Task‐related gamma‐band dynamics from an intracerebral perspective: Review and implications for surface EEG and MEG

The Clinical Setting

In patients with pharmacologically resistant epilepsy, iEEG is used to identify cortical regions critical for seizure onset and identify others that need to be spared at the time of surgery [e.g., Kahane et al., 2004]. The intracerebral electrodes sometimes stay in place for more than two weeks in order to localize the origin of fast electrophysiological rhythms that precede seizure onset and that are at the core of the epileptogenic network. Surgical resection of the focus of the seizure has been shown to yield up to 70% success rates for drug‐resistant temporal lobe epilepsy [Engel, 1996]. Intracranial EEG is used to test one or several hypothesis regarding the anatomical organization of the epileptogenic network. This sometimes implies that intra‐cerebral electrodes are positioned in widely distributed brain regions that might include pathological but also healthy tissue. Therefore, such a clinical context can also provide a unique opportunity to study fundamental questions about neural coding and cognition. The rest of this review is dedicated to iEEG data acquired in the clinical context of epilepsy.

Stereotactic‐EEG and Electrocorticography: Two Intracranial Recording Techniques

Although microelectrodes have been used in humans to acquire single‐neuron spiking data [Fried et al., 1997; Heit et al., 1988; Ojemann et al., 2002; Ward and Thomas, 1955], clinical recordings in epilepsy patients are generally performed using macroelectrodes that measure coherent activity of local neuronal populations in the vicinity of the recording site. The most common choice in the clinical routine is to use either stereotactic‐electroencephalography (SEEG) or electrocorticography (ECoG) which acquire intracranial data using multilead depth electrodes or subdural grid electrodes respectively (see Fig. 1). Subdural grids consist of 2D arrays (or sometimes one‐dimensional strips) of electrodes positioned directly on the lateral surface of the brain, with a typical inter‐electrode distance of 1 cm [Engel et al., 2005]. In contrast, depth electrodes are semi‐flexible one‐dimensional linear arrays, shaped as narrow needles that penetrate deep into the brain. Such depth electrode implantations are often referred to as Stereotactic EEG because a stereotactic technique developed by Talairach and Bancaud is used to localize the electrodes [Kahane et al., 2006]. While subdural grids provide widespread cortical coverage and cortical maps of gyral activity, the multilead depth electrodes record from both sulci and gyri and go beneath the cortical surface to probe deep cortical structures, such as the cingulate gyrus, and occasionally subcortical structures, such as the lateral geniculate nucleus [Krolak‐Salmon et al., 2003] or the nucleus acumbens [Münte et al., 2008]. A further difference between SEEG and ECoG is that while depth electrodes require small burr holes for implantation the implantation of two‐dimensional subdural grids involves a larger craniotomy.

Converging evidence suggests that both grid [Menon et al., 1996] and depth‐electrodes recordings [Lachaux et al., 2003] provide sufficient spatial resolution to localize neural activity at the gyral level, a precision that is as good as, if not better, than what is achieved with fMRI. In addition, the spatial precision of the analysis also depends on the accuracy of electrode localization [Dalal et al., 2008a] and choice of reference electrode. In SEEG data, the local precision is highest when each recording site is referenced to its nearest neighbor (bipolar montage) than when one remote site is used as reference for all channels (common reference) [Lachaux et al., 2003].

Unless otherwise stated in the manuscript, the experiments performed by our group that are reported in this review were carried out using multi‐channel video‐EEG acquisition and monitoring system (Micromed, Treviso, Italy) that simultaneously records the intracerebral activity from up to 128 depth‐EEG electrodes (up to 1024 Hz sampling rate and a 0.1–200 Hz bandpass filter). As a general rule, twelve to fourteen semi‐rigid multi‐lead electrodes are stereotactically implanted in each patient. The stereotactic‐EEG (SEEG) electrodes used have a diameter of 0.8 mm and, depending on the target structure, consist of 5–18 contact leads 2 mm wide and 1.5 mm apart (Alcys, Besançon, France).

Strengths and Limitations of iEEG Research

At first sight, iEEG can be seen as combining the best of two worlds: the spatial resolution of fMRI and the temporal accuracy of MEG/EEG. In addition, given that the recording sites are within the brain, iEEG signals are immune to a large range of external artifacts (e.g., eye blinks, face and neck muscles contractions). However, it is important to bear in mind two significant limitations: first, the data acquired report on patients with epileptic disorders and, second, the invasive electrodes only provide a limited coverage of the brain.

Indeed, in recent years, iEEG research has begun to establish a set of methods and specific practices to limit the risk of making physiological interpretations based on phenomena that actually reflect the pathological condition of the patients. It is common iEEG practice to discard signal epochs that show any type of epileptiform activity. Furthermore, most iEEG studies include post‐surgical information about the spatial organization of the epileptogenic network and do not report on task‐related responses within that network. Also, when possible, an additional caution is to only report observations that are reproducible across several patients with different types of epileptogenic networks and anticonvulsant medication [Fell et al., 2002; Halgren et al., 1994]. Nevertheless, an important body of iEEG research provides evidence that such issues can indeed be overcome to an extent that warrants the physiological significance of the physiological insights and novel hypotheses drawn from the data [Axmacher et al., 2008b; Brovelli et al., 2005; Canolty et al., 2006; Crone et al., 1998a, 2006; Fell et al., 2002; Jung et al., 2008; Lachaux et al., 2003, 2005, 2006a; Lakatos et al., 2008; Mainy et al., 2007].

The second limitation of many iEEG investigations is the fact that the electrode implantation scheme only provides a restricted sampling of selected cerebral structures. A complete 3D coverage of the brain with a spatial resolution of 3.5 mm has been estimated to require about 10,000 recording sites [Halgren et al., 1998], that is roughly 100 times more than the number of sites we actually record (approx. 100). A more comprehensive view of the large‐scale networks involved in various cognitive tasks would therefore require combining data from multiple subjects with both overlapping and complementary electrode positions. In practice, collecting data from comparable anatomical origin across several patients is of non‐negligible difficulty since the target structures for intracerebral electrodes placement is determined in each patient individually and independently of research objectives. The standards of invasive and non‐invasive research are thus inherently different. Furthermore, to a certain extent, iEEG studies share some common aspects with monkey electrophysiology studies. The latter report in most cases on two or three animals with recordings in comparable regions and with a particularly high signal‐to‐noise ratio. In addition, the limited extent of probed cerebral structures (spatial coverage) and the question of generalization to the healthy human brain are undoubtedly two issues that are also relevant to findings reported invasively in monkey studies. Animal studies provide valuable observations and hypotheses that ultimately need to be explored and validated in humans. Similarly, the results of iEEG studies in human patients may be further validated in healthy subject by non‐invasive techniques such as MEG and EEG.

Gamma‐Band Activity Modulations: A Robust Correlate of Local Neural Activation

Several early animal microelectrode studies pointed towards the putative functional importance of synchronized neural oscillations at high frequencies (>40 Hz), i.e., within the gamma band [Eckhorn et al., 1988; Freeman, 1978; Gray et al., 1989; Rougeul‐Buser et al., 1975]. Neural synchronization in the gamma band has been proposed as a fundamental mechanism for neural communication underlying both perceptual and motor processes [Murthy and Fetz, 1992; Singer and Gray, 1995]. In addition, several scalp EEG studies reported task‐related responses in the gamma band in a variety of visual paradigms [Gruber et al., 1999; Tallon‐Baudry et al., 1996] suggesting a link to the neural synchronization phenomenon recorded invasively in animals.

Forty years after Chatrian's seminal report [Chatrian et al., 1960], task‐related gamma band increases were observed in direct recordings of the human brain using tightly controlled experimental paradigms and advanced signal processing analysis [Crone et al., 1998a; Lachaux et al., 2000]. Ever since, an increasing number of iEEG studies in humans (epilepsy patients for the large part) report gamma‐range activations in various brain structures and in various experimental conditions. Interestingly, although research into the functional significance of gamma activity is still in its early days, converging findings from these studies point to a number of properties of gamma‐band responses (GBRs) that appear to be reproducible across a variety of brain structures and in various experimental conditions. Previous work including numerous ECoG and SEEG studies by our group and others suggest that these common characteristics could be tentatively summarized as follows:

• 1

  Broad frequency extent: gamma‐band responses (GBR) consist of spectral power increases (or suppressions) that are time‐locked to sensory or motor events, in a broad range of frequencies (40– 200 Hz). Note also that task‐related gamma‐band suppressions (GBS) have recently been reported with iEEG [Lachaux et al., 2008].

• 2

  Functional specificity: the fine spatial organization of GBRs strongly depends on the cognitive process at hand. For example, Figure 2 illustrates this specificity within Broca's area during a word recognition task (semantic versus phonological processing). Similarly, we have also reported using SEEG in an implanted subject performing an auditory experiment, that switching the auditory stimulus from speech to music causes a seven millimeters shift in the focus of the GBR locus in the Superior temporal gyrus [Lachaux et al., 2007a].

• 3

  Spatial specificity: Although sources of GBRs have been found across distributed brain areas, each GBR source per se generally shows spatially focal activation. This contrasts with the wider spatial distribution of power modulations in lower frequency bands such as the alpha (8–13 Hz) or beta (15–30 Hz) bands. This general observation has been shown to be particularly clear in the sensorimotor cortex [Crone et al., 1998b], but is also visible during visual search tasks for instance (cf. Fig. 3).

• 4

  Generalization: large‐scale networks exhibiting GBRs have been found in a large variety of cognitive tasks involving working memory [Axmacher et al., 2008b; Howard et al., 2003; Mainy et al., 2008], declarative memory [Axmacher et al., 2008a; Fell et al., 2001, 2002, 2008; Sederberg et al., 2007], reading [Mainy et al., 2008], speech perception and production [Crone et al., 2001b], attention [Brovelli et al., 2005; Jensen et al., 2007; Jung et al., 2008; Ray et al., 2008], visual [Lachaux et al., 2005; Tanji et al., 2005], auditory [Bidet‐Caulet and Bertrand, 2005; Crone et al., 2001a; Edwards et al., 2005], somatosensory [Aoki et al., 1999] and olfactory perception [Jung et al., 2006], motor processes [Crone et al., 1998a; Lachaux et al., 2006].

Although the above properties should by no means convey the impression that GBRs provide by themselves a comprehensive view of task‐related neural dynamics, they do highlight, in our opinion, the importance of exploring the functional significance of gamma range activity. Whether GBRs represent a phenomenon that is intrinsic to cognition is still debated. A theoretical framework which draws upon single‐unit recordings in animal studies suggests that gamma‐band synchronization facilitates neural communication [Fries, 2005; Singer, 1999]. Compared to a resting baseline level, gamma‐band activity recorded within a neural population would increase when its neurons are recruited by a cognitive process, and conversely, gamma power might be expected to decrease when the population is no longer necessary or when its activity is inhibited [Lachaux et al., 2008]. The view that engagement in a cognitive act triggers a large‐scale network of distributed and spatially focal enhancements and suppressions (i.e., positive and negative modulations) of gamma‐band responses compared to the resting state provides an attractive framework for comparison with fMRI studies. Indeed, recent studies both in animals and in humans report prominent spatial overlap between the networks of gamma‐band activations and those of the BOLD signal suggesting a direct relationship between the two [Kayser et al., 2004; Lachaux et al., 2007b; Logothetis et al., 2001; Niessing et al., 2005; Nishida et al., 2008].

Oscillatory Modulations in Lower Frequency Bands

In addition to GBRs, time‐frequency analysis of iEEG data has also revealed robust task‐related response patterns in lower frequency bands, mostly in the theta (4–7 Hz), alpha (7–14 Hz) and beta (15–30 Hz) bands. Results obtained with iEEG are largely in line with reports from scalp recordings, and often reveal alpha and beta power suppressions during the task compared to the baseline period. For instance, such findings were reported in the sensorimotor cortex and described as event‐related mu and beta desynchronization [Crone et al., 1998b; Pfurtscheller et al., 2003] as well as in the parieto‐occipital cortex [e.g., Adrian, 1941]. Furthermore, alpha and beta suppressions often occur simultaneously with GBRs. Although there has been, to our knowledge, no extensive iEEG data meta‐analysis that specifically explores this relationship, it has often been reported in humans with iEEG studies [Crone et al., 1998a, b; Lachaux et al., 2005] and MEG recordings [de Lange et al., 2008] as well as in animal studies [e.g., Rickert et al., 2005]. Additionally, alpha and beta desynchronizations have been found to occur across cortical areas that often include, but also extend beyond the origin sites of GBR. Figure 3 shows an illustration of this pattern in a visual search task. This has also been reported in the motor cortex, where the spatial organization of movement‐related GBR provides a precise focal spatial mapping, while alpha and beta suppressions yield spatially smoothed maps [Crone et al., 1998a]. As shown in Figure 2, such a reversal between high and low frequency activation patterns is also visible in a word recognition paradigm contrasting semantic and phonological processing [Mainy et al., 2008].

Modulations of theta (4–7 Hz) activity have also been consistently reported in iEEG [Kahana et al., 2001; Raghavachari et al., 2001]. In contrast to alpha and beta activity, and much like the gamma band, theta activity appears to be a rhythm often observed to increase in amplitude during active involvement in a task. For instance, theta power increases have been reported in the hippocampus during spatial navigation [Ekstrom et al., 2005] and with memory load during the delay of a verbal working memory task in multiple frontal, temporal and parietal areas [Raghavachari et al., 2001]. The observation that both gamma and theta power increases during cognitive processing may suggest that the two rhythms interact with each other. Indeed, theta–gamma interactions have been reported in animals and in humans [Canolty et al., 2006; Jensen and Colgin, 2007].

Intracranial Event‐Related Potentials

Event‐related potentials measured at the intracerebral level have been explored extensively by numerous fundamental iEEG studies. A detailed review of this literature is beyond the scope of this review. Interestingly, however, ERPs and cerebral oscillations are often investigated separately. Only a few studies have sought to establish links between the two [Karakas et al., 2000; Mazaheri and Jensen, 2008; Nikulin et al., 2007]. A handful of studies have compared the task‐sensitivity and spatial organization of ERPs and gamma‐band responses and revealed a degree of overlap between the two systems. For instance, in a paradigm that required participants to detect faces hidden in high‐contrast figures, the presented stimuli triggered both ERPs and gamma‐band responses in the same region of the fusiform area [Lachaux et al., 2005]. However, this particular study did also reveal that, on some instances, the ERP was identical whether the face was actually detected or not, while GBR was stronger for detected faces. This is in line with other iEEG results that report functional dissociations between ERPs and gamma‐band responses [Tallon‐Baudry et al., 2005].

Statistical Considerations for iEEG Data Analysis and Interpretations

Intracranial EEG data analysis is complicated by the fact that populations are often small, and the recording sites are highly variable across patients. As a consequence, group statistics, as they are commonly used in MEG/EEG studies, are not suitable to many iEEG studies. Instead, task‐related effects can be shown at the individual level and localized precisely on the cortical anatomy of each patient. The attempt to represent statistically significant task‐related modulations of iEEG data across subjects, time, space and frequency is in itself quite a unique neuroimaging challenge. One approach used in our group to deal with this complexity consists of the following three‐step procedure: First, a statistical comparison is performed between experimental conditions at the individual level, on each recording site, to identify regions of interest with task‐related neural responses. This includes a correction for multiple comparisons and may involve either a direct comparison between two conditions or a comparison between the task and a pre‐stimulus resting baseline. The second step is the identification of brain regions of interest which show similar effects across several patients. The third step is to display the individual data for each region of interest, displaying the task‐related effect on each patient's anatomical MRI. The above steps are illustrated in Supporting Information Figure 1. Such a procedure has proven to be particularly well‐suited for analyzing depth‐electrode recordings in populations of ten to twenty patients [Jung et al., 2008; Mainy et al., 2007].