Abstract
In contrast to scalp EEG, our knowledge of the normal physiological intracranial EEG activity is scarce. This multicentre study provides an atlas of normal intracranial EEG of the human brain during wakefulness. Here we present the results of power spectra analysis during wakefulness. Intracranial electrodes are placed in or on the brain of epilepsy patients when candidates for surgical treatment and non-invasive approaches failed to sufficiently localize the epileptic focus. Electrode contacts are usually in cortical regions showing epileptic activity, but some are placed in normal regions, at distance from the epileptogenic zone or lesion. Intracranial EEG channels defined using strict criteria as very likely to be in healthy brain regions were selected from three tertiary epilepsy centres. All contacts were localized in a common stereotactic space allowing the accumulation and superposition of results from many subjects. Sixty-second artefact-free sections during wakefulness were selected. Power spectra were calculated for 38 brain regions, and compared to a set of channels with no spectral peaks in order to identify significant peaks in the different regions. A total of 1785 channels with normal brain activity from 106 patients were identified. There were on average 2.7 channels per cm3 of cortical grey matter. The number of contacts per brain region averaged 47 (range 6–178). We found significant differences in the spectral density distributions across the different brain lobes, with beta activity in the frontal lobe (20–24 Hz), a clear alpha peak in the occipital lobe (9.25–10.25 Hz), intermediate alpha (8.25–9.25 Hz) and beta (17–20 Hz) frequencies in the parietal lobe, and lower alpha (7.75–8.25 Hz) and delta (0.75–2.25 Hz) peaks in the temporal lobe. Some cortical regions showed a specific electrophysiological signature: peaks present in >60% of channels were found in the precentral gyrus (lateral: peak frequency range, 20–24 Hz; mesial: 24–30 Hz), opercular part of the inferior frontal gyrus (20–24 Hz), cuneus (7.75–8.75 Hz), and hippocampus (0.75–1.25 Hz). Eight per cent of all analysed channels had more than one spectral peak; these channels were mostly recording from sensory and motor regions. Alpha activity was not present throughout the occipital lobe, and some cortical regions showed peaks in delta activity during wakefulness. This is the first atlas of normal intracranial EEG activity; it includes dense coverage of all cortical regions in a common stereotactic space, enabling direct comparisons of EEG across subjects. This atlas provides a normative baseline against which clinical EEGs and experimental results can be compared. It is provided as an open web resource (https://mni-open-ieegatlas.research.mcgill.ca).
Introduction
The scalp EEG during wakefulness in healthy individuals is well defined (Chang et al., 2011), but the knowledge accumulated on normal physiological intracranial EEG activity outside of the context of intracranial event-related potentials, task-related responses or intracranial stimulation (Jerbi et al., 2009; David et al., 2013; Matsumoto et al., 2017) is surprisingly scarce. This contrasts with the vast literature on epileptic activity in intracranial EEG (for a review see Frauscher and Dubeau, 2018). The knowledge of physiological intracranial EEG activity is based mainly on studies performed in the pre-MRI era using visual analysis of the recorded signals (Gastaut 1949; Jasper and Penfield, 1949; Petersen et al., 1953; Sem-Jacobsen et al., 1953, 1955, 1956; Chatrian et al., 1960 a, b; Ribstein, 1960; Perez-Borja et al., 1962; Graf et al., 1984). Evidence suggests that normal alpha, beta and gamma rhythms, and slow wave activity in the theta and delta range are observed in the intracranial EEG following a spatially distributed pattern. For instance, alpha rhythm was described in the occipital lobe, the parietal lobe just posterior to the postcentral gyrus, and the posterior temporal neocortex (Gastaut, 1949; Jasper and Penfield, 1949; Sem-Jacobsen et al., 1953, 1956; Chatrian et al., 1960a; Perez-Borja et al., 1962), and non-reactive alpha activity was found in the frontal lateral neocortex (Jasper and Penfield, 1949). Beta rhythm is most frequently seen over the anterior head regions. In particular, well sustained beta frequencies of ∼25 Hz are described in the precentral and postcentral gyri with lower frequencies present in the postcentral region (Jasper and Penfield, 1949). Other normal activities such as mu rhythm (Graf et al., 1984) and lambda waves (Chatrian et al., 1960a, b; Perez-Borja et al., 1962; Chatrian, 1976) have also been described.
Location-related differences in the occurrence of these rhythms and patterns remain purely descriptive without attempts to determine the exact localization of the positions of the electrode contacts in a common space. Furthermore, quantitative studies of normal activity during resting wakefulness are either lacking or are restricted to certain brain areas, often with small patient numbers. The best evidence exists for the hippocampus where an activity in the theta delta range has been described in quiet wakefulness (Brazier, 1968; Huh et al., 1990; Meador et al., 1991; Nishida et al., 2004; Moroni et al., 2012). In an attempt to create an atlas of normal human brain activity during wakefulness, Kahane (1993) described the quantitative EEG distribution across several cortical areas using intracranial EEG performed in epilepsy patients during presurgical epilepsy work-up. Coordinates of the individual electrode positions were given in Talairach space. In this work, the author already highlighted the methodological difficulties to determine normality in an epileptic brain.
Patients with refractory focal epilepsies are the only human subjects where extensive intracranial cortical EEG studies are undertaken, and thus allow access not only to pathological but also to normal brain neurophysiology. We have a poor knowledge of the normative electrophysiological data of brain activity. This is explained by the relatively rare placement of electrodes in healthy brain tissue and the challenge in identifying healthy brain regions, and by the difficulty of standardization of the electrode placement compared to scalp EEG, resulting in problems performing interindividual comparisons of EEG patterns (each patient has at most a few contacts in healthy brain, with great variability in electrode location from patient to patient). Moreover, even in large tertiary epilepsy centres, only a small proportion of patients (∼15–25 patients) are explored each year with intracranial electrodes.
The current work is a multicentre study aiming to provide an atlas of normal intracranial EEG activity recorded with both stereo-EEG electrodes and cortical grids/strips. These data are provided as an open web resource (https://mni-open-ieegatlas.research.mcgill.ca) developed under the LORIS framework (Das et al., 2016). Localization of the contacts in a common stereotactic space allows the accumulation and superposition of results from a large number of subjects and a direct comparison of EEG activity across subjects. The results of power spectra analysis during wakefulness are presented in this paper.
Materials and methods
Selection of intracranial EEG recordings
Charts and recordings of patients who underwent intracranial EEG investigation as part of their clinical evaluation for epilepsy surgery at the Montreal Neurological Institute and Hospital (MNI), Centre Hospitalier de l’Université de Montréal (CHUM), and Grenoble-Alpes University Hospital (CHUGA) were screened, starting with the most recent patients at time of data collection (September 2015 for MNI and CHUM, April 2016 for CHUGA), and moving consecutively backward to January 2010 or earlier, to identify cases with <40 recordings fulfilling selection criteria.
Recordings of patients included for this study had to fulfil the following four inclusion criteria:
presence of at least one channel with normal activity. Such channels are not common (on average 11% of channels per patient as shown in one of our previous studies) (Frauscher et al., 2015). A channel with normal activity is defined as a channel localized in normal tissue as assessed by MRI, is located outside the seizure onset zone, does not show at any time of the circadian cycle interictal epileptic discharges (according to the clinical report of the complete implantation and to a careful investigation of one night of sleep by a board-certified electrophysiologist), and shows the absence of overt slow-wave anomaly;
presence of peri-implantation imaging (CT or MRI) for exact localization of individual electrode contacts (contacts located in the white matter were excluded);
availability of a controlled intracranial EEG recording obtained after a minimum of 72 h after insertion of stereo-EEG electrodes or 1 week after placement of subdural grids or strips (medications are usually not yet lowered), and at least 12 h after a generalized tonic-clonic seizure, 6 h in case of focal clinical seizures, or 2 h in case of purely electrographic seizures, and not after electrical stimulation, as done in our previous work (Frauscher et al., 2015); and
sampling frequency of a minimum of 200 Hz. This minimum sampling frequency was chosen to include as many patients as possible for the analysis of frequencies in the classical Berger frequency bands (0.3–70 Hz).
The data collected in the three centres complemented each other well, as different implantation strategies are used (stereo-EEG electrodes versus subdural grids/strips, and frame-assisted versus frameless electrode placement), and provide a broad coverage of different cortical regions. Ethical approval was granted at the MNI as lead ethics organization (REB vote: MUHC-15-950).
Co-registration and anatomical localization of electrodes and electrode contacts
Registration to stereotaxic space and anatomical localization of electrodes was performed using minctools (www.bic.mni.mcgill.ca/ServicesSoftware/Services SoftwareMincToolKit) and the IBIS framework (Drouin et al., 2017). Peri-implantation CT/MRI images showing electrode positions were linearly registered to pre-implantation MRI images (preMRI) of each patient. In turn, preMRIs were non-linearly registered to the ICBM152 2009c non-linear symmetric brain model (Mazziotta et al., 2001; Fonov et al., 2011). The combined transformation allowed estimation of the electrode positions in a common stereotaxic space in order to visualize the standardized atlas of normal EEG activity.
Anatomical structures were fully automatically segmented. We used an atlas created with an unbiased method (Fonov et al., 2011) from manually segmented data of 20 normal subjects (Landman and Warfield, 2012). It consists of 132 grey matter labels (66 per hemisphere after excluding basal ganglia, cerebellum, midbrain and brainstem). We chose this atlas because it was detailed and contains deep as well as surface cortical regions. The preMRI was non-linearly registered to the atlas’ averaged MRI and then the inverse transformation was applied to warp the labels back to the patient’s space. In case of lesional epilepsy, segmentations were checked visually by an epileptologist in order to ensure that selected electrode contacts are indeed outside of lesional tissue. Patients with large cortical malformations such as extended polymicrogyria, hemimegalencephaly, agenesis of corpus callosum, or extended encephalomalacic lesions were excluded.
Each bipolar channel was represented by a volume made of a cylinder of length equal to the distance between its two electrode contacts, 10 mm diameter and one half sphere at each end. The anatomical localization of each channel was estimated by assessing in which segmented grey matter region laid the majority of its volume. The findings were grouped according to the topographical localization of the channels and plotted onto the standard ICBM152 template.
Selection of EEG sections
We visually selected 60 s sections (either continuous or consecutive discontinuous >5 s segments after artefact exclusion; 5 s were chosen as a compromise between minimizing the number of segments and obtaining the 60 s sections) of resting wakefulness EEG with eyes closed recorded during standardized conditions. The controlled ‘eyes closed and eyes opened’ recordings were part of the controlled intracranial EEG evaluation at the three participating sites. Analysis of EEG activity of the non-epileptic physiological channels was performed using Matlab (Mathworks). The signals were bandpass filtered at 0.5–80 Hz and downsampled to 200 samples per second if the original sampling rate was higher (original sampling rates were 200, 256, 512, 1000, 1024, and 2000 samples per second).
Analysis of the oscillatory component of the signal
The spectral density in each channel was estimated with Welch’s method, i.e. averaging the magnitude of the discrete time Fourier transform of 59 overlapping blocks of 2 s duration and 1 s step, weighted by a Hamming window. In each channel the resulting spectral density was normalized to a total power equal to one, making it independent of the EEG signal amplitude.
To determine the presence of peaks in the spectrum, we first defined the ‘no-peak set’ as a set of channels with no peak in their spectrum. We defined this set in a data-driven procedure. The normalized spectra were classified into k-groups using the k-means algorithm, with 160 features given by the values at each frequency step (0.5 Hz steps between 0.5 and 80 Hz), Euclidean distance, and 100 repetitions. This separated channels according to their most prominent peaks in their spectrum. By repeating the classification with increasing number of groups k, eventually a group without peaks was found. We determined the presence of this group by requiring that its mean normalized spectrum be lower than the maximum among the other groups. From this group, we took the 50% of the channels closest to its mean to define the no-peak set.
Because as the frequency increases, the spectrum becomes more correlated between equally spaced frequency steps, we defined a set of unequal frequency intervals in which to test for the presence of peaks in the spectrum. The intervals were selected so that each of them had at least 4% of the power of the average normalized spectrum of all the studied channels. This resulted in 22 frequency intervals (0.5–0.75–1.25–1.75–2.25–3.25–3.75–4.25–5.25–6.25–6.75–7.75–8.25–9.25–10.25–11.75–13.25–15.25–17.25–20.25–24.25–31.75–80 Hz). The presence of a peak at a given frequency interval in any particular brain region was determined by comparing the distribution of the normalized spectra of all the channels in the region at the desired frequency interval, to the distribution of the no-peak set at the same frequency interval. A one-sided two-sample Kolmogorov-Smirnov test was used, at 5% significance level with Dunnett’s correction for multiple comparisons (for the number of tested anatomical regions times the number of frequency intervals). The two-sample Kolmogorov-Smirnov test is sensitive to differences in location and shape of the empirical cumulative distribution functions. A second test was performed for brain regions and frequency intervals in which significant differences were found, to determine the percentage of channels in the brain region of interest that had a significantly higher relative power than the no-peak set at the given frequency interval. We performed an uncorrected one-sided Wilcoxon rank-sum test (at 5% significance level), comparing the distribution of the no-peak set to the spectral density of each individual channel.
Analysis of the non-oscillatory component of the signal
To investigate the non-oscillatory component of the spectrum, we explored the scale-free dynamics of the brain activity by fitting a 1/fβ model to the high frequency end of the spectrum. We adapted the methodology used in previous studies (Yamamoto and Hughson, 1993; He et al., 2010). We used non-overlapping Hamming windows of 1 s duration for the computation of the spectrum of the scale-free activity, and fitted the β coefficient in the 10–40 Hz range. We adopted a low end of 10 Hz instead of 1 Hz as in He et al. (2010), because we observed that in most of the channels the scale-free component of the spectrum had a shoulder between 1 and 10 Hz. The upper end of the range (40 Hz) is limited by the sampling rate of our data (the highest frequency not affected by the antialiasing down sampling filters is 80 Hz, but the computation of the scale-free component of the spectrum is correct only up to half this frequency). To determine if some regions had a different β coefficient, we computed the median value of β and performed a permutation test (100 000 permutations of region labels, Bonferroni corrected for 38 regions).
For exploring the nested frequencies in the brain activity, we performed a phase-amplitude coupling analysis. We computed the modulation index (Tort et al., 2008) for phases at frequencies between 0.5 and 12 Hz, with 0.5 Hz steps. We explored the coupling to amplitude in 17 frequency bands (4–5–6–7–8–10–12–14–16–20–24–30–36–42–48, 52–58, 62–70–80 Hz). For computation of the modulation index we used eight phase bins. The Hilbert transform was implemented between 0.5 and 80 Hz, with an FIR filter of order 200. The bandpass filters were elliptic IIR filters of order 6 applied in two passes to achieve zero-phase. To determine the statistical significance of the observed modulation index, a set of 1000 surrogate amplitude signals were analysed for each channel, by partitioning the original time series in eight sections of random length and shuffling them. The mean and standard deviation (SD) of these surrogate data were used to determine the z-score of the modulation index. The proportion of channels with significant modulation index at different phase-frequency bins was computed, and a permutation test was performed to determine regional differences (100 000 permutations of region labels, Bonferroni corrected for 24 × 17 phase-frequency bins).
Results
Here we report on the cortical coverage for this project, the different electrode types used, as well as the automatic classification of the no-peak set, which is used subsequently to test for the presence of peaks in the spectrum of channels from different brain regions. We then describe the lobar differences in spectral density distribution, highlight the intracranial EEG signature of some cortical brain regions, stress unexpected findings, and analyse the non-oscillatory component of the signal.
Demographical information of the study sample
The study sample consisted of 106 patients (54 males; CHUGA: 49 patients, MNI: 40, CHUM: 17) with therapy-refractory focal epilepsy who underwent invasive intracranial EEG investigation during presurgical epilepsy work-up. The mean age was 33.1 ± 10.8 years. Eighty-nine patients (84%) were investigated with stereo-EEG electrodes, and 17 (16%) with grids/strips. Of the patients who were explored with grids/strips, all but two had additional stereo-EEG electrodes for exploring deep structures such as the insula or the mesiotemporal lobe.
Number and density of cortical channels with normal EEG activity
A total of 1785 intracranial EEG channels (left hemisphere, 1066; right hemisphere, 719) recording presumably normal physiological brain activity were identified. The positions of the investigated channels after co-registration in stereotaxic space are given in Fig. 1. There were on average 0.9 channels per cm2 of cortical surface, and 2.7 channels per cm3 of cortical grey matter volume. Restricting the channels to have an equal number of channels per region in both hemispheres (695 channels), we found no differences in the distribution of normalized spectral density between the channels from homologous regions of both hemispheres (two-sided two-sample Kolmogorov-Smirnov test with Dunnett’s correction for 22 comparisons for the tested frequency intervals). Therefore, we grouped left and right hemisphere channels together. The figures show all channel positions flipped to one of the hemispheres of the symmetric atlas.
Localization of the 1785 EEG channels with normal physiological activity analysed for this study. The 1520 channels from stereo-EEG electrodes are visualized in blue, and the 265 channels from cortical grids and strips are in yellow. Note that for the ‘inflated’ brain display at the bottom, the electrodes are projected on the cortical surface.
The spectral density is similar in stereo-EEG electrodes and cortical grids/strips
Eighty-five per cent of the 1785 channels (1520) were from stereo-EEG electrodes, and 15% (265) from cortical grids and strips (Fig. 1). To compare the spatial coverage achieved with intracerebral electrodes and with cortical grids and strips we classified the cortical surface according to curvature, and compared regions with convex curvature (likely gyri) to regions of concave curvature (likely sulci). We found that grid channels were 5.3-times more frequent in the former, and stereo-EEG electrode channels 1.6-times more frequent in the latter (180 versus 34 and 412 versus 673 in convex versus concave regions, respectively; the remaining channels were in regions with no clear dominant curvature). This indicates that channels from cortical grids/strips are typically placed atop of gyri, and that stereo-EEG electrodes record not only from the sulci, but also from a significant fraction of the gyri. The normalized spectral density of the neocortical channels recorded with stereo-EEG (commercial DIXI electrodes, home-made MNI electrodes, or Ad-Tech electrodes) or cortical grids/strips (Ad-Tech electrodes) is similar across electrode types (Fig. 2A).
Comparison of different electrode types (stereo-EEG electrode types: DIXI CHUGA, DIXI MNI, MNI MNI, AdTech CHUM, and grids: CHUM). (A) Median spectral density for each electrode type. (B) Signal amplitude (and number of channels) according to electrode type and centre. Note the ∼100% increase in absolute amplitude (root mean square, RMS) in channels from cortical grids/strips compared to stereo-EEG electrodes. CHUGA = Centre hospitalier de l’université de Grenoble-Alpes; CHUM = Centre hospitalier de l’université de Montréal; MNI = Montreal Neurological Institute.
Interestingly, there is a ∼2-fold increase in absolute amplitude (root mean square) in channels from cortical grids/strips compared to stereo-EEG electrodes (Fig. 2B). The difference is maintained when only neocortical channels are analysed (i.e. excluding mesiotemporal lobe structures). This difference does not affect subsequent results, as normalized spectra are used.
Automatic classification of channels and no-peak set
The procedure to separate a set of channels with no peaks in the spectrum from sets of channels with various peaks resulted in eight channel groups (no peak, one group and various peaks, seven groups). The median normalized spectral density of the groups can be observed in Fig. 3A, and the groups with spectral peaks correspond to peaks in the low delta band, high delta band, theta band, alpha band, low beta band, high beta band, and gamma band. Channels with delta or gamma activity usually do not show sustained rhythms, but simply an increase in the low or high frequency activity. The spectral density of what we define as ‘the no-peak set’, derived from the group of channels with no peak in their spectrum, is shown in Fig. 3B. This no-peak set is used in subsequent sections to test for the presence of peaks in the spectrum of channels from different brain regions. Figure 3C and D shows the distribution of the groups in the inflated cortex. Note that even though the frequency resolution of this unsupervised classification is poor, it shows a frequency increase from the posterior to the anterior cortical regions.
Unsupervised classification of channels and no-peak set. (A) Median normalized spectral density of the eight different groups obtained by the classification. Channels with delta or gamma activity usually do not show sustained rhythms (no peaks), but simply an increase in the low or high frequency activity. (B) The spectral density of the no-peak set, derived from the group of channels with no peak in their spectrum. (C and D) Distribution of the groups in the inflated cortex (lateral and mesial views). Even though the frequency resolution of this unsupervised classification is poor, the figure gives a rough global idea of the distribution of rhythms in the brain, and shows a posterior to anterior gradient of increasing frequencies.
Lobar frequency differences in intracranial EEG
Figure 4 shows the difference in the spectral density of the standard EEG frequencies across the different brain lobes (758 frontal channels, 415 temporal, 355 parietal, and 105 occipital). The occipital lobe shows, as expected, a clear peak in alpha activity (significant in the 7.75–10.25 Hz, in 47% of the channels at ∼10 Hz), whereas faster activity peaks in the beta and gamma band were recorded in the frontal lobe (16.75–80 Hz, in 48% of the channels at ∼24 Hz). The parietal lobe shows peaks at intermediate frequencies in the alpha (8.75–9.75 Hz, in 21% of the channels) as well as in the beta band (16.75–31.75 Hz, in 24% of the channels at ∼20 and 31.5 Hz). The temporal lobe shows a borderline alpha peak (7.25–9.25 Hz, in 30% of the channels at ∼8 Hz), and some channels show a peak in delta activity (0.75–2.25 Hz, in 24% of the channels at ∼1 Hz). Figure 4B shows the difference in the distribution for some of these cases, in which the difference is not clearly appreciated by looking at the median spectral density in Fig. 4A.
Lobar differences in EEG frequences. (A) Spectra of the different brain lobes in semi-logarithmic graph. The red line corresponds to the median spectral density of all the channels in the region. The 25 and 75 percentiles are indicated by the pink shaded region. The broken red lines show the upper and lower bounds of the spectral distribution at every frequency. The thin black line shows the median spectrum of the no-peak set used to determine the presence of peaks. Vertical black lines separate the common clinical frequency bands indicated by Greek letters. The coloured horizontal segments in the upper part of each graph indicate the presence of peaks. If the segment is present, it indicates that the distribution of the channel spectral densities is significantly higher than the distribution of the no-peak set. The colour of the line indicates the percentage of channels that have a significant deviation compared to the no-peak set at each frequency. (B) Illustration of differences not clearly visible in A. Each panel shows in black the distribution of the energy in the no-peak set for the given frequency interval, and the red line shows the distribution in different brain regions (frontal lobe and temporal lobe). The broken vertical lines indicate the median value, and the coloured area under the curve indicates the percentage of channels that has significantly higher power than the no-peak set in the corresponding frequency interval. The figure illustrates situations in which the test can find differences, not always obvious looking only at the median (in A). Left: Different position (mean, median) in the gamma band of the frontal lobe. Middle: Equal median but different dispersion (SD) in the delta band of the temporal lobe. Right: Equal median but asymmetric distribution (skewness) in the alpha band of the temporal lobe.
Some cortical brain regions have a specific EEG signature
We analysed the spectrum of channels in different anatomical brain regions. Some neighbouring anatomical regions of the MICCAI atlas were joined in order to attain at least five channels per analysed region. We formed eight regions joining channels from 20 regions (the entorhinal cortex had no electrodes): (i) superior and middle occipital gyri; (ii) inferior occipital gyrus and occipital pole; (iii) lingual gyrus and occipital-fusiform gyrus; (iv) postcentral gyrus and its medial segment; (v) gyrus rectus and anterior, medial, and posterior orbital gyri; (vi) superior frontal gyrus and frontal pole; (vii) temporal pole and planum tempolare; and (viii) fusiform and parahippocampal gyri. We analysed 38 cortical regions (median 40.5 channels per region, range 6–178 channels). Thus, the total number of analysed regions used in the multiple comparison correction was 42, the 38 cortical regions, plus the four lobes presented in the previous subsection.
Power spectra of all investigated cortical regions are provided in the Supplementary material. Brain regions such as the posterior insula, the anterior cingulate, and others did not show deviations from the no-peak set. In contrast, other brain regions demonstrated a specific signature: significant peaks present in more than 60% of the channels within a given region were found in the precentral gyrus (lateral segment: 64% of 123 channels at ∼20–24 Hz; medial segment: 72% of 18 channels at ∼24–31.5 Hz), the opercular part of the inferior frontal gyrus (72% of 39 channels at ∼20–24 Hz), the cuneus (68% of 19 channels at ∼8 Hz), and the hippocampus (72% of 36 channels at ∼1 Hz) (Fig. 5).
Spectral density plots of cortical regions showing a significant spectral density peak in >60% of investigated channels..
Some cortical areas have more than one spectral peak within the same channel
From all analysed channels, 141 (7.9%) had more than one peak in the spectrum (comparing each channel to the no-peak set, without correcting for multiple comparisons). The percentage of channels with two or more peaks was significantly higher than the average of 7.9% in eight brain regions (cumulative binomial distribution with parameters theta = 0.079 and k = number of channels in the region, uncorrected). Most of these regions are sensory/motor regions: medial segment of the precentral gyrus (5/18 or 28% of channels), mid-cingulate region (8/40 or 20%), supplementary motor cortex (11/47 or 23%) and the cortex anterior to the supplementary motor cortex (3/16 or 19%), transverse temporal gyrus (3/14 or 21%), inferior occipital gyrus (4/20 or 20%), precentral gyrus (23/123 or 19%), and postcentral gyrus (11/65 or 17%).
In one-third of these channels (48/141) the second peak is at twice the frequency of the first (second harmonic), and in 8% (11 channels) there was a significant peak at the third harmonic. Half of the channels with second harmonics (24/48) correspond to the precentral gyrus, its medial segment, or the postcentral gyrus. Nine of the 11 channels with significant third harmonic belong to these regions.
The statistical test to determine if there are two significant peaks at different frequencies requires a multiple-comparison correction for a high number of 231 frequency pairs; therefore the results are not significant for the available number of channels.
Unexpected findings
In this section, we present some highlights of the findings for a few brain regions. For findings from all 38 brain regions, see the Supplementary material as well as the raw data, which are made available through our webpage (https://mni-open-ieegatlas.research.mcgill.ca).
The anterior insula cortex shows a beta peak
The insula cortex did not show a significant difference compared to the no-peak set. When subdividing the insula into an anterior and posterior portion, we found a peak in beta activity for the anterior insula in 34% of 71 channels (∼20 Hz). There was no significant difference in the spectral power density between the posterior insula and the no-peak set (Supplementary material).
The middle cingulate gyrus shows a beta peak
The middle cingulate gyrus showed a beta peak (40% of 40 channels ∼24 Hz, range: 19.75–24.25 Hz), but there was no difference of the power spectral density compared to the no-peak set for the anterior and posterior cingulate gyrus (Supplementary material).
Generation of alpha activity in occipital, parietal and temporal lobes
Alpha activity was present in a large portion, but not all regions of the occipital lobe (Fig. 6). The cuneus was the region with the highest number of channels showing a clear alpha peak (68% of 19 channels ∼8 Hz, range: 7.75–9.75 Hz). This was even more expressed than in the calcarine cortex (58% of 12 channels ∼9 Hz, not significantly different than the no-peak set, presumably because of the relatively low number of channels in the region). There was no significant alpha peak in the inferior occipital gyrus.
Presence of peaks in alpha activity across the brain. (A) Power spectral density plots of brain regions in the occipital lobe. (B) Power spectral density plots of brain regions outside the occipital lobe with significant alpha activity.
Apart from the occipital lobe, alpha activity was present, albeit in a lesser degree, in the superior parietal lobule (38% of 53 channels ∼9 Hz, range: 8.25–9.25 Hz) and in some temporal regions, but at a lower frequency and at the limit of the traditional theta and alpha bands. These findings were significant for the superior temporal gyrus (35% of 79 channels ∼8 Hz, range: 6.75–9.25 Hz) and the fusiform and parahippocampal gyri (49% of 45 channels ∼8 Hz, range: 6.75–9.25 Hz) (Fig. 6). No significant peak in alpha activity was seen in the middle and inferior temporal gyrus, the temporal pole and planum temporale, the transverse temporal gyrus and the mesial temporal lobe structures, or in the angular gyrus, supramarginal gyrus, the pars opercularis, precuneus and posterior cingulate gyrus.
Some regions show significant peaks in delta activity during wakefulness
The most expressed peak in the delta range was found for the hippocampus (72% of 38 channels ∼1 Hz, range: 0.5–3.25 Hz). Other regions with a delta peak were the inferior occipital gyrus and occipital pole (48% of 23 channels ∼1.5 Hz, range: 1.25–3.25 Hz), the angular gyrus (38% of 53 channels ∼1 Hz, range: 0.5–1.25 Hz), the medial frontal cortex (42% of 19 channels ∼1 Hz, range: 0.5–1.25 Hz), the gyrus rectus/orbital cortex (42% of 45 channels ∼1 Hz, range: 0.5–1.75 Hz), and the middle temporal gyrus (31% of 129 channels ∼1 Hz, range: 0.5–1.75 Hz) (Fig. 7).
Power spectral density plots of brain regions with significant delta activity..
Analysis of the non-oscillatory component of the signal
We explored the scale-free brain activity by fitting a 1/fβ model to the high frequency end of the spectrum (10–40 Hz). We observed a median value of the β coefficient of 2.29 (range 0.68–4.26; mean 2.35; SD 0.52). We found different median β coefficients at a 5% significance level for the following regions: the β coefficient was higher (steeper decrease of the spectrum for increasing frequency) in the fusiform and parahippocampal gyri (45 channels, median β = 2.84, corrected P = 0.0004) and in the inferior occipital gyrus and occipital pole (23 channels, median β = 2.78, corrected P = 0.024), whereas it was lower in the middle frontal gyrus (178 channels, median β = 2.11, corrected P = 0.001). Figure 8A shows the distribution of the β coefficient in the different brain regions.
Analysis of the non-oscillatory component of the signal. (A) Distribution of the β coefficient of the 1/fβ model of the scale free component of the spectrum (10–40 Hz), i.e. the spectrum after removing peaks associated to oscillatory activity. The asterisks indicate regions with coefficient significantly different from the rest. (B) Percentage of channels showing statistically significant phase-amplitude coupling at different frequencies (uncorrected). (C) Percentage of channels showing statistically significant phase-amplitude coupling at different frequencies (corrected for false discovery rate of 5%).
Regarding the nested-frequencies analysis, we found a low proportion of channels with significant phase-frequency coupling, likely because of the limited duration of the recordings. Figure 8B shows the proportion of channels with statistical significance at 5% without correcting for multiple comparisons (330 phase-amplitude bins), and Fig. 8C corrected for false discovery rate of 5%. In both cases the highest proportion is found between low delta (0.5 Hz) and low alpha band (8–10 Hz). But the proportion of channels in which this coupling was significant is low after correcting for multiple comparisons (28 channels, 1.5% of the 1785 available channels). We found no significant differences between regions.
Discussion
This is the first atlas of normal intracranial EEG of the adult human brain. Data from physiological channels of many patients with epilepsy were collected to overcome the limited sampling of the brain with intracranial electrodes and the unusual recording from presumably normal brain regions. Furthermore, to superimpose the results from all subjects on one brain, intracranial electrode positions were standardized in a common 3D environment using a stereotactic template providing the exact coordinates of each recording contact. This atlas of normal brain activity will aid in the better differentiation between physiological and pathological brain activity for clinical work and research in humans. In this paper, results of the quantitative analysis obtained from a large population of human subjects are illustrated. This atlas will be an open resource available for consultation on the web developed under the LORIS framework (Das et al., 2016).
Selection of channels from presumably normal brain regions
We are aware that selecting ‘true’ normal healthy cortex represents an inaccurate task, since we essentially analyse EEGs from the brains of epileptic patients. In most of these patients, however, some electrodes are placed in non-epileptogenic zones devoid of structural or physiological anomalies. These electrodes are necessary for comparison of cortical physiology and help to define the limits of the eventual surgical resection. Some presumably normal superficial neocortical regions are also recorded as a result of the need to reach a deep structure with a multi-contact electrode. Selecting the most normal brain regions in these patients is as close as we can ever get to a ‘true’ atlas of the normal brain. To ensure as much as possible the selection of channels with normal physiological EEG activity, we followed a strict protocol that involved the consensus of two epileptologists for selection of all the imaging and neurophysiological data for each subject (see ‘Materials and methods’ section). Even if a few channels from pathological regions might have slipped through our careful screening process, it is unlikely to have occurred for many, and the large number of channels in each region makes it likely that our average results are representative of the healthy brain.
Intracranial EEG data from different brain regions
This atlas extends our knowledge on cortical neurophysiology in humans by providing quantitative and accurate EEG data for 38 different brain regions. In each region, an average of 2.7 channels per cm3 of cortical grey matter was used for EEG recording. The best coverage was obtained in the areas most frequently explored during epilepsy surgery evaluation, the fronto-temporo-parietal regions. Less frequently explored areas such as the occipital cortex were, however, sufficiently covered. We were also able to identify channels with physiological activity in the mesiotemporal lobe structures, an area often explored, but usually showing epileptic activity.
Superimposition of all channels in a common stereotactic 3D environment
The lack of standardization in electrodes insertion and placement results in problems performing inter individual comparisons of EEG patterns and, hence, obtaining normative electrophysiological data. To overcome this problem, we used a common 3D environment (IBIS) for semi-automatic co-registration and superimposition of the results from all electrodes from all subjects on one brain template (Mazziotta et al., 2001). To avoid potential bias by a human scorer in the assignment of electrode contacts to anatomical regions, anatomical structures were automatically segmented using an atlas created with an unbiased method (Fonov et al., 2011). This atlas provided detailed volumetric tissue classification of cortical and deep structures (e.g. hippocampus, amygdala). Full brain volume coverage with 66 labels per hemisphere allowed proper assignment of electrode contacts to the anatomical region from which they most likely record.
Comparison of stereo-EEG electrodes and cortical grids or strips
Compared to scalp EEG, intracranial recordings are not standardized and may show interindividual variations. Stereo-EEG electrodes are in direct contact with the neuronal generators, with a spatial organization that is highly variable and does not follow a standardized electrode positioning like for the 10-20 or 10-10 International EEG System (Jasper, 1958; Nuwer et al., 1998). In contrast, grids/strips have the same distance to the cortical surface and are therefore in the same position with respect to the cellular organization of the cortex.
Our study demonstrated that there was no difference in the spectral power distribution between EEG recordings performed with stereo-EEG and cortical grids/strips. The absolute amplitude was two times higher using cortical grids/strips as compared to stereo-EEG. This might be because of the location and orientation of the electrodes with respect to the cortical sources. In case of grids/strips the electrode contacts are always localized atop of the cortex, whereas there are different orientations in stereo-EEG regarding the cellular organization of the cortex.
Region-specific EEG signatures as assessed with spectral power analysis
After discussing methodological issues we now discuss the main findings. The present work aimed at identifying rhythms characteristic for the different brain regions. To do so we calculated spectral density plots for the different regions investigated. Significances in spectral frequencies were calculated by performing a comparison with a group of automatically selected channels exhibiting no spectral peaks (‘no-peak’ channels). This is a first step in analysing the available data and identifying region-specific characteristics to improve our knowledge on the cortical structural-functional interrelations. Indeed, different methods could have been used. The availability of the present data will allow others to apply different methods and provide further characterization of the human normal iEEG.
Distribution of rhythmic activity across different brain regions
Current evidence suggests that rhythms are important for the communication between different brain regions (Steriade, 2006; Buzsaki and Schomburg, 2015). To study such rhythms and interactions systematically, it is important to have knowledge of the basic rhythms of each brain region at rest, which is provided by the current atlas. A recent study using MEG in 22 healthy subjects investigated whether rhythmic brain activity is characteristic for different cortical areas. The authors found that individual human brain areas can be identified from their characteristic spectral activation patterns. Moreover, clustering of brain areas according to similarity of spectral profiles reveals known brain networks (Keitel and Gross, 2016). Data of the present atlas could be used to validate these results, and extend them because of the difficulty of MEG to assess deep sources. Another recent study examined brain rhythms of the sensorimotor region in 20 healthy subjects using 64-channel scalp EEG. This study is interesting, although limited to the central region only, as it is a first step towards the identification of cortical areas based on neuronal dynamics rather than on cytoarchitectural features (Cottone et al., 2017). An extension of this work could be to apply this methodology to the intracranial atlas.
Focal generation of alpha activity
Early work in the canine model demonstrated that the alpha rhythm is generated in the cerebral cortex (Lopes da Silva and Storm Van Leeuwen, 1979). These data obtained by invasive EEG are complemented by more recent non-invasive studies using source analysis showing that the alpha rhythm is due to more than one generator in the posterior cortex. Moreover, a thalamic modulating influence has been suggested (Tyvaert et al., 2008; Chang et al., 2011). Findings from electrocorticography and intracerebral depth EEG studies described alpha activity in the occipital lobe, the parietal lobe just posterior to the postcentral gyrus, and in the posterior temporal neocortex (Frauscher and Dubeau, 2018). Our present work confirmed in part these findings. Interestingly and in contrast to what we initially suspected, a clear peak in alpha activity was not found in all occipital lobe structures; it was most prominent in the mesio-occipital lobe and in particular in the cuneus. Alpha activity was further seen in the superior parietal lobe, as well as the temporal lobe. In the latter, alpha activity was of a lower frequency compared to the occipital lobe. Moreover, there was a peak in alpha activity in the postcentral gyrus. We did not find alpha activity in the frontal lobe, as previously described (Jasper and Penfield, 1949).
Peak in beta activity in motor function-related cortical areas
In the motor system, oscillations in the beta frequency band have been suggested to play a role in sensorimotor integration (Baker, 2007). Of note, a vast majority of motor function-related eloquent neocortical areas such as the precentral gyrus, the supplementary motor cortex, the mid-cingulate, the anterior insula, the pars triangularis of the inferior frontal gyrus, and the operculum showed a significant peak in the beta frequency band during resting state. This beta peak was most evident in the precentral gyrus, as suggested by previous work during acute electrocorticography (Jasper and Penfield, 1949). In line with our findings, studies examining the execution of voluntary movements showed, in primary sensorimotor areas, a desynchronization of beta activity with the motor response, followed by the appearance of an event-related synchronization in the gamma range between 40 and 60 Hz, and reappearance of beta activity (Sem-Jacobsen et al., 1956; Pfurtscheller and Lopes da Silva, 1999; Szurhaj and Derambure, 2006).
Presence of frequency peaks in the different bands depends on the brain region
A new finding of this study is that there is more than one peak in certain brain regions, even within the same EEG channel. One example is the precentral gyrus. All channels in this cortical region showed more than one peak in the delta, theta and high beta range. Previous work using acute electrocorticography showed a fast beta rhythm of 25 Hz similar to that in our study in the precentral area (Jasper and Penfield, 1949), but did not refer to underlying slower frequency bands. We further confirmed peaks in alpha and beta activity in the postcentral gyrus. This finding is in line with the notion that there is less beta activity in the postcentral gyrus when performing a direct comparison to the precentral gyrus (Jasper and Penfield, 1949).
It is noteworthy that there are certain brain regions, particularly in the temporal lobe or in deep structures such as the anterior and posterior cingulate gyrus as well as the posterior insula, which did not show any difference (statistically significant or perceptible visually) with the set of channels with no identifiable spectral peaks. For other regions, such as the calcarine cortex, there was a clear peak in certain frequency bands, which, however, was not significant, most likely because of the relatively small number of investigated channels.
Presence of focal cortical peaks in the delta band in the awake brain
Slower frequencies in the theta delta range were found in frontobasal, as well as frontomesial and temporomesial structures. Delta activity in the ventral medial frontal brain portions was already suggested by Sem-Jacobsen et al. (1955) and for the orbitofrontal cortex by Nishida et al. (2004). Slow wave activity in the hippocampus, consisting of irregular slow waves in the delta and theta frequency range was also already reported in the past (Jasper and Penfield, 1949; Brazier, 1968). Interestingly, previous studies showed that the EEG activity in the hippocampus is task-dependent (Bódizs et al., 2001; Cantero et al., 2003; Clemens et al., 2009; Lega et al., 2012; Moroni et al., 2012; Watrous et al., 2013; Billeke et al., 2017). Findings of the present study are in line with the described ‘large irregular activity’ characterized by hippocampal sharp waves occurring during immobility. In this context, it is noteworthy to stress that in humans, this slow wave activity is in the delta range; in animals it was shown to be in the theta range (Bódizs et al., 2001; Clemens et al., 2009; Watrous et al., 2013).
Analysis of the non-oscillatory component of the signal
The classical frequency bands in EEG are described based on the various oscillatory rhythms that appear in the EEG. Apart from this oscillatory component, the EEG has a significant, if not major, component of non-oscillatory or scale free activity. We analysed both components and observed differences in the non-oscillatory EEG signal in different brain regions, with higher values in the fusiform and parahippocampal gyri, as well as the inferior occipital gyrus and occipital pole. All these regions are involved in visual processing. Interestingly, this is in keeping with a study analysing scale free activity with functional MRI that showed increased values for the visual cortex (He et al., 2010). We further analysed our data for nested frequencies. We found the strongest phase amplitude coupling between the phase of very low frequency activity at 0.5 Hz and the low alpha band amplitude. The low proportion of channels with significant coupling compared e.g. to He et al. (2010), might be explained by the short length of the recordings in the awake/eyes closed condition.
Conclusion and future prospects
This is the first atlas of normal intracranial EEG activity in a common stereotactic space that allows the accumulation of small amounts of data from each of a large group of subjects, to obtain an atlas covering well most cortical regions. It will aid the neurophysiologist to understand the EEG frequency distributions across the different cortical regions of the human brain better. This atlas provides a normative baseline against which clinical EEGs and experimental results can be compared. It will be an open resource available for augmentation and consultation on the web (https://mni-open-ieegatlas.research.mcgill.ca). We are currently preparing an atlas on sleep, as sleep data are available from most of the subjects and we will be able to compare wake and sleep across the different investigated regions. Also, we still have some regions that are not fully covered, but we intend to allow the atlas to grow by inclusion of data from the original centres as well as from other qualified centres, and to continue the current project in a prospective way. This will also allow to collect longer segments of controlled awake/eyes closed condition for different types of analysis, such as that of nested frequencies.
Acknowledgements
The authors wish to express their gratitude to Louis Collins, PhD, Vladimir Fonov, PhD, as well as Alan Evans, PhD and the LORIS team from the McConnell Brain Imaging Centre of McGill University, as well as the staff and technicians at the EEG Department at the Montreal Neurological Institute and Hospital, Lorraine Allard, Nicole Drouin, Chantal Lessard and Linda Ménard, the staff and technicians at the Neurophysiopathology Laboratory of Grenoble-Alpes University Hospital, Dr. Anne-Sophie Job-Chapron, Patricia Boschetti, and Marie-Pierre Noto, and the staff and technicians at the Centre Hospitalier de l’Université de Montréal.
Funding
This work was supported by the Savoy Epilepsy Foundation (project grant to B.F. and post-doctoral fellowship to R.Z.), the Botterell Powell’s Foundation (grant to B.F.), and the Canadian Institute of Health Research (grant FDN-143208 to J.G.).
Conflicts of interest
The authors have no potential conflict of interest with the present study. Outside of the submitted work, B.F. received speaker/advisory board fees sponsored by UCB. J.G. and N.v.E. have received fees for consultancy from Precisis Inc. P.K. received speaker/advisory board fees from UCB and Eisai. D.K.N., D.H., F.D., J.H., L.M., A.O., and R.Z. have nothing to declare.
Supplementary material
Supplementary material is available at Brain online.
