An Electro-Encephalogram beta gap after induction with diazepam: A localization method in epileptogenic lesions

Clinical Neurophysiology 120 (2009) 1235–1244

Contents lists available at ScienceDirect

Clinical Neurophysiology

journal homepage: www.elsevier .com/locate /c l inph

An Electro-Encephalogram beta gap after induction with diazepam:A localization method in epileptogenic lesions

Steven Claus a,*, Frans Leijten d, Paul Kallansee a, John Klepper a, Fernando H. Lopes da Silva f,Hanneke Ronner e, Demetrios Velis a, Max A. Viergever c, Stiliyan Kalitzin b

a Department of Clinical Neurophysiology, Stichting Epilepsy Instellingen Nederland (Epilepsy Institutes of The Netherlands Foundation), SEIN, Achterweg 5, 2103 SW,Heemstede, The Netherlandsb Department of Medical Physics, Stichting Epilepsy Instellingen Nederland (Epilepsy Institutes of The Netherlands Foundation), SEIN, Heemstede, The Netherlandsc Image Sciences Institute, University Medical Center Utrecht, Utrecht, The Netherlandsd Department of Clinical Neurophysiology, University Medical Centre Utrecht, Utrecht, The Netherlandse Department of Clinical Neurophysiology, VU Medical Centre, Amsterdam, The Netherlandsf Swammerdam Institute of Life Sciences, University of Amsterdam, Amsterdam, The Netherlands

a r t i c l e i n f o

Article history:Accepted 11 April 2009Available online 22 May 2009

Keywords:EEGLocalizationEpilepsy surgeryBeta-activityGamma-activity

1388-2457/$36.00 � 2009 Published by Elsevier Ireladoi:10.1016/j.clinph.2009.04.007

* Corresponding author. Tel.: +31 23 5588182; fax:E-mail addresses: [email protected], stephanotus@gm

a b s t r a c t

Objective: We clinically tested a quantitative EEG method to localize abnormal variations in benzodiaz-epine-induced fast rhythms to localize focal epileptogenic lesions, assuming altered quality/quantity ofGABA receptors in the lesions.Methods: During a 64-channel-EEG (sampled at 1 kHz) recording benzodiazepines were administered tofive patients with localization related epilepsy associated with an MRI visible focal lesion. We determinedthe post-injection dominant spectral modulation using Gabor wavelets and analysed the symmetry ofspatial distribution. This was compared to the localization of the lesion on the MRI scan.Results: The principal component was found in the beta/gamma band. In all patients one region ofdecreased change was associated with the lesional hemisphere, and overlapped with the site of the lesionin four. Three patients underwent surgery: interictal corticographic findings concurred with the area ofdecreased benzodiazepine response.Conclusions: This simple method localized abnormal function associated with epileptogenic lesions. Fur-ther methodological validation is now justified. Final clinical validation must be done in MRI-negativecases as well.Significance: This research may lead to techniques for non-invasive easy localization of epileptogenic tis-sue that is not visible on a structural MRI scan.

� 2009 Published by Elsevier Ireland Ltd. on behalf of International Federation of ClinicalNeurophysiology.

1. Introduction

Localization of aberrant cortical areas that are pivotal to thegeneration of seizures is critical to diagnosis, classification and cer-tain types of therapy, but it often consumes valuable time and clin-ical resources. A case in point is the localization of the ictal onsetzone in the preoperative evaluation for resective epilepsy surgery.That zone is usually not restricted to the anatomical lesion (as de-tected on the structural neuroimaging), if located there at all(Holmes et al., 1999). However, successful localization of the onsetzone is a decisive condition for favourable post-surgical outcome interms of seizure reduction (Engel et al., 1996). Moreover, some-times there is strong evidence of a focal cortical lesion that maycontain the onset zone of seizures, but it is not visible on the struc-

nd Ltd. on behalf of International F

+31 23 5588179.ail.com (S. Claus).

tural MRI scan. Also, the area critical to generation of seizures usu-ally extends beyond the borders of the onset zone. Forapproximation of this total area – the epileptogenic zone – anextensive battery of diagnostic procedures, including very invasiveones, is needed. Only the post-surgical outcome defined by thereduction of seizures (Engel et al., 1996) clinically determines ifthe approximation or demarcation was successful or not. Thereforenew techniques are welcome that prospectively define epilepto-genic areas and reduce the number of auxiliary techniques whenattempting to localize the epileptogenic lesion.

Scalp Electro-Encephalography (EEG) could again be a valuableplatform. It measures a close consequence of neuronal function:the integrated excitatory and inhibitory post-synaptic field poten-tials (EPSP/IPSP) as transmitted to the scalp through volume con-duction. One need not make indirect assumptions that wouldreduce the validity of a conclusion. This is in contrast to other tech-niques used for localization purposes such as BOLD-fMRI or SPECT

ederation of Clinical Neurophysiology.

1236 S. Claus et al. / Clinical Neurophysiology 120 (2009) 1235–1244

that assume changes in blood flow to be correlated to changes inneuronal function. EEG still has a broad clinical use, because manyolder limitations are successfully negotiated by new technicaldevelopments. Recognition of interictal spikes and spike waveshas improved through clever use of mathematical analysis of thesignal (Adjouadi et al., 2004a,b, 2005; Jung et al., 2000). Otherdevelopments – such as digital recording, associated high samplefrequencies, high(er) density of electrodes and higher capacitiesfor data storage – may further help to increase the clinical potentialof the EEG.

Interesting in context of localizing aberrant cerebral tissuethrough measurement of function is the development of a tech-nique to localize epileptogenic lesions using induced fast brainactivity. Perturbation – application of an external disrupting im-pulse to a system – of the brain with barbiturates almost immedi-ately produces fast activity in healthy tissue, but less or none inlesional brain tissue. This asymmetry in distribution might thuspoint to a lesion (Pampiglione, 1952) (Hufnagel et al., 1990,1992). Nevertheless, after the first introduction in the sixties, theimpact of this technique has been minimal. We attribute this tonot having an appropriate hypothesis at the time and, to a certainextent, the advent of anatomical imaging less than two decades la-ter. Regarding the first issue in particular: neither the concept ofnetwork based EEG signal generation nor knowledge of the dynam-ics of the system generating that signal existed at that time. Fur-thermore the effect-mechanisms of barbiturates on the cellularlevel were unknown when Pampiglione published his seminalwork.

Several decades later ligand-based FMZ (Flumazenil)-PET scan-ning has provided a possible explanation for the lack of or changein benzodiazepine/barbiturates induced fast activity in epilepto-genic lesional tissue, and consequentially for the localizing valueof this activity for epileptogenic lesions. FMZ is a competitive ben-zodiazepine antagonist that binds to the benzodiazepine receptorson the ligand-gated ion chloride (Cl�) channels where benzodiaze-pines have a changing effect on the allosteric binding properties of

Fig. 1. Gabor time–frequency representation (activity over all electrodes averaged) whilines), ‘post-benzodiazepine-epoch’ (between second set of yellow lines) and target freqmoment at which the benzodiazepine was administrated. X-axis: recording time in minufree of artefacts and to enable the patient to assume a comfortable supine position the tfrequencies found in the signal. Colour pseudo scale ranging from deep blue (frequencyappear in bursts most likely represent muscle activity. The continuous narrow frequencyin this figure caption the reader is referred to the web version of the article.)

GABA through which the channel opens more often to Cl�-ions.This hyperpolarizes the cell more than without the binding of ben-zodiazepines. The hyperpolarization is associated with the induc-tion of beta-band rhythms (sometimes called the ‘beta-buzz’) inthe EEG (Haenschel et al., 2000). FMZ-PET scanning thus enablesvisualization of the benzodiazepine receptor sites on the GABAcomplex (Heiss and Herholz, 2006). It has shown the GABAreceptor to be abundant in the cortex and very sensitive to damage(Heiss and Herholz, 2006) which supports the assumption that thequantity and quality of the receptors is altered in lesional neuronaltissue. As has been shown, certain types of lesions indeed do nothave the same quantity of benzodiazepine/barbiturate-GABAreceptors as healthy tissue, also the structure of these receptorsmay have been altered (Richardson et al., 1998; Arnold et al.,2000; Hammers et al., 2001a,b; D’Antuono et al., 2004; Morimotoet al., 2005). Finally, decreased FMZ binding in specific lesionshas been associated with epileptogenicity of these lesions (Arnoldet al., 2000). It could thus be expected that this method pointsmore specifically to epileptogenic lesions.

With the advent of more powerful quantitative EEG measure-ment localization of epileptogenic zones based of paucity onbeta/gamma activity after global induction holds a theoreticalpromise of being non-invasive and very time efficient. Moreoverit is cheap and available to a large number of hospitals. It thereforebecomes interesting to retest this old method, now supported by atheoretical model. In this pilot study we tested if it is possible tolateralize to and to localize an epileptogenic lesion (as seen on astructural MRI scan) in a small number of patients, by measuringthe distribution of benzodiazepine-induced brain activity.

2. Materials and methods

2.1. Patients

In this study we recruited patients who were eligible for epi-lepsy surgery because of localization related epilepsy (LRE) associ-

ch was used to select the ‘pre-benzodiazepine-epoch’ (between first pair of yellowuency band (between the set of horizontal blue lines). The red line represents the

tes (counting from zero). Because some time was needed to make the EEG recordingotal recording time exceeds the necessary recording time of 15 and 20 min. Y-axis:least prevalent) to red (frequency most prevalent). The broad frequency bands thatbands probably represent cerebral activity. (For interpretation of colour mentioned

Fig. 2. Representation of the computed spectral modulations. Top diagram, ‘PC (principal component)-time’: the decomposition into zero-mean and the mean over the timesegment; X-axis length of the epochs in seconds (the separation between the two epochs is represented by the red line). Second diagram from the top, ‘PC frequency’: thesubsequent factorization and the representation of the leading time-evolution component – LTC – (frequencies of 56 Hz and higher, but a second important component isfound around 30 Hz). The diagram in the bottom left-hand corner represents the channel-frequency specific weights; notice the activity in the gamma band in the rangebetween the two ‘R’ marks on the X-axis (channels over the right hemisphere). The diagram in the lower right-hand corner represents the topographical representation of thechannels. The colour pseudo scale represents the intensity of modulation of the leading frequency component (the colour red represents the most while the colour bluerepresents least modulation). Note the conglomeration of dark blue coloured electrode positions in the field with uneven numbered electrodes (coding for electrodes on theleft side of the head); this represents the so-called ‘cold spot’ where the dominant spectral modulation – DSM – is least present. (For interpretation of colour mentioned in thisfigure caption the reader is referred to the web version of the article.)

S. Claus et al. / Clinical Neurophysiology 120 (2009) 1235–1244 1237

ated with a lesion identified on the structural-MRI (sMRI) concor-dant with the results of presurgical scalp EEG/CCTV (Closed CircuitTeleVision) seizure monitoring (for localization and type of the le-sion and patient characteristics see Figs. 3–7). These patients werenot prescribed benzodiazepines regularly, and had not used thesepreparations in the two weeks prior to recording. Five consecutivepatients were recruited for this study.

2.2. Clinical protocol

We used a protocol for data acquisition (approved by two insti-tutional review boards). A peripheral intravenous catheter wasbrought in place before the EEG recording started. The patient as-sumed a supine position with the eyes open. If patients becamedrowsy (concurring with slowing of the EEG) they were keptawake by voice stimuli from the investigator; the patient had beeninstructed not to reply to such stimuli. This way we minimizedfluctuations in the background pattern and occurrence of muscleor movement artefacts in the EEG (Sinai et al., 2005).

2.3. EEG measurement

The high resolution EEG was recorded continuously from atleast 15 min prior to administration of the benzodiazepines(0.1 mg/kg bolus of Diazemuls� i.v., Alpharma BV, Baarn, The Neth-erlands) until 20 min after. The investigators were blinded to theEEG at the time of recording; quality control of the signal wasthe responsibility of a trained and experienced registered EEGtechnician. We used a 66 channel (65 EEG + one EKG channel) Sch-warzer amplifier (Schwarzer GmbH, Munich, Germany) opticallylinked to a Windows PC running HarmonieTM (Stellate Systems,Montreal, Quebec, Canada, software version 6.0). We only used

the hardware filters of the amplifier (low-pass at 300 Hz and a highpass at 0.016 Hz). The sampling frequency was at 1 kHz, at an ana-log-to-digital conversion rate of 16 bits per sample. The 65 chan-nels were connected to Ag/AgCl disc electrodes with a 5 mmdiameter. The symmetrically configured coverage was based onthe assumption that use of more than 64 electrodes does not seemto contribute much to accuracy of localization compared to mea-surements with 128 electrodes (Lantz et al., 2003; Michel et al.,2004), and on the assumption that the activity of interest will beinduced symmetrically.

2.4. Signal analysis

2.4.1. Gabor time–frequency representation (see Fig. 1 for graphicalrepresentation)

To quantify the (possibly non-stationary) spectral properties ofthe signal we selected a set of Gabor aperture functions are givenas

gðt � t0; mÞ ¼ 1Nm

e�p2a2m2ðt�t0Þ2�i2pmðt�t0 Þ � Om ð1Þ

where m is the central frequency and the product am is thebandwidth of the filter. The normalization factors Nm and the offsetfactors Om are chosen so that the functions have zero mean and unit1-norm:

X1t¼�1

jgðt; mÞj ¼ 1: ð2Þ

This normalization choice for the Gabor functions is selected sothat when convolved (see Eq. (3) below) with a quasi-stationary(compared to the aperture of the Gabor function) sinusoidal

1238 S. Claus et al. / Clinical Neurophysiology 120 (2009) 1235–1244

rhythm of fixed amplitude and of the same frequency m, the resultis a complex amplitude with a magnitude independent of the fre-quency band m (up to infinitesimal additive correction). Or, in otherwords, we have chosen equal response for rhythms of equal ampli-tude. As we shall see later, (see Eq. (10)) the normalization of theGabor functions plays negligible role in the subsequent results.We have also selected the factor a with a constant value of 0.1.The result of our choice for the Gabor set (1) is therefore a se-quence of scale-transformed filters with bandwidths of 10% ofthe corresponding central frequencies.

In this study we used a set consisting of 100 filters with centralfrequencies exponentially increasing from 4 Hz to the maximal fre-quency of 500 Hz. In other words the sequence m1; m2; . . . ; m100 waschosen so that

ðvk � vk�1Þ=ðvk þ vk�1Þ ¼ 0:1; k ¼ 2; . . . ;100:

For each electrophysiological trace (channel) FcðtÞ we define itsGabor time–frequency dependent amplitude as

Gcðt; mÞ �X

t0gðt � t0; mÞFcðt0Þ: ð3Þ

Our analysis is based on the square-root of the averaged squaremagnitude of (3) within given time interval T defined as:

Wðc; m; TÞ �ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffihjGcðt; mÞj2it2T

q: ð4Þ

For simplicity we will refer below to this quantity below as the‘‘spectrum”.

In the sequel we have used a sequence Tn of continuous timeintervals of 10 s with 50% overlap. In formula (4) therefore we shallsimply replace T � Tn with the number of the corresponding timewindow-n,

Wðc; m; nÞ �Wðc; m; TnÞ: ð5Þ

We shall refer also to the discrete number of the window-n as‘‘time” as it carries the remaining time evolution information inthe data.

2.4.2. Computing the spectral modulations (see also Fig. 2 for agraphical representation)

Our strategy is to reduce the data represented in the spectralweights (5) to one single number per channel representing themodulation carried by the signal in this channel. To this end we as-sume the presence of a particular structure in the spectral volume(5). More specifically, we expect that due to the modulating effectof the BDZ the relatively stationary activity acquires a perturbationof the form:

Wðc; m; nÞ �W0ðc; mÞð1þMðc; mÞHðnÞÞ;hHðnÞin ¼ 0; hH2ðnÞin ¼ 1: ð6Þ

In the above equation W0ðc; mÞ � hWðc; m;nÞin is the time aver-aged spectrum and Wmðc; m;nÞ �W0ðc; mÞMðc; mÞHðnÞ representsthe modulated part. The leading hypothesis here is that thedominant temporal change in the spectrum is caused by the ben-zodiazepine-perfusion process and is common for all channelsand frequency bands involved in the process. The normalizationof the time evolution vector HðnÞ is purely arbitrary.

To extract the modulation coefficients Mðc; mÞ, after removingthe averaged spectrum we perform an approximate factorizationof the three dimensional data as a product of channel andfrequency dependent coefficients and a common time evolutionfactor:

Wmðc; m;nÞ �Wðc; m; nÞ �W0ðc; m;nÞ � Uðc; mÞHðnÞ: ð7Þ

In the last representation the two factors can be uniquely deter-mined up to a mutual normalization constant by requiring that theleading time-evolution component HðnÞ – further in this work re-ferred to as LTC (leading temporal component) – and that the chan-nel-frequency specific weights Uðc; mÞ minimize the cost-function

CW ¼Xc;m;tðWmðc; m; nÞ � Uðc; mÞHðnÞÞ2: ð8Þ

This can be achieved by taking HðnÞ as the eigenvector corre-sponding to the largest eigenvalue of the matrixVðn;n0Þ �

Pc;mWmðc; m;nÞWmðc; m;n0Þ. The complimentary factor

Uðc; mÞ can then be obtained as

Uðc; mÞ ¼P

nWmðc; m; nÞHðnÞPnHðnÞ

2 : ð9Þ

To validate our hypothesis we can use the ‘‘goodness of fit” forthe factorization (7) defined as eW � 1� Cw=

Pc;m;nW2

mðc; m;nÞ. Wehave accepted data selections with eW P 0:5.

Next we define according to (6) the temporal modulation of thesignal’s spectrum for each channel as

Mðc; mÞ �Uðc; mÞ H2ðnÞ

D En

W0ðc; mÞ: ð10Þ

In the last equation the normalization choice from (6) has beenused. We have chosen to first factorize as in (7) and then to normalize(10) to the averaged spectrum. Although the modulation extractioncan be done in a reverse order as long the representation (6) is closeto exact, we have chosen for the above procedure in order to avoid thedivision by infinitesimal numbers which may affect the factorizationalgorithm in case the recordings contain ‘‘empty” channels.

The next step is to extract channel-specific coefficients from themodulations (10). To this end we apply once again the same factor-ization technique as with Eqs. (7)–(9).

Mðc; mÞ � SðcÞFðmÞ: ð11Þ

In the formula above, FðmÞ is the dominant spectral modulation(DSM). The coefficients SðcÞ represent the relative amount of mod-ulation affecting the corresponding EEG channel. This factorizationis derived from the assumption that the same spectral content ismodulated over all channels. The representation (10) is then re-quired to minimize the function CM ¼

Pc;mðMðc; mÞ � SðcÞFðmÞÞ2

and is unique up to mutual scaling between the two factors. Theacceptance criteria for the validity of the above assumption andfor the exactness of our data reduction are selected byeM � 1� CM=

Pc;mMðc; mÞP 0:8.

To remove the arbitrariness of common multiplication factor in(11), we rescale the channel weights as follows:

�SðcÞ � SðcÞmedianðSðcÞÞ : ð12Þ

Finally, we define the asymmetry index as

AðcÞ � �SðcÞ � �Sðc0Þ; ð13Þ

where c0 is the label of the contra-lateral to c electrode position. Forthe central electrodes such as Fz, Cz, etc. the asymmetry index ispostulated as zero.

On some occasions in the figures we use also the term principalcomponents (PC) for both LTC and DSM as the factorization tech-niques described above are closely related to the methodology ofthe principal component analysis where only the largest principalcomponent is retained.

The algorithms for analysis were written and sequenced in Mat-labTM (Mathworks, Natick, MA, USA, version 6.6), which enabledquick processing per case (approximately 15 min from conversion

S. Claus et al. / Clinical Neurophysiology 120 (2009) 1235–1244 1239

of the raw EEG signal to production of the final graphical represen-tation of the distribution of the principal component over all elec-trodes as represented in Fig. 2).

2.5. Validation of the paradigm

To calibrate and test our methods in a simulated 10/10-elec-trode configuration we generated synthetic signals with individu-ally selected modulations and applied the analytical procedurecomparing epochs of simulated EEG without these modulationsto epochs with the added modulations. To this end, mixtures ofseveral frequency bands components were used. Each componentwas generated as white noise with Gaussian distribution, andband-filtered with a fifth-order butter filter. The components werethen normalized to selected standard variation as a measure oftheir amplitude and subsequently added to form a mixture.

To minimize the influence of confounding non-cerebral activitythe epochs of measurement had to be as artefact-free as possible.For the first patient the original EEG signal was compared to thepower spectrum to better understand how known artefacts are re-flected in the selection window (see also Fig. 1). For selection ofepochs in the power spectra of the remaining four patients the inves-tigators were blinded to the original EEG record at the time of analysis.

2.6. Preoperative localization

In three patients, results of EEG localization by beta-asymmetryquantification after induction were correlated to the results ofintraoperative localization of epileptogenic zones by means ofAcute ElectroCorticography (ACoG) with intracranial electrodegrids (Ad-Tech, Racine, Wisconsin, USA).

3. Results

3.1. Validation of the model

As proof of principle we performed validation tests. The meth-odology described was able to reproduce – and localize – correctlythe relative changes in the amplitude for any frequency band thatwas sufficiently isolated from the rest of the components.

3.2. Epoch selection

The average power spectrum gave reliable representation of thesignal making it easy to select epochs with the least amount ofartefact (Fig. 1). Comparison between the power spectrum andthe original EEG trace of the first patient showed narrow horizontalcontinuous bands of frequencies to be correlated with cerebralactivity, whereas bursts of vertical broad band frequencies corre-lated with muscular activity.

3.3. Spectral factorization analysis

The leading time-evolution component or the LTC (Fig. 2; topdiagram) shows how much of this component was present in theepochs before and after administration of the benzodiazepines. Astepwise increase (compare the epochs before and after adminis-tration of benzodiazepines) indicates that the principal componentis attributable entirely to events associated with the administra-tion of benzodiazepines. In all five patients this stepwise increaseis (re)produced best when the epochs for comparison contain theleast artefacts.

The frequency modulation diagram depicts the dominant com-ponents of modulation, the DSM, associated with administration ofbenzodiazepines (Fig. 2, second diagram from the top). In all pa-

tients these were found in the beta or the gamma-band. Usuallythe leading component was found in the gamma band, whereasthe following components would be found in the beta-1 andbeta-2 range (15–35 Hz). This profile was characteristic for all fivepatients, also when the target frequency band in the power spec-trum selection window had been chosen up to maximum band-width (Fig. 1 the frequency range between the two blue lines).

The asymmetry index is depicted topographically in Fig. 2 (rightlower-hand corner). In all five patients the area of least modulationof the dominant component (subsequently referred to as the ‘coldspot’) was found ipsilateral to the lesion; see Figs. 3–7. The locali-zation of the ‘cold spot’ overlapped with the neocortical lesions asidentified on the MRI scan of each patient in all five cases. Theseresults were reproducible, provided the LTC showed the character-istic stepwise increase after the moment of administration of thebenzodiazepines. When analysis was performed of two epochschosen only from the pre- or the post-administration period, theseconsistent asymmetries were not reproducible.

3.4. Preoperative localization

Actual surgical treatment was performed in patients 1, 3 and 4.For patients 1 and 3 the approach was (partial) lesionectomy, tai-lored by means of Acute Corticography (ACoG) by means of sub-dural electrode grids. Patient 4 had temporal lobe epilepsyassociated with left sided Mesial Temporal Sclerosis, which wastreated with an ACoG-tailored temporal lobe resection withamygdalohippocampectomy.

For patient 1 spikes were detected in the first 4.5 cm from thetemporal pole which overlapped with the ‘cold spot’. In patient 3ACoG showed a few spikes that were located in the postcentralgyrus approximately 2 cm above the tumour. These findings con-cur with the ‘cold spot’ (parasagittal localization), although some-what more posterior to the cold spot with beta induction. The grid,however, had not been positioned over the latter area. For patient4, spikes were found in the hippocampus, the basal temporal area,and the lateral temporal neocortex both on the left in an area thatextended up to 9 cm from the temporal pole. This overlaps withthe localization of the cold spot as indicated by our quantitativeEEG (QEEG) measurements.

4. Discussion

4.1. Significance of the findings

We demonstrated lateralising and localizing potential of focalpaucity in globally induced gamma/beta activity after i.v. benzodi-azepines as measured with interictal scalp EEG and analysed withGabor function based power spectrum estimation and leadingcomponent factorization techniques in five patients. The EEGrhythms that were modulated were consistently from the samefrequency bands (the beta and gamma-bands), and the ‘cold spots’showed overlap with an epileptogenic MRI lesion and the results ofintracranial localization during Acute Corticography. This suggeststhat this easily applicable, non-invasive technique is able to detectthe presence of epileptogenic lesions in patients with LRE.

A reproducible lateralizing modulation of scalp EEG frequenciescould only be measured when a pre-benzodiazepine-administra-tion-epoch was compared to an epoch after the time of administra-tion, indicating that the benzodiazepines are associated with thechanges, and that these frequencies consistently lie within the betaand gamma bands. We took pains to prevent frequency selectionbias by choosing a wide enough target band for analysis (usuallyranging from 7 up to 90 Hz) while the sample frequency was highenough to meet with the demands brought forth by the Nyquist

Fig. 3. Patient 1. The image at the lower left-hand corner shows the MRI-slice through the tumour in the left temporal lobe. The image at the lower right-hand corner imageshows the ‘cold spot’ in the left frontolateral area. The top image shows that the DSM was generated after administration of benzodiazepines, the second image from the topshows the principal components to be 14, 30, and 56 Hz up to the bandwidth upper limit.

Fig. 4. Patient 2. The lesion was located in the left posterior temporo/occipito/parietal region. The ‘cold spot’ can be found image in the lower right-hand corner (TP7 en P9electrodes). The DSM components were found in the gamma band (around 51 Hz and between 64 and 82 Hz).

1240 S. Claus et al. / Clinical Neurophysiology 120 (2009) 1235–1244

theorem. The measured frequencies concur with the theoreticallyexpected effect of benzodiazepines on brain rhythms (Whittingtonet al., 2000) and the temporal context proves that the administra-tion of the benzodiazepines is the crucial factor for the frequencyband-related modulation to take place. No such modulations weredetected when the epochs for measurement are taken before themoment of benzodiazepine administration.

It is likely that the measured EEG effect is the reflection or con-sequence of GABA receptor manipulation (Whittington et al.,2000). Since GABA receptors are widely distributed throughout

the brain, and since certain types of lesions show a variation inGABA receptor density or quality as shown with FMZ (Flu-mazenil)-PET studies (Richardson et al., 1998; Arnold et al., 2000;Hammers et al., 2001a,b; D’Antuono et al., 2004; Morimoto et al.,2005), any consistent variations in scalp EEG frequency band mod-ulation could very well be associated with such a lesion. Thiswould place our findings of EEG analysis after beta-induction onthe same line as FMZ-PET. Therefore, one of the tasks at hand willbe to compare the results of new studies with FMZ-PET measure-ments in the same patients.

Fig. 5. Patient 3. A lesion in the left frontolateral areas with concurring with a ‘cold spot’ in the anterior frontolateral area (image in the lower right-hand corner). The DSMcomponent of modulation was found in the gamma band (i.e., at 48 Hz and between 58 and 71 Hz).

Fig. 6. Patient 4. The lesion is in the left hippocampal area (mesiotemporal sclerosis). The ‘cold spot’ is found in the left hemisphere and extends from frontopolar region tothe frontocentral and midtemporal regions.

S. Claus et al. / Clinical Neurophysiology 120 (2009) 1235–1244 1241

We are, however, aware of the limitations of this pilot study. Itprovides a proof of principle, but the true clinical significance ofthe technique has to be sought out further. We investigated a small

number of patients with single and discrete lesions on MRI andcongruent findings of EEG/CCTV seizure monitoring who were allawaiting epilepsy surgery. In the current set up it was also not pos-

Fig. 7. Patient 5. The lesion is seen in the right frontocentral region. The ‘cold spot’ is found in the right anterior frontolateral regions of the topographical map. An Electro-Encephalogram beta gap after induction with diazepam: a localization method in epileptogenic lesions.

1242 S. Claus et al. / Clinical Neurophysiology 120 (2009) 1235–1244

sible to evaluate the discriminating power between lesions that areepileptogenic as opposed to lesions that are not. A case-controlledstudy would be needed for such purposes. Nevertheless the resultsobtained are in agreement with the current concepts of epileptog-enicity in LRE besides being both consistent and intuitive.

4.2. Technical remarks

The technique has some technical limitations that will need tobe addressed before subjecting a large number of patients to thetechnique in case-controlled studies.

Firstly, it is too early to tell whether the technique we reporteddemarcates aberrant tissue in anatomical concordance with topo-graphical localization of the aberrant rhythms. The principal limi-tation that would have to be resolved is the spatial resolution ofour EEG derivations. We experimented with a reduction of elec-trode coverage (as compared to the original number of 64) and ob-tained poorer results. We have not recorded with more than 64scalp EEG electrodes, because other groups reported that an in-crease in density did not seem to add much value (Michel et al.,2004). Since we successfully lateralised, but somewhat ambigu-ously localized, increasing the number of scalp EEG electrodesmay yet be worthwhile to see if it improves localization anddemarcation in a follow-up study.

A second problem that influences the localization potential ofpharmacologically modulated EEG rhythms seems to be the (skull)tissue impedance. We have shown now that it is possible to detectrhythms in the gamma range that seem to have significance in thelocalization of (patho)physiological processes. It is interesting tofocus more on these higher frequency band rhythms. Since thetechnique we report on is well-suited for short-term, interictalmeasurements, another approach would be to use whole-head

magneto encephalography (MEG). Measurement of magnetic fieldfluctuations generated from neuronal ion fluxes is not impededby tissue that lies between the brain and the MEG sensors, whereasit offers the bonus of a much higher spatial sampling (i.e., 151 MEGsensors versus 64 electrodes in our case). We have demonstratedthat MEG measurements constitute an alternative to the EEG, freeof certain limitations imposed on the EEG by the anatomy of hu-man head as volume conductor (Parra et al., 2004).

A concerning technical dilemma is caused by the inevitableinterference of extra cerebral activity with the cerebral signal inthe (summed) EEG signal. Most of this interference can be attrib-uted to activity from the scalp muscles. Since the introduction ofEEG to measure cerebral activity, muscle activity has been knownas the most important confounder. Most of the frequencies reflect-ing muscular activity show spectral overlap with cerebral frequen-cies (O’Donnell et al., 1974) and seemingly most notably in the betaand the gamma part of the spectrum (Whitham et al., 2007). Detec-tion of rhythm and asymmetry in its distribution with the Gaborfilters in the artificially generated perturbed EEG signals with ourmodulation detection technique is unambiguous, however, whenworking with our in vivo generated data we met some degree ofambiguity. A case in point is shown in Fig. 4. In this patient thestructural MRI shows the lesion in the left temporo-occipital area,which is associated with a ‘cold spot’ in that same location. How-ever, one can also find areas of less modulation over the rightfrontolateral areas. Usually such ambiguities arose when the epochfor comparison included ‘burst like broad band frequencies’ whichare likely to have arisen from biological artefacts in the EEG e.g.muscle artefacts. The electrodes used with EEG measurements lievery close to the scalp muscles. It seems therefore interesting tostudy differences in reflection of muscle and cerebral activity whendistance of the sensors is varied. One way to do this is comparing

S. Claus et al. / Clinical Neurophysiology 120 (2009) 1235–1244 1243

EEG/MEG co recordings. Of course this idea is not new, but ourtechnique of analysis together with forced modulation of activitymay shed new light on matters. Additionally, muscle activitymay not be the only source of such ambiguities the occurrenceand probable causes of which may only be established by a moreextensive follow-up study.

Although the technique of pharmacological perturbation of theEEG works well, improvements are desirable especially in specific-ity and sensitivity studies to follow. A definite improvement wouldbe to generate stronger perturbations that improve the signal-to-noise ratio. We now used a ‘slow release’ form of diazepam (Diaz-emuls�). It reaches plasma peak levels after 20 min following intra-venous administration. We did not measure plasma levels,however, which make it impossible to know whether measure-ments have been obtained at the optimal time. Due to the manyartefact-related bursts it has been impossible to test variation ofmodulation effect with variation in epoch length. We were simplydepending on the few artefact-free epochs in the power spectrumto prevent confounding influence from muscle activity. Othermeans of optimizing pharmacologically induced perturbation isto administer barbiturates (Pampiglione, 1952; Hufnagel et al.,1990, 1992) instead of benzodiazepines since the former directlyopens the GABA/Cl�-channels, whereas benzodiazepines merelyenhance the effect of GABA on the channels. Barbiturates cantherefore be expected to give a more pronounced perturbation ofthe system generating the high-frequency EEG bands and will bethe preferred method in our next project.

The analytic methodology used in this work has been basedessentially on two assumptions. First we have assumed that therelevant power spectrum time evolution is governed, in firstapproximation, by a single time curve common for all channelsand spectral amplitudes involved. Second, we have assumed thatthe resulting spectral modulation defined by Eq. (10) consists ofa dominant channel-independent spectral component. These twoassumptions have allowed us to reduce the initial amount of datato a set of modulation weights distributed over the channels. Aban-doning one or both of the above assumptions may extend the ap-proach to cases where the spectral content of different channelschanges differently. The restriction we have chosen for however al-lows for better statistics and artefact rejection certainty.

Finally, neither normalization (12) nor definition (13) are un-ique. Different normalizations will produce different values forthe asymmetry quantity. A normalization-invariant definition suchas A(c) = (S(c) � S(c0))/(S(c) + S(c0)) would avoid the ambiguity butspecial attention in case of infinitesimally small modulations mustbe paid. As from this point on our analysis in this article is morequalitative than quantitative we do not explore all the possibilities.In a future study we will explore the quantitative associations be-tween the asymmetry index and the topological closeness to theseizure onset zone.

4.3. Historical remarks

Localization of discrete brain pathology by means of the EEG re-cord obtained through pharmacological challenge is certainly notnew. It was successfully introduced in the pre-CT era (Pampiglione,1952), but has subsequently been largely abandoned due to theintroduction of neuroimaging techniques that enabled localizationof lesional tissue. However, in case of LRE the ictal onset zone is notalways associated with the lesion visible on neuroimaging. There-fore throughout time the technique of pharmacologically activatedEEG has been used to localize epileptogenic lesions, although thiswas often limited to invasive measurement, as subtle changes inbeta-band EEG activity was hard to detect on conventionally re-corded scalp EEG performed at lower sample frequencies and re-corded non-digitally (Hufnagel et al., 1990, 1992). More recent

attempts to localize tumours by means of spontaneous occurringactivity as measured with the MEG and analysed using an equiva-lent dipole model have not succeeded (de Jongh et al., 2003). Appli-cation of more sophisticated spectral methods, such as matchingpursuit using the same Gabor filters as used in unmodulated data(Jouny et al., 2003) will probably result to an even better spectralresolution. Our current focus on pharmacologically obtained EEGbeta-band modulation is an exciting new component in view ofthe published literature of matching pursuit (Jouny et al., 2003,2004).

4.4. Clinical implications

We report on the clinical use of a non-invasive EEG techniquefor the lateralisation and localization of an epileptogenic lesion.This technique has advantages over some of the existing tech-niques. The approach is quick, cheap and relatively non-invasive.This means it can be made widely available even to hospitals innon-affluent countries that operate on small budgets. It holds fur-ther promise if it will be able to demarcate cerebral lesions that arenot detectable by the other techniques, especially structural MRI.The results of this study therefore merely mark the beginning ofa prospective study designed to test its value in MRI-negative pa-tients who are to undergo chronic invasive monitoring. Shouldthe value of this technique be established, application is also imag-inable in other fields such as classification of epilepsies, and tar-geted intervention in intractable epilepsies by means of e.g.repetitive Transcranial Magnetic Stimulation and intracerebralelectrical stimulation.

5. Conclusion

The technique of using spectral analysis and simple factoriza-tion algorithms enables to detect modulation of EEG frequenciesassociated with pharmacological EEG perturbation shows promiseof lateralization and possible localization of epileptogenic cerebrallesions. Adaptations are proposed to further refine the method.

Funding

The authors report no conflicts of interest in carrying out andreporting the results of this study.

Acknowledgements

This study was carried out with the support of Stichting Epilep-sie Instellingen Nederland (SEIN; The Epilepsy Institutes of TheNetherlands Foundation, Heemstede/Zwolle, The Netherlands).We thank the charity organisation known as ‘Christelijke Vereeni-ging voor de Verpleging van de Lijders aan Epilepsie’, Haarlem, TheNetherlands, for their annual unrestricted grant to SEIN fromwhich this projected is financially supported.

Our appreciation goes out to the colleagues who invest theirtime in patient acquisition. Our special thanks go out to René M.Ch. Debets, MD, attending neurologist at SEIN, Heemstede Campus,for referring most of the patients who participated in thisstudy.

References

Adjouadi M, Cabrerizo M, Ayala M, Sanchez D, Yaylali I, Jayakar P, et al. A newmathematical approach based on orthogonal operators for the detection ofinterictal spikes in epileptogenic data. Biomed Sci Instrum 2004a;40:175–80.

Adjouadi M, Sanchez D, Cabrerizo M, Ayala M, Jayakar P, Yaylali I, et al. Interictalspike detection using the Walsh transform. IEEE Trans Biomed Eng2004b;51(5):868–72.

1244 S. Claus et al. / Clinical Neurophysiology 120 (2009) 1235–1244

Adjouadi M, Cabrerizo M, Ayala M, Sanchez D, Yaylali I, Jayakar P, et al. Detection ofinterictal spikes and artifactual data through orthogonal transformations. J ClinNeurophysiol 2005;22(1):53–64.

Arnold S, Berthele A, Drzezga A, Tolle TR, Weis S, Werhahn KJ, et al. Reduction ofbenzodiazepine receptor binding is related to the seizure onset zone inextratemporal focal cortical dysplasia. Epilepsia 2000;41(7):818–24.

D’Antuono M, Louvel J, Kohling R, Mattia D, Bernasconi A, Olivier A, et al. GABAAreceptor-dependent synchronization leads to ictogenesis in the humandysplastic cortex. Brain 2004;127(Pt. 7):1626–40.

de Jongh A, de Munck JC, Baayen JC, Puligheddu M, Jonkman EJ, Stam CJ. Localizationof fast MEG waves in patients with brain tumors and epilepsy. Brain Topogr2003;15(3):173–9.

Engel J, Van Ness Jr PC, Rasmussen TB, Ojemann GA. Outcome with respect toepileptic seizures. In: Engel J, editor. Surgical treatment of the epilepsies. NewYork: Raven Press; 1996. p. 609–21.

Haenschel C, Baldeweg T, Croft RJ, Whittington M, Gruzelier J. Gamma and betafrequency oscillations in response to novel auditory stimuli: a comparison ofhuman electroencephalogram (EEG) data with in vitro models. Proc Natl AcadSci USA 2000;97(13):7645–50.

Hammers A, Koepp MJ, Labbe C, Brooks DJ, Thom M, Cunningham VJ, et al.Neocortical abnormalities of [11C]-flumazenil PET in mesial temporal lobeepilepsy. Neurology 2001a;56(7):897–906.

Hammers A, Koepp MJ, Richardson MP, Labbe C, Brooks DJ, Cunningham VJ, et al.Central benzodiazepine receptors in malformations of cortical development: aquantitative study. Brain 2001b;124(Pt. 8):1555–65.

Heiss WD, Herholz K. Brain receptor imaging. J Nucl Med 2006;47(2):302–12.

Holmes MD, Wilensky AJ, Ojemann GA, Ojemann LM. Hippocampal or neocorticallesions on magnetic resonance imaging do not necessarily indicate site of ictalonsets in partial epilepsy. Ann Neurol 1999;45(4):461–5.

Hufnagel A, Elger CE, Boker DK, Linke DB, Kurthen M, Solymosi L. Activation of theepileptic focus during intracarotid amobarbital test. Electrocorticographicregistration via subdural electrodes. Electroencephalogr Clin Neurophysiol1990;75(6):453–63.

Hufnagel A, Burr W, Elger CE, Nadstawek J, Hefner G. Localization of the epilepticfocus during methohexital-induced anesthesia. Epilepsia 1992;33(2):271–84.

Jouny CC, Franaszczuk PJ, Bergey GK. Characterization of epileptic seizure dynamicsusing Gabor atom density. Clin Neurophysiol 2003;114(3):426–37.

Jouny CC, Franaszczuk PJ, Adamolekun B, Bergey GK. Gabor atom density as ameasure of seizure complexity. Conf Proc IEEE Eng Med Biol Soc 2004;1:310–2.

Jung TP, Makeig S, Humphries C, Lee TW, McKeown MJ, Iragui V, et al. Removingelectroencephalographic artifacts by blind source separation. Psychophysiology2000;37(2):163–78.

Lantz G, Grave de Peralta R, Spinelli L, Seeck M, Michel CM. Epileptic sourcelocalization with high density EEG: how many electrodes are needed? ClinNeurophysiol 2003;114(1):63–9.

Michel CM, Murray MM, Lantz G, Gonzalez S, Spinelli L, Grave de Peralta R. EEGsource imaging. Clin Neurophysiol 2004;115(10):2195–222.

Morimoto K, Tamagami H, Matsuda K. Central-type benzodiazepine receptors andepileptogenesis: basic mechanisms and clinical validity. Epilepsia2005;46(Suppl. 5):184–8.

O’Donnell RD, Berkhout J, Adey WR. Contamination of scalp EEG spectrum duringcontraction of cranio-facial muscles. Electroencephalogr Clin Neurophysiol1974;37(2):145–51.

Pampiglione G. Induced fast activity in the EEG as an aid in the location of cerebrallesions. EEG Clin Neurophysiol 1952;1:79–82.

Parra J, Kalitzin SN, Lopes da Silva FH. Magnetoencephalography: aninvestigational tool or a routine clinical technique? Epilepsy Behav2004;5(3):277–85.

Richardson MP, Koepp MJ, Brooks DJ, Duncan JS. 11C-flumazenil PET in neocorticalepilepsy. Neurology 1998;51(2):485–92.

Sinai A, Bowers CW, Crainiceanu CM, Boatman D, Gordon B, Lesser RP, et al.Electrocorticographic high gamma activity versus electrical cortical stimulationmapping of naming. Brain 2005;128(Pt. 7):1556–70.

Whitham EM, Pope KJ, Fitzgibbon SP, Lewis T, Clark CR, Loveless S, et al. Scalpelectrical recording during paralysis: quantitative evidence that EEGfrequencies above 20 Hz are contaminated by EMG. Clin Neurophysiol2007;118(8):1877–88.

Whittington MA, Traub RD, Kopell N, Ermentrout B, Buhl EH. Inhibition-basedrhythms: experimental and mathematical observations on network dynamics.Int J Psychophysiol 2000;38(3):315–36.