenced to the adjacent electrode on a deeper side. (Laplacian montage) Each electrode was referenced to the average of adjacent electrodes on both sides (17, 18, 36).

We measured task-related high-gamma activity, as reported in our previous studies (14, 18, 32). We used the FieldTrip toolbox² to perform the Morlet wavelet time-frequency analysis on SEEG signals in the common average reference, bipolar, and Laplacian montages. We transformed SEEG voltage signals from −200 to 5,000 milliseconds period centered on each stimulus onset into time-frequency bins at a 60–140 Hz broadband range [2-Hz frequency bins; a given frequency band divided by seven cycles; sliding in 10-millisecond steps (14);]. Next, we computed the percent change of high-gamma (60–140 Hz) amplitude at each electrode site, compared to the baseline period 200–500 milliseconds prior to stimulus onset for each trial. Finally, we created movies animating the spatiotemporal dynamics of high-gamma amplitude changes in the standard Montreal Neurological Institute brain (Figures 3, 4 and Supplementary Video S1).

Mixed model analysis tested the following hypothesis. [1] Electrodes located outside the brain parenchyma and in the white matter would show lower task-related signal changes in the bipolar and Laplacian montages than in the common average reference montage, and electrodes located in the gray matter would maintain signal changes in the bipolar and Laplacian montages. [2] During the response phase, overt responses would elicit higher signal changes outside the brain parenchyma than covert responses. The dependent variable was the maximum high-gamma amplitude change in deep white matter, shallow white matter, gray matter, and outside the brain parenchyma during the stimulus (0–1.8 s) and response (1.81–5 s) phases in a given trial, respectively. The fixed effect predictors included types of electrode montages (common average reference montage, bipolar montage, or Laplacian montage) and vocalization conditions (Overt response or covert response). The random factors included patient and intercept.

Statistical analyses were performed using “IBM SPSS Statistics version 27” (IBM Corp., Armonk, NY, United States) and “Statistical and Machine Learning Toolbox MATLAB 2018b” (Mathworks, Natick, MA, United States). The significance was set at an FDR p-value of 0.05.

All SEEG data, as well as the Matlab-based code used in the present study, are available upon request to the corresponding author.

We analyzed the empirical data from 12 patients who underwent high-gamma modulation language mapping during extraoperative SEEG recording (Table 1). SEEG recording was done with 47.8 [±13.6] analyzed depth electrodes per patient on average (standard error). According to the consensus following group discussions, 67 depth electrodes were located at the deep white matter, 86 at the shallow white matter, 309 at the gray matter, and 111 outside the brain (Figure 2).

Figures 3, 4 and Supplementary Video S1 demonstrate the spatiotemporal dynamics of high-gamma modulations during the auditory naming task. Electrodes outside the brain parenchyma showed amplitude augmentation in the common average reference montage, especially during overt responses in the response phase (i.e., later than 1,800 milliseconds after the stimulus onset). In bipolar and Laplacian montages, these augmentations decreased, and amplitude augmentation in the adjacent cortex became prominent. Figure 5 shows a representative spectrogram during auditory naming with overt responses. The amplitude changes outside the brain parenchyma seen in the common average reference decreased in the bipolar and Laplacian montages, while those in the gray matter were maintained.

Figure 6 presents how well bipolar and Laplacian montages reduced the auditory naming-related high-gamma amplitude recorded at contacts outside the brain parenchyma and in the deep white matter during the response phase. Compared to those measured on common average reference montage, the median amplitude changes during overt naming in the response phase at contacts outside the brain was 62.1% smaller on bipolar montage and 48.8% smaller on Laplacian montage. The median amplitude changes at contacts in the deep white matter were 24.0% smaller on the bipolar montage and 44.6% smaller on the Laplacian montage. Conversely, the high-gamma amplitudes in the gray matter were well maintained on bipolar and Laplacian montages (1.8% smaller on bipolar montage and 11.4% smaller on Laplacian montage). During the stimulus phase, the high-gamma amplitudes in the deep white matter were 12.9% smaller on the bipolar montage and 36.5% smaller on the Laplacian montage. Neither contacts outside the brain parenchyma nor contacts at the grey matter showed more than 10% differences between the montages.

The mixed model analysis (Tables 2, 3) likewise demonstrated that during the response phase, bipolar montage and Laplacian montage reduced the high-gamma amplitudes recorded at contacts outside the brain parenchyma (mixed model estimate: −90.6 to −91.2; p < 0.001; t = −13.921 to −10.158) and those in the deep white matter (mixed model estimate: −85.4 to −31.5; p < 0.001; t = −14.827 to −6.254). The high-gamma amplitudes at the gray matter were well maintained on bipolar montage and Laplacian montage (mixed model estimate: +7.3 to +7.8%; p = 0.099 to 0.085; t = +1.786 to +1.724). During the stimulus phase, bipolar montage and Laplacian montage reduced the high-gamma amplitudes recorded at contacts in the deep white matter (mixed model estimate: −64.5 to −21.2; p < 0.001; t = −14.652 to −5.499). At the contacts outside the brain parenchyma, the Laplacian montage did not significantly reduce high-gamma amplitudes, and the bipolar montage only reduced mixed-mode estimates by about 20%.

Figure 7 shows the changes in auditory naming-related high-gamma amplitudes in each vocalization condition and montage at the contacts outside the brain parenchyma. During the response phase, the overt responses elicited larger amplitude changes than the covert responses. These amplitude changes were smaller in bipolar and Laplacian montages compared to the common average reference montage.

The mixed model analysis (Table 3) demonstrated that overt responses elicited larger amplitude changes outside the brain parenchyma during the response phase than covert responses (mixed model estimate: +17.0; p < 0.001; t = +2.808). The electrodes in the deep white matter did not show the difference in the high-gamma amplitude between vocalization conditions (mixed model estimate: +5.8; p = 0.232; t = +1.359).

The present study suggests that applying bipolar and Laplacian montages may be optimal for analyzing auditory-naming related high-gamma modulations recorded on SEEG. The present study shows that bipolar and Laplacian montages effectively reduce the signal deflection at contacts in the deep white matter and outside the brain parenchyma, as observed in the common average reference montage, and maintain the high-gamma augmentation in the gray matter. We assumed that in the deep white matter, bipolar and Laplacian montages mainly reduced the far-field potentials. They reduced amplitude changes to the same extent during the stimulus (21.2 and 64.5%, respectively) and response (31.5 and 85.4%, respectively) phases, independent of the presence of speech. A previous SEEG study showed that the white matter signal was mainly a mixture of volume conduction from nearby gray matter (17). SEEG studies analyzing cortico-cortical evoked potentials and spectral responses showed that bipolar and Laplacian montage reduced contaminating far-field potentials in the white matter and outside the brain parenchyma as in the present study (18, 30). We then assume that the reduced signal deflections outside the brain parenchyma mainly consist of EMG artifacts. Bipolar and Laplacian montages reduced amplitude changes by only 20.1 and 0.8% in the stimulation phase and 90.6 and 91.2% in the response phase, respectively. Contacts outside the brain parenchyma also showed a 17% reduction in amplitude change in covert responses. Previous SEEG studies showed the risk of muscle artifacts in the temporal surface region and lateral rectus muscle activity contamination in the temporal pole in the vicinity of the muscle (16, 37).

One may consider that bipolar montage would also reduce significant local activity across two electrodes (24). Nevertheless, the present study showed that bipolar montage reduced gray matter activity in the response phase by only 1.8%, less than deep white matter (24.0%) and outside the brain parenchyma (62.1%).

One may also consider that if MRI and CT can be accurately fused, there is no need to select a montage because electrodes outside the brain parenchyma or in the white matter can be visually excluded. However, the mean cortical thickness is about 2.5 mm, and previous studies reported that MRI was distorted by magnetic field inhomogeneity, resulting in a coordinate variation of about 2 mm compared to CT (38, 39). We believe that the selection of an appropriate montage would complement these image fusion pitfalls and help in the identification of cortical activity.

Our findings support the opinion that bipolar montage is suitable for SEEG interpretation in clinical practice (35, 40). We believe our findings will be useful in developing guidelines for functional brain mapping using high-gamma modulation in SEEG.

We assumed that covert response may be suitable for SEEG mapping in brain regions where the covert response elicits sufficient high-gamma augmentation. The advantage of covert response is that it makes it easier to identify gray matter activity by reducing signal deflection outside the brain parenchyma. In the present study, covert response reduced signal deflections outside the brain parenchyma independent of the montage setting during the response phase. Previous EEG and MEG studies mainly examined covert speech production or pre-vocalization processes to avoid the influence of EMG artifacts associated with vocalization (28). The disadvantage of a covert response is that it may only elicit low amplitude changes and risk missing the language-related areas. Previous ECoG studies showed that the covert response elicited lower high-gamma augmentation in the motor and language areas of the frontal lobe and lower sustained high-gamma augmentation in the sensory language area of the temporal lobe than did the overt response in word repetition tasks (41, 42). Functional areas identified by ECoG high-gamma language mapping using covert responses matched those identified by electrocortical stimulation, excluding the oral motor cortex (43). By exploring the physiological SEEG high-gamma modulation of each brain region for each vocalization condition, further research will determine the optimal vocalization conditions for each brain lobe and region, improving the accuracy of functional localization in the high-gamma language mapping.

The present study did not perform simultaneous EMG recordings of muscles for speech production. Jerbi et al. (16) simultaneously recorded electrooculograms and showed the saccade-related gamma band activity in SEEG. Further studies will support our results by showing a direct correlation between artifacts and EMG by simultaneous EMG recordings of muscles involved in speech production, such as the temporal muscle.

This study failed to determine whether bipolar or Laplacian montage is better in high-gamma language mapping. Each electrode montage has the pros and cons (18). Laplacian montage does not allow us to measure neuronal responses at the end of depth electrodes accurately. Bipolar montage may not tell us which pair of contacts is responsible for a given neuronal response of interest.

Most patients in this study did not undergo electrocortical stimulation mapping, making it difficult to determine the accuracy of high-gamma mapping in each montage with respect to electrocortical stimulation mapping. Further studies including patients who underwent both electrocortical stimulation mapping and high-gamma mapping are warranted to provide direct evidence for the superiority of bipolar and Laplacian montages in localizing language areas.

The present study failed to determine the appropriate montage for each language area [i.e., language comprehension-related and language production-related areas (44),] due to the difficulty in determining to which language area the electrodes outside the brain parenchyma belonged. Each language area shows different high-gamma modulation dynamics in the stimulus and response phases (14). Further studies comparing the results of electrocortical stimulation and high-gamma mapping at cortical electrodes in each language area and task phase may help determine the optimal montage and task phase to localize each language area.