Combining EEG and fMRI: A multimodal tool for epilepsy research



Patients with epilepsy often present in their electroencephalogram (EEG) short electrical potentials (spikes or spike-wave bursts) that are not accompanied by clinical manifestations but are of important diagnostic significance. They result from a population of abnormally hyperactive and hypersynchronous neurons. It is not easy to determine the location of the cerebral generators and the other brain regions that may be involved as a result of this abnormal activity. The possibility to combine EEG recording with functional MRI (fMRI) scanning opens the opportunity to uncover the regions of the brain showing changes in the fMRI signal in response to epileptic spikes seen in the EEG. These regions are presumably involved in the abnormal neuronal activity at the origin of epileptic discharges. This paper reviews the methodology involved in performing such studies, particularly the challenge of recording a good quality EEG inside the MR scanner while scanning is taking place, and the methods required for the statistical analysis of the combined EEG and fMRI time series. We review the results obtained in patients with different types of epileptic disorders and discuss the difficult theoretical problems raised by the interpretation of an increase (activation) and decrease (deactivation) in blood oxygen level dependent (BOLD) signal, both frequently seen in response to spikes. J. Magn. Reson. Imaging 2006. © 2006 Wiley-Liss, Inc.

THE ELECTROENCEPHALOGRAM (EEG) was used in the context of epileptic disorders soon after it was discovered. This test remains today the gold-standard for the diagnosis of epilepsy, the classification of seizure types and epileptic disorders, and the localization of the generators of epileptic activity. The EEG is usually recorded on the scalp with 20 to 40 electrodes and interpreted visually by assessing which electrode locations, and hence brain regions, show some pattern of interest. This diagnostic method is very specific but lacks spatial resolution. It would be ideal if one could determine the intracerebral sources of the EEG patterns seen on the scalp, but the solutions to this problem (the so-called inverse problem) are very complex. Many methods have been developed (1). They are usually divided into the dipole-based methods, which assume that sources are localized, and the distributed sources methods, which assume extensive sources with specific characteristics (for instance smoothness). Both types of method require assumptions that are sometimes difficult to justify, and to confirm, since it is impossible to know the complete distribution of the intracerebral potential. Dipole-based methods have been applied to the localization of interictal epileptic activity (2–4), but also to analyze the onset of seizure discharges (5, 6). Distributed methods have also been applied in epilepsy (7). The validation of most studies is indirect, for instance the spike generator is found to be in the same location as a lesion seen on MRI, but it has been possible in some cases to confirm partially some results by comparison to intracerebral EEG studies of the same events (4, 6, 8).

Although we will not discuss magnetoencephalography (MEG) further, it should be noted that it is not fundamentally different from EEG since it records the same electromagnetic field generated by cerebral structures and uses the same methods and assumptions as EEG for its analysis. Several reviews comparing EEG and MEG have been published (9–11).

Functional magnetic resonance imaging (fMRI) is a powerful and noninvasive method that allows the localization of brain regions in which there is a change in the level of neuronal activity during an experimental condition compared to a control condition. The change in the level of neuronal activity is accompanied by a change in the ratio of concentration of oxy- and deoxyhemoglobin in the blood. This change can be measured through the blood oxygen level dependent (BOLD) effect (12, 13). fMRI is mostly used in the study of sensory, motor and cognitive functions, in which the experimental condition differs from the control condition in a way that is controlled by the experimenter. In the context of epilepsy or physiological changes in brain state, one can consider the control condition to occur at a time when the EEG is at baseline and the experimental condition to correspond to the presence of an epileptic discharge or of an electrophysiological phenomenon such as a sleep spindle.

To define such an experimental condition, it is necessary to record the EEG while the subject is in the MR scanner. Given the intensity of the magnetic field inside a scanner (1–7 T), it would seem impossible to record a signal such as the EEG, which is of very low amplitude and therefore very sensitive to external electromagnetic interference. The pioneering study of Ives et al (14) demonstrated that it is possible to record the EEG in such a hostile environment. It is now possible to combine EEG and fMRI in the study of epileptic disorders, and therefore to determine the region of the brain in which there is a change in the BOLD signal as a result of an epileptic discharge seen on scalp EEG, wherever that change takes place in the brain. One can hypothesize that this region is where the spike originates, in a similar way that single photon emission computed tomography (SPECT) studies are performed at the time of epileptic seizures to determine the regions of increased blood flow (15, 16).

We will present the methods that have been developed for recording and analyzing the EEG in the MR scanner, the methods required to analyze the BOLD signal resulting from epileptic discharges, and results from studies of different types of epileptic disorders. Finally we will discuss the difficult issue of the interpretation of EEG-fMRI results in epilepsy.


EEG Recording in the Scanner

Patient Safety and EEG Quality

The recording of EEG during fMRI scanning raises several questions concerning the safety of the patient and the quality of the resulting EEG and MRI signals. EEG electrodes are metallic and therefore it is possible that the rapidly changing magnetic fields associated with scanning will induce a current which could lead to heating and localized burns to the patient's scalp. Although this is a potential issue, it has been shown that using nonferrous electrodes and leads, perhaps with current-limiting resistors (17), and avoiding current loops involving the patient (18), result in safe recordings, certainly at 1.5 T. Simulations suggest elevated specific absorption rates when high density EEG is recorded at very high field strengths (up to 124 electrodes at 3 T and 7 T) (19), and this is clearly an issue that will have to be periodically reinvestigated as MRI technology develops. Using a head transmit coil would probably be safer than a whole body coil, since the power deposition would be reduced, and this is more important when doing EEG-fMRI since the presence of the electrodes increases the specific absorption rate.

The quality of the EEG within the scanner is reduced compared to the EEG outside, because of the presence of conducting electrodes and wires within the static magnetic field itself. Several factors are involved in this reduction, such as small movements of the electrode wires caused by subtle head movements or vibrations of the scanner. We found however that with a careful EEG setup, the increase of root mean power (RMP) was only 7% between EEGs outside and inside the scanner, with no scanning (20). The main factor in reaching this quality was the immobilization of the wires between the head and the amplifier with sand bags (which, if not immobilized, produced an increase of 29% of the RMP), followed by immobilization of the head with a vacuum cushion (10% increase after removal) and immobilization of the wires on the head by a bandage (6% increase).

During scanning the situation is different. While the currents induced in the electrodes and leads by the rapidly changing magnetic fields may be sufficiently small to be safe for the patient, they result in artifacts on the EEG which can be of the order of 50 times the background EEG (“gradient artifacts”). This can be avoided by only scanning after the observation of an EEG event of interest (EEG-triggered fMRI, see below), but this is not a particularly efficient experimental paradigm and requires an expert observer to be monitoring the EEG at all times. On the other hand, continuous scanning requires an amplifier with a sufficiently large dynamic range that it will not saturate. Several commercial amplifiers are now available, that can be placed in the scanner room and allow short cables from the electrodes to the amplifier (minimizing noise pickup), and are connected to a recording computer outside of the scanner room via an optic fiber cable. The optic cable ensures the absence of an electrically conductive bridge between the outside and the inside of the scanner room, which would break the magnetic shielding of the scanner room and deteriorate the quality of MR images. In general, good quality MR images can be achieved despite the EEG apparatus (21).

Gradient and Pulse Artifact Removal

The most widely used method to remove the gradient artifact was originally proposed by Allen et al (22) and consists of estimating the artifact and subtracting it from each frame, followed by adaptive noise cancellation. For this approach to be valid, the assumption is made that the recorded EEG consists of artifact plus physiological EEG, which is reasonable assuming that the amplifier remains within the linear range (i.e., does not saturate). Moreover, the artifact must have been faithfully recorded and be stable over the scanning period, which in addition to having an amplifier that does not saturate requires a sampling rate of several kilohertz. An example of the use of this method is shown on Fig. 1. Another method takes advantage of the frequency structure of the gradient artifact, i.e., a ray spectrum. A Fourier filter is constructed by zeroing the signal at frequencies with high power with respect to baseline EEG (23).

Figure 1.

Eight channels bipolar EEG samples to illustrate quality of recordings. a: Routine EEG with normal electrodes in the EEG laboratory, showing a right temporal spike, with equipotentiality at F8–T4. b: EEG acquired inside the scanner with Ag/AgCl electrodes, during the acquisition of EPI sequences, during which EEG activity is not visible. c: Same segment as in b, after removal of the MRI artifact and filter, disclosing a spike similar to that shown in a.

Several variations on the subtraction scheme have been proposed. In Goldman et al (24), it was suggested to acquire the EEG time-locked to the fMRI acquisition, which permitted a lower sampling rate to be used and still have a stable artifact between frames. Anami et al (25) have, in addition, customized their gradient sequence in order to have periods with low artifact amplitude, and to sample the signal at these periods. This considerably reduces the artifact and renders the task of filtering much easier. Bénar et al (20) proposed grouping the artifact frames by value of jitter with respect to the fMRI acquisition, which permitted a low sampling rate and did not require the use of a trigger from the scanner. Negishi et al (26) suggested the use of principal component analysis (PCA) in order to capture variations of the artifact over time. Garreffa et al (27) have shown that it is possible to filter the artifact in real time, thereby allowing monitoring of the EEG while the subject is being scanned.

Recently, a method has been demonstrated using two subdural electrodes in an animal model of epilepsy that does not require post-processing of the EEG and allows real-time EEG recording (28). This requires closely spaced active electrodes and an amplifier designed to maximize the common mode rejection ratio. Whether this method will prove suitable for multiple electrode recordings of human scalp EEG remains to be seen.

Another artifact that is often observed on the EEG recorded within the scanner is referred to as the pulse or ballistocardiogram artifact. This consists of deviations following each heartbeat and possibly originates from small movements of the head or the electrodes following each pulsation because of fast movement of the blood in the arteries. It has been noted from the first report as one of the main problems when recording EEG in an MR scanner (14). It can be removed by averaging and subtraction (29), adaptive filtering (30), wavelet filtering (31), or independent component analysis (ICA) (20, 32). An example is shown in Fig. 2.

Figure 2.

Results of ballistocardiogram removal. a: Original recording. b: After ICA filtering. The epileptic spikes were left intact and ICA filtering essentially eliminated ballistocardiographic activity (20).

Acquisition and Analysis of MR Images

fMRI Scanning Protocol: EEG-Triggered or Continuous

As mentioned above, one method of avoiding the problems associated with the gradient artifact is to only scan after the observation of an EEG event, triggering fMRI acquisition three or four seconds after each spike, when the hemodynamic response is presumed to be close to its maximum. An equivalent number of baseline frames also has to be acquired to allow statistical comparison of the postevent and baseline periods. This technique was widely used in the first reports of combined EEG-fMRI (33–35), but has several drawbacks: Only as many frames as spikes can be acquired in total; low frequency drifts cannot be taken into account; and an experienced observer needs to be actively monitoring the EEG during the whole session. With this method, it has also been proposed to inject a short-acting anticonvulsant (for instance, clonazepam or lorazepam) in order to obtain a long baseline period completely free of spikes (36), which can then be compared to a period with spikes. This can be particularly interesting in the case of patients with a high spiking rate, for whom it is difficult to obtain spike-free baseline sections. Drugs such as benzodiazepines may however alter the cellular metabolism and complicate the interpretation of the results.

The most commonly used method now is to scan continuously and filter the gradient artifact afterwards as described above, thus revealing the timing of epileptic spikes. An electroencephalographer identifies epileptic events after offline artifact removal. This has a higher sensitivity to detect spike-related signal changes than triggered acquisition (37). Continuous recording also has the advantage that the hemodynamic response function (HRF) can be calculated, which allows more detailed investigation of the link between neuronal activity and vascular response (neurovascular coupling). For example, since the HRF associated with a spike lasts 10 to 15 seconds (38, 39), the issue of linearity becomes relevant if spikes are separated by less than this duration (40). Also, very little work has been done to determine whether the hemodynamic response to epileptic activity is equivalent to that to normal brain function, although there are suggestions that neurovascular coupling is preserved in patients with epilepsy (41, 42).

Patients' Movements and Pre-Processing of fMRI Data

Patient movement can have a severe effect on the quality of fMRI data, particularly over the course of a long scanning session such as those required for obtaining a reasonable number of spikes. The patients have to remain as still as possible throughout the scanning session, which is much more likely if they are cooperative and comfortable. Significant discomfort is caused by the pressure points at the electrodes. This can be in large part alleviated by the use of a plastic bag (50 × 70 cm, 10 liter capacity) filled with very small polystyrene spheres, in which a vacuum is obtained by air suction (S&S X-Ray Products, Brooklyn, NY, USA).

However, even the most cooperative patient will move slightly over the course of two hours, and some image realignment is usually necessary and provided by most statistical processing packages. This can only go some way towards correcting the problem and scans with more than a few millimeters of movement are unlikely to be usable. Realigning the images is not sufficient for complete recovery of the fMRI signal though, as pointed out by Friston et al (43) and Salek-Haddadi et al (44), who used the parameters of motion correction as additional regressors in the statistical analysis. Following motion correction, images are smoothed, which helps to reduce noise. A Gaussian filter of 6–8 mm full width at half maximum (FWHM) is generally used.

Statistical Analysis and Interpretation

In event-related fMRI (45), the statistical methods most widely used are probably those developed by Friston et al (46), Friston and Worsley (47), and Worsley et al (48). In this approach the signal at each voxel is statistically compared with a model constructed by convolving impulses corresponding to the timing of the spikes (the events) with one or more basis functions, representing the hemodynamic response. A statistical map (t or F statistic) is constructed where the value at each voxel reflects the resemblance between the model and the data, and therefore the plausibility that this region is activated in response to the spikes. A wide variety of basis functions have been used to describe the hemodynamic response. The simplest uses a standard HRF, which is the measured response to a brief stimulus, such as an auditory tone (49). The model assumes that following each spike the BOLD signal will change according to the shape of the HRF. Clearly, this does not take into account differences between patients, between different brain regions or between sessions within a single patient, although it is known that these effects can be considerable (50). There is also no real reason to assume that the response to an epileptic discharge, or to any other pathological activity, will be the same as that observed in normal subjects performing sensory or cognitive tasks. Standard HRFs have been widely used and are capable of detecting responses even when the actual HRF differs considerably from the model (39, 51, 52) (Fig. 3). This approach, however, is likely to fail if the actual HRF differs too much from the standard HRF, or if the signal to noise ratio in the activated area is too low.

Figure 3.

HRFs measured at voxels corresponding to local peaks in the statistical map. Our population consisted of 13 patients, 12 are shown. The first row shows HRFs with high amplitude (scale up to 5%); the second row HRFs with amplitude up to 2% and the last two rows HRFs of 1% or less. The statistical maps were obtained with a standard HRF peaking at 5.4 seconds. The collected HRFs were obtained by projecting the fMRI signals on a basis of sine and cosine functions, defined from the spike time to 64 seconds after the event (only 45 seconds are shown). There is large interpatient variability. The intrapatient variability in shape and amplitude does not seem to be related to the distance between points. For patient 2, HRFs (a) and (b) have different amplitude and timing but correspond to points at a distance of 19 mm (left basal temporal and lateral temporal). For patient 7, HRFs (a) and (c) are very similar but correspond to a distance of 47 mm (temporooccipital and parietal). This is also the case for the negative HRFs for patient 11, which are distant of 100 mm (symmetric posterior parietal deactivations) (39).

There are several approaches that allow for some variability in the shape and timing of the modeled HRF. One option is to add the derivative of the HRF to capture small departures from a standard model (53). The efficiency of this approach depends on the adequacy of the HRF model itself (54, 55). Another approach is to analyze the data with a series of standard HRFs with different latencies (52) or, when even more flexibility is required, a set of basis functions can be used, such as sines and cosines (“Fourier set”) (45) or consecutive impulse functions (“finite impulse response”, or FIR) (56). This has the advantage of allowing much larger variations in the shape of the actual response. The corresponding maps are based on an F statistic, which does not distinguish between positive and negative BOLD responses. An exploratory approach has also been proposed where the data are first processed with a standard HRF. Then, the actual HRFs are obtained at the peaks in the statistical map, modeled as a sum of two gamma functions, and the data are reanalyzed with these new models (51). Another exploratory approach lies in the use of data-driven approaches such as ICA (57)), which do not require the specification of a model of response. It has been proposed recently to combine the data-driven approaches with paradigm-based information (58–60).

Low frequency drifts in the fMRI data, which possibly originate from scanner instability (61), must also be modeled. This can be accomplished by using a third order polynomial, although this may not always be flexible enough (39), or by using a basis set of sine and cosine functions (62), which is similar to Fourier filtering.

Whatever the specifics of the method of analysis, it usually results in a statistical parametric map. This shows the probability of the data from each voxel being correlated with the model, usually in the form of the t or F statistics. In order to interpret the map, it is usual to apply a statistical threshold so that only voxels above the threshold remain. This reveals the regions that are presumably involved in the generation of the spikes. The threshold can either be calculated based on a Bonferroni correction, which takes into account the multiple comparisons inherent in the massively univariate approach of testing on a voxel wise basis, or by examining the magnitude of random excursions in the t or F fields (63). Alternatively, the spatial extent of the responses can be taken into account (64), or a combination of their magnitude and extent (65). In the final interpretation of the map, it is worth remembering that BOLD fMRI is prone to recording signal changes from large draining veins (66), which may be at some distance from the actual site of neuronal activation (67).

In addition, there are regional differences in the sensitivity of fMRI to detect task- or spike-related signal changes (68, 69), particularly as a result of inhomogeneities in the static magnetic field caused by susceptibility effects at tissue boundaries. This can be a particular problem in the inferior aspect of the temporal lobe and in the orbitofrontal region (70). In patients where the potential epileptogenic region is located in these areas, the signal loss may be an important issue. To assess the effect of optimization of the sequence acquisition to improve the raw signal in these areas, we scanned eight patients with temporal lobe spikes using z-shimming. Comparing the z-shimmed images with those created using the nominal gradient, the mean signal increase in the temporal lobes was 46% and the percentage of temporal lobe voxels above a brain intensity threshold rose from 66% to 78%. However, this increase did not lead to any significant differences in the statistical maps created with the two sets of functional images, suggesting that an even larger increase may be required (71).


In this section, we will review the EEG-fMRI studies of patients with focal or generalized epilepsy, first in an overview of the series and then discussing important findings in the most relevant publications. It has to be emphasized that different studies used different equipment (magnet strength, EEG amplifier sampling rate) or protocol (spike-triggered or continuous recording, statistical analysis), and this may explain some of the differences between studies. In addition, early studies did not report on the presence of negative changes in BOLD signal (deactivations), and their importance has been highlighted in more recent EEG-fMRI publications (72–74).

Almost all studies deal with interictal epileptic activity rather than seizures. Seizures are unpredictable and occur rarely while a patient is in the scanner. fMRI during epileptic seizures have nevertheless been acquired without concomitant EEG recording in a few cases (75–79), usually showing activation in the presumed seizure focus. There are also some case reports of focal seizures recorded during EEG-fMRI allowing a better analysis of the relationships between the different phases of spread of the seizure and the regions showing BOLD activation (80, 81).

In the first EEG-fMRI recordings, Warach et al (33) studied interictal activity in two patients, one of them with generalized epilepsy that showed, surprisingly, a focal response. Most of the subsequent studies looked at spikes in series of patients with different focal epilepsy syndromes and MRI findings (35–37, 41, 51, 52, 82–89).

While studying these patients, researchers described the BOLD responses, and added considerations such as concordance with EEG spike topography (all studies), replicability of the responses within subjects (35), different statistical treatment (41, 51, 52, 83), electrode placement according to the prior knowledge from EEG recordings (86), and spatial relationship between lesions and responses (37, 51, 52, 82, 88, 89) Some interesting cases of focal epilepsy have also been reported where different modalities of investigation were compared in individual patients (80, 90–92), and rarer conditions such as reading epilepsy (93) and discharges related to fixation-off sensitivity (94) have also been studied.

The first series of generalized epilepsy patients (95) explored photosensitivity, but later on, emphasis was placed on spontaneous generalized discharges, showing involvement of the thalamus and the presence of widespread frontal, parietal and posterior cingulate region deactivation (72, 96). Posterior cingulate deactivation had already been observed by Hill et al (95), but they considered it an undershoot response to the stimulus.

Focal Epilepsy Studies

In focal epilepsy patients, many of whom have the potential to become surgical candidates, the use of EEG-fMRI to evaluate the epileptic focus becomes an issue of the concordance between the BOLD response with the EEG field and the presumed anatomical regions involved in the generation of the discharges. In lesional focal epilepsy, the spatial relationship between the response and the boundaries of an existing structural lesion is also important to consider.

One of the first series of EEG-fMRI in focal epilepsy patients (35) included patients with various structural MRI findings such as atrophy and malformations of cortical development (MCD), patients who had been operated, and patients with normal MRI (nonlesional focal epilepsy). The study was performed in a 1.5-T scanner using a spike-triggered protocol, and was the only one so far systematically to include replicability of BOLD responses in all 10 patients that were scanned (at least two sessions per patient, with a total of 24 EEG-fMRI studies): six of the 10 patients had reproducible responses in close spatial concordance with the epileptic discharges seen on EEG.

Another early publication by Patel et al (83) included a larger series of patients (20 patients) with focal epilepsy, the majority with temporal lobe epilepsy, with no mention of structural MRI findings. This was also a spike-triggered study in a 1.5-T scanner where agreement of fMR response with EEG localization was determined, and that explored the use of three types of analysis regarding computation of the events in the statistics. The method that gave the best results was to weight the images corresponding to individual spikes, taking into account the mean and standard deviation of the baseline images. They performed a visual analysis and considered the regions with a “bright signal”. The authors found BOLD responses in nine of 12 patients that could be fully evaluated, and these responses were consistent with the EEG discharges observed in the scalp. Lazeyras et al (36) studied 11 patients with focal epilepsy and a variety of structural abnormalities (MCD, atrophy or porencephaly) or non-lesional epilepsy, also with spike-triggered acquisition at 1.5 T. An important validation to the BOLD responses was added: in eight of 11 patients, the response was concordant with EEG, and in five this was further confirmed with depth electrodes recording data.

The first study to specifically address the spatial relationship between EEG-fMRI responses and lesions was performed by Krakow and colleagues (82). They studied 24 patients with focal epilepsy, 15 of whom had structural brain abnormalities (MCD, hippocampal atrophy or tumor). Twelve patients (50%) had activations that were concordant with the EEG, and in the seven patients who had structural lesions, responses were concordant with the lesion. Two patients showed discordant responses, and 10 had no response, with a significantly lower mean spike amplitudes compared to those with positive fMRI results.

Jäger et al. (84) studied 10 patients with focal epilepsy, in the first series using continuous EEG-fMRI recording after the original description of the method by Lemieux et al (38). However, only five patients could be analyzed, all showing BOLD responses concordant with the EEG. The only lesional case (left occipital lobe MCD), showed no involvement of the lesion. The authors stressed that a high spiking rate was not necessary to obtain fMRI responses, with an average of 17 events per study in their series.

Al-Asmi et al (37) reported the largest group of patients to date, and the series comprised 38 patients with various MRI findings (atrophy, postoperative, MCD and normal anatomical scans). Sixteen studies were performed with the original method of spike-triggered fMRI and 32 with the continuous scanning method described above. From the total 48 studies, 17 could not be analyzed because patients did not show any spikes during scanning or there were technical problems. Activation was obtained in 39% of 31 studies, concordant with EEG localization in almost all studies and confirmed with intracranial depth electrodes in four patients. In this study, two patients had two scanning sessions, and in one of them there was no response. Therefore, the total number of patients with no response was 17 of 31, giving 55%. The rate of response was independent of the presence of a lesion and in lesional cases, responses could be identified inside or near the structural anomaly. Interestingly, these investigators found that bursts of spikes were more likely to generate an fMRI response than isolated spikes (76% vs. 11%). Also, the number of spikes tended to be higher in patients who showed activation than in those who did not, but there was a great dispersion of values and the difference was not statistically significant. Although spiking rate is a factor influencing the presence of BOLD response, it is clearly not the only one since there were responses in patients with very few spikes (seven in one case) and a lack of response in others with numerous spikes.

In a subsequent study (52) this research group analyzed a series of focal epilepsy patients using HRFs with peaks ranging from three to nine seconds, as opposed to the use of the “standard” HRF or Glover response (49) that peaks at 5.4 seconds. The proportion of patients showing a BOLD response with this multiple HRF assessment increased from 45% to 62.5%. In the vast majority of cases, the BOLD response was in the same lobe as the peak of the epileptic spike, but in some cases included regions remote from the spiking region. The standard HRF was good at detecting activations (positive BOLD response), but less appropriate for deactivations (negative response), which were more accurately modeled by an HRF that peaked later than the standard. Coregistration of statistical maps with gadolinium-enhanced MRIs suggested that the detected fMRI responses were not in general related to large veins.

Kikuchi et al (86) tried to limit the number of electrodes (six) used in the EEG-fMRI session using a priori knowledge of the spatial distribution of the spikes in the EEG, and studied six patients with nonlesional focal epilepsy. Activations were found in concordance with the spikes in half of the patients, and the ones that did not show any responses tended to have a low number of spikes and low spike amplitude.

We had the opportunity to study a large series of temporal lobe epilepsy patients with a variety of underlying structural abnormalities (87), and assessed the involvement of the temporal and extra-temporal regions as the result of a temporal lobe spiking activity. We analyzed 35 EEG-fMRI studies derived from 27 patients who showed spikes during scanning. BOLD responses occurred in 83% of studies, predominantly in the spiking temporal lobe, and manifested as activation or deactivation. Responses often involved also the contralateral homologous cortex at the time of unilateral spikes and frequently involved extratemporal regions, suggesting a widespread effect of temporal lobe epileptic spikes (Fig. 4).

Figure 4.

a: Activation clearly predominates in the left temporal lobe in a patient with nonlesional partial epilepsy and generalized spikes; activation is also seen in other brain regions and particularly in the homologous right temporal region. b: Patient with temporal lobe epilepsy and left temporal spikes. The robust activation seen in the left temporal region is not focal, and there is also activation in other brain areas, including the homologous contralateral temporal lobe.

Two studies looked specifically at patients with benign epilepsy with centrotemporal spikes (BECT): a series of seven patients by Boor et al (85) and a case report by Archer et al (91). The first study found activation that was concordant with the spikes, in the perisylvian central region, in three patients. In the second, with left-sided spikes, activation was seen in the inferior left sensorimotor cortex near the face area, consistent with the facial sensorimotor involvement in BECT seizures. Deactivation however was described in the anterior cingulate region supporting the possibility that the scalp recorded field of this patient with BECTs may reflect electrical change in more than one brain region.

Special attention has been given to the study of malformations of cortical development, although the first descriptions were in the context of series of patients with other pathologies (35–37, 51, 82, 84). Two recent papers addressed the involvement of the malformation itself, the perilesional area and the distant brain regions, in the pattern of BOLD responses. Kobayashi et al. (89) looked at nine patients with polymicrogyria and Federico et al (88) studied five patients with focal cortical dysplasia and one with a ganglioglioma. All patients showed BOLD responses. Activation within the malformations, but not involving the whole lesion, was the main observation in these two series supporting intrinsic epileptogenicity of those malformations (Fig. 5). However, other areas in the vicinity of the lesions, and at a distance, frequently in homologous cortical areas and in subcortical regions, showed BOLD responses, often as deactivations (Fig. 6). This indicates again that spiking may activate an abnormal network, in the lesion and at a distance.

Figure 5.

Activation in a patient with bilateral perisylvian polymicrogyria (89): very widespread activation (b) is maximum in the lesion (better visualized in a, the anatomical scan) bilaterally.

Figure 6.

Bilateral occipital deactivation concordant with the spikes, observed at T5–O1 and T6–O2, in a patient with bilateral periventricular nodular heterotopia.

From these studies on focal epilepsy patients, where we have prior knowledge of potential anatomical regions involved in spike generation, some issues should be highlighted. The percentage of studies that shows a response is variable and depends, among other factors, on: magnet strength (the signal to noise ratio increases and the likelihood of a response is higher with higher magnet strength) and the number of spikes analyzed (although we have seen good responses with just a few events, a large number of spikes is more likely to show a response). Among the studies that did not show any response, a methodological issue relates to modeling the HRF: it is possible that a more appropriate model could be more adequate to disclose BOLD responses in these individuals. There are, however, other issues since some patients that were rescanned for not showing a response had one in the second scan.

Analyzing only the studies with responses, the majority shows some degree of concordance with the spike field and potential generators, but often in the context of a more widespread response. The difference between these two scenarios may be related to the choice of statistical threshold. The concordance with the EEG has been confirmed with intracerebral recordings in a few cases, but in general work remains to be done to obtain a better understanding of the meaning of many of the responses.

Generalized Epilepsy Studies

Hill et al (95) published the first EEG-fMRI study with generalized epilepsy patients, which included 16 epileptic subjects (nine had photosensitivity and were also reported separately in reference 97). The authors performed visual stimulation to elicit photoparoxysmal spike-wave activity and used a spike-triggered EEG-fMRI protocol to assess responses to these paroxysms. With a flash stimulation lasting two seconds, prominent visual cortex activation was seen in all normal subjects and patients. There were no fMRI responses in relation to the brief photoparoxysmal spike-wave activity evoked in photosensitive patients. Patients with photosensitivity, however, had increased areas of cortical activation with photic stimulation, simultaneous involvement of noncontiguous areas (most prominent in perirolandic regions), and, immediately after the photic stimulation, a decrease in the fMRI signal in the occipital cortex and in the posterior cingulate gyrus, which they interpreted as an undershoot from the earlier increase.

Another study, performed by Archer et al (96) in a 3-T scanner, included five patients with generalized epilepsy, four of whom showed both activation and deactivation. Activations involved different cortical areas and the thalami (in two studies), and deactivations were seen in the posterior cingulate regions. In a group analysis, they confirmed this consistent posterior cingulate deactivation and found scattered activation in the precentral sulci, bilaterally, but no response in the thalami. The single patient who showed no response in the individual analysis had only four epileptic discharges.

Aghakhani et al (72) studied a group of patients with idiopathic generalized epilepsy (IGE), selected on the basis of frequent bursts of generalized spike or polyspikes and wave activity. In this group of patients, 14 of 15 studies in which EEG bursts were present showed a response. Responses were found in the thalamus, most often in the form of activation, and were widespread, bilateral and symmetrical in the cerebral cortex, predominantly in the form of deactivation (although activations were also commonly seen). This study clearly demonstrated an involvement of the thalamus in generalized spike-wave bursts of IGE, providing further evidence for the thalamocortical circuit involvement in the generation of interictal and ictal spike-and-wave activity. Secondly, in contrast to the predominantly frontal EEG distribution of the spike-and-wave activity, the BOLD response was more diffuse and the posterior head regions were almost as involved as the frontal areas. A group analysis later performed in this series of IGE patients (71), showed bilateral activations in the thalamus, mesial frontal region, insulae, midline cerebellum and on the borders of the lateral ventricles. Deactivations were found bilaterally in the anterior frontal and parietal regions and in the posterior cingulate gyri, in a pattern very similar to that seen in the “default” state of the brain (98–100).

Salek-Haddadi et al (100) described ictal EEG-fMRI results in a patient with juvenile absence epilepsy, and who had four absence seizures with generalized spike-wave discharges during the 35-minute scanning period. They also found bilateral thalamic activation and widespread symmetrical cortical deactivation with however frontal predominance.

In summary, EEG-fMRI studies in generalized epileptic discharges show a widespread activation involving both anterior and posterior head regions, with important involvement of the thalami. Deactivation seems to follow the spatial distribution seen in the default mode of brain function, and might represent the effect of the bursts of discharges on normal brain function (Fig. 7). As compared to focal spikes, generalized discharges show responses more frequently, and responses are always widespread.

Figure 7.

Deactivation in a patient with bilateral spike and wave activity (72). The deactivation is in the same spatial distribution as that seen in the default state of the brain (98).


EEG (and MEG) is commonly used to localize epileptic discharges, with visual inspection or with inverse solutions based on dipole modeling or distributed sources (2–4, 7). We described how fMRI can be used to image the metabolic changes resulting from epileptic discharges. The integration of EEG and fMRI data is a promising avenue for obtaining a spatiotemporal picture of brain activity with high spatial (on the order of millimeters) and high temporal resolution (millisecond). With this in mind, attention has been directed towards designing a methodological framework that would allow such integration.

Any attempt to integrate these data should start with a word of caution (101, 102). The two modalities do not measure the same physical quantities, may not be sensitive to the same neural populations and are not subject to the same uncertainties. On the one hand, EEG measures voltages at the millisecond scale, and is mostly sensitive to the synchronous activity of neocortical pyramidal cells. Uncertainty in source localization originates from errors in the head models used to compute the potentials produced by given currents. It also derives from the ill-defined nature of the inverse problem where volume currents are reconstructed from surface measurements. On the other hand, fMRI measures the effect of deoxyhemoglobin concentration in veins and is linked to the metabolic needs of all cells drained by these veins. Uncertainty in fMRI originates from the fact that the signal can be at a distance from the activated neural population (67), and because the link between neural activity and fMRI signal is far from being completely understood (103).

The first step before integration is to compare the spatial findings of the EEG with the results of the fMRI. Many techniques of EEG source localization can be used, that differ by the source model (dipoles, distributed sources and cortical patches) or by the inverse problem technique (e.g., choice of regularization). One option is to consider EEG statistical maps that can be seen as a parallel to the classical fMRI parametric maps (8, 104). Using this technique in five patients with focal epileptic spikes, we found that data from EEG source localization and fMRI were partially concordant, but were all in correspondence with a region of epileptic activity in intracerebral recordings (Bénar, 73). This suggests that the two modalities are complementary in the spatial localization of epileptic spikes, and that the correspondence between fMRI areas and dipoles should not be considered as one-to-one.

Efforts have been directed to the use of fMRI data as independent spatial information that could be injected into the ill-defined inverse problem of EEG source localization. The simplest constraint is to place (“seed”) an EEG dipole in each fMRI activated area, representing focal magnetic or electric sources (105, 106). Ahlfors et al (107) recommended, as a precaution, performing EEG source localization independently from fMRI information, and then refining the EEG model based on the fMRI responses. Alternatively, the weights of each source in linear distributed inverse solutions can be biased in order to give more weight to regions with fMRI activity (108). Liu et al (109) verified that electromagnetic sources with no fMRI response could still be recovered by using fMRI as a partial constraint.

In fact, the linear distributed inverse solutions can be seen as a particular case of the Bayesian formulation of the inverse problem of EEG (110). In this formulation, fMRI results can be seen as a priori information (111). Moreover, the Bayesian framework is flexible, as it permits statistical inferences on the number and size of sources (112) or the incorporation of other a priori knowledge such as temporal smoothness (113). Daunizeau et al (114) used this formulation to assess the relevance of incorporating the fMRI results as a priori spatial information for each particular data set. Trujillo-Barreto et al (115) proposed to go further by using a symmetrical model that gives equal weight to EEG and fMRI data in the localization process.

The next step in integration of EEG and fMRI is to consider the temporal links between the two modalities. Simultaneous EEG-fMRI allows the simultaneous fluctuations of the two signals to be followed. This can be done by grouping events and observing the relative variations in amplitude of EEG and fMRI (116, 117). A promising advance is to use the information from continuous EEG as a marker for fMRI analysis. This can be done for single-event variations in EEG amplitude (118) or continuous fluctuations in the power of some EEG bands (99, 119, 120).


In most fMRI studies of sensory, motor and cognitive functions, the paradigm implies that the experimental condition activates a part of the brain more than the control condition. For instance, there can be a contrast between auditory stimulation and silence, motor activity and rest, hearing words in contrast to non-word sounds. It is expected that the target condition results in a higher level of neuronal activity, thus an increase in the BOLD signal (or activation). Increases in postsynaptic activity, whether excitatory or inhibitory, result in an increase in oxidative metabolism (121) and eventually in an increase in the BOLD signal. Action potentials, on their own, appear to play a lesser role in explaining changes in the BOLD signal. In this view, an increased BOLD signal in a given region results from increased synaptic activity, because action potentials arrive in that region and create post-synaptic activity or because of a change in local circuitry.

Sensory and cognitive tasks may, however, result in a decrease in the BOLD signal (or deactivation), when comparing the experimental condition to the control condition (122). Such deactivations are common in response to epileptic discharges. Four mechanisms can be envisaged to explain this phenomenon: 1) a relative reduction of cerebral blood flow (CBF) in the deactivated areas is caused by a steal phenomenon secondary to the increased CBF in activated regions. This can only explain a decrease that is adjacent to a region of increased BOLD signal. 2) There is an abnormal coupling between neuronal activity and regional CBF (rCBF). If increased neuronal activity is not accompanied by the usual increase in blood flow, for instance in pathological conditions involving the cerebral circulation, a decrease in the BOLD signal will be observed. A preserved neurovascular coupling has in fact been demonstrated in epileptic patients, during a motor task and at the time of interictal epileptic discharges (42). 3) Regions of deactivation correspond to decreased synaptic activity, such as that caused by reduced neuronal input or by functional deafferentation, (123) compared to the control condition. 4) GABAergic-mediated inhibition results in a profound decrease in neuronal firing and at a very low cost of energy (124, 125). On balance, this would result in a decrease in energy requirements (neuronal firing is not the major contributor to energy demands but is nevertheless a significant factor), and, hence, in a reduction of O2 consumption and rCBF.

We found that both focal and generalized epileptic discharges are frequently associated with deactivation (74). This is a surprising finding as it is usually expected that an epileptic discharge results from intense neuronal activity, which should cause an increase in BOLD signal. We believe that this could be explained either by a reduced synaptic activity or by a low energy requirement GABAergic inhibition (hypotheses 3 and 4 above).


These studies clearly illustrate the great potential of combining EEG and fMRI in the understanding of the pathophysiological mechanisms of epileptic discharges, as outlined in a recent editorial (126). The method of continuous scanning followed by the removal of the scanning artifact from the EEG is probably the most efficient for data acquisition. Inconsistencies are present in the results but it is important to keep in mind that one cannot expect a strict one-to-one correspondence between the EEG and fMRI observations. First, the EEG records only activity from superficial cortical layers. Second, we are measuring very different types of activity, one electrical and the other based on a venous response to metabolic changes (where BOLD results from complex interactions between blood flow, blood volume and O2 consumption). It is likely that in some cases one modality measures activity to which the other is blind. For this reason EEG and fMRI should be regarded as providing complementary information.

The majority of patients with focal and generalized epileptic activity now show BOLD responses in relation to the EEG events, although the optimal method of data analysis remains to be specified. A future challenge will be to adapt fMRI analysis techniques to the specific requirements of epileptic activity, in contrast to using techniques developed for functional activation. Some approaches do not rely on the linear assumption, such as that presented by Friston et al (127). Others, such as temporal clustering, attempt to analyze the BOLD signal independently of the occurrence of spikes in the EEG (128, 129). In the independent component analysis approach (57), fMRI data sets are decomposed into spatially independent components. One or several of these components could be related to epileptic activity. If these methods were successful, it would be possible to detect epileptic activity anywhere in the brain, whether or not this activity is visible on the scalp EEG.

The significance of the different BOLD responses also remains to be fully evaluated. In particular the presence of both activation and deactivation is puzzling in the context of epilepsy, where one expects that excessive neuronal firing is the predominant phenomenon and that it would lead exclusively to activation. In addition, it will be important to evaluate the specific responses of the different types of epileptogenic lesions, such as mesial temporal sclerosis, malformations of cortical development, atrophic lesions and brain tumors. The extent of the response may give an indication of the extent of the epileptogenic zone and of the regions that may be affected by an epileptic discharge, beyond the limits of the anatomical lesion.

In generalized spike-wave discharges, the combination of activation and deactivation may be explained in a different way and may give a clue regarding the interpretation of deactivations in some situations. Activation was found in the thalamus, insula, and mesial midfrontal region, as well as in the cerebellum, bilaterally; deactivation was observed in anterior frontal and parietal regions, also bilaterally, in a global pattern strongly reminiscent of the proposed default state of the brain (98). This would imply that this default state, present during the rest period between epileptic discharges, is suspended during these discharges. The mechanism by which this occurs may be related to thalamic activity but this deactivation does not result directly from firing of neurons that are part of the epileptic discharge, but from the indirect effect of the discharge on attention mechanisms. It is now possible to perform similar studies in experimental animals and these may shed some light on the human results (130, 131).

When BOLD responses occur in multiple regions, particularly in focal epilepsy, the question arises as to the possibility that these different areas could correspond to propagation of the interictal discharge, or to distant sites particularly sensitive to the effect of epileptic discharges. The time resolution of the fMRI does not allow the measurement of propagation times of a few milliseconds. Using EEG source modeling, it may be possible to assess propagation if the model includes EEG sources in the same regions as BOLD responses. It may also be possible that the pattern of BOLD response itself is different in the primary epileptogenic region and in the region in which the activity propagated (39). It would be interesting to investigate methods of assessing functional connectivity from fMRI data (132).

Most studies have been done with 1.5-T MR machines, although a few have used 3 T. As more studies are performed with 3 T, one might expect a better sensitivity in detection of BOLD changes, but also worsening of some problems: the susceptibility artifact will result in higher signal loss, the ballistocardiogram and head movements will result in worse artifacts in the EEG. Since there are good methods for artifact removal, one can hope that studies at 3 T will increase the yield of EEG-fMRI studies in epilepsy.

In conclusion, combining EEG and fMRI appears to be a very promising tool in the study of epileptic discharges. Although the technique is not simple, it is now very feasible and opens a new way to investigate the source and the effect of epileptic activity. Its application to individual patients for the purpose of localizing epileptogenic regions is not yet warranted, as more needs to be learned about the meaning of the various responses. It could be considered, however, in the context of indicating potential regions for further investigation, such as focused anatomical MRI analysis or possibly electrode implantation.