Unsupervised Classification of EEG from Subdural Seizure Recordings

Brain Topography, Volume 10, Number 2,1997 121

Unsupervised Classification of EEG from SubduralSeizure Recordings

Werner G. Hofmann* and Manfred P.J. Spreng*

Summary: Whereas the visual EEG-inspection of epileptic seizures draws the attention to the waxing and waning of specific graphoelements inmulti-channel recordings, the domain of computerized EEG-analysis for epilepsy diagnosis is detection of transients (i.e., spikes) and the quantificationof background activity (i.e., mapping procedures). We present an approach to identify relatively fast changes of background activity by use of anautomatic classifier. This algorithm is independent of the occurrence of any specific single type of graphoelement. The EEG is segmentated into shortepochs of 0.64 sec duration each. For every segment a set of parameters (Hjorth, spectral power in classical frequency bands) is extracted, which takentogether build elements of a vector-space. The elements are clustered in an automatic and unsupervised manner by use of a cosine-classifier, suchthat every EEG-epoch belongs to one class. Changes of brain activity as seen with the EEG are marked by transitions from one class to another. Theclass occurrence density is defined as the number of different classes that occur within a pre-defined number of EEG-epochs. It gives a new measureof variability of the EEG-signal. Comparing the epochs when class transitions take place in different channels, the class transitions coincidence betweentwo channels is a measure of functional coupling of brain areas.

Key words: EEG; Epilepsy; Seizure description; Unsupervised classification; Coincidence analysis.

IntroductionThere is an unfortunate gap between visual and

computerized EEG analysis of epileptic seizures. Visualinspection draws the attention to the chronological orderof the epileptic seizure with emphasis upon the waxingand waning of specific graphoelements (i.e., spikes, pre-dominance of specific rhythmical activity such as theta),especially when comparing several EEG-channels. Com-puterized EEG-analysis puts its emphasis upon spikedetection and localization on the one side and mappingprocedures on the other. But concerning seizure descrip-tion, spikes are only one indicator, whereas power mapsof background activity can usually mark only the resultsand not the onset time of EEG changes as averaging oflonger epochs is necessary, comparing i.e., pre- vs. post-ictal state.

An approach is presented here (figure 1) to identifyand classify relatively fast changes of background activ-

*Institut fur Physiologic und Experimentelle Pathophysiologie,Arbeitsgruppe Biokybernetik, Friedrich-Alexander-Universitat Er-langen-Niirnberg, Erlangen, Germany.

Accepted for publication: August 16,1997.Correspondence and reprint requests should be addressed to Dr.

Werner G. Hofmann, Institut fur Physiologic und ExperimentellePathophysiologie, Arbeitsgruppe Biokybernetik, Friedrich-Alexander-Universitat Erlangen-Niirnberg, D - 91054 Erlangen, Universitatsstr. 17,Germany.

Copyright © 1997 Human Sciences Press, Inc.

ity by use of an automatic and unsupervised classifier.This algorithm is independent of the occurrence of anyspecific type of graphoelement selected for evaluation,as several different EEG-parameters (in this study:Hjorth-activity, Hjorth-first-order-complexity, spectralpower in the theta- and beta-band) serve as input. Inprinciple, the number of parameters is not limited andothers (i.e., from non-linear dynamics) can be added tothe list. They are derived from the segmentated rawdata (figure 1, step I) and normalized according to theextremal values for every parameter separately (figure1, step II). For every epoch, a four-dimensional vectoris formed with the explicit values of the four normalizedparameters. The classifier (figure 1, step III) formsclasses of similar vectors that describe overlappingEEG-epochs. The explicit values for a single parameterwithin a class might differ very much, so combinedclassification to some degree is independent from thetime course of a single parameter. The classes are notpredefined in any way by the expectation of the elec-troencephalographer (Spreng et al. 1993; Spreng et al.1995). Changes of EEG activity are marked by oc-curence and disappearance of classes. As the EEGchanges, the EEG-epochs within a given time slice willbelong to different classes. We define the class occur-rence density (figure 1, step IV) as the number of classesthat occur within a given time, here set to 10 epochs of0.64 sec each. The epileptogenic zone as well as thefunctional deficit zone will show low values for the

122 Hofmann and Spreng

STEPS OF ANALYSIS

I. Segmentationepoch length 0.64 sec = 128 values, single channel

"EEG printout", fig. 2a-d

II. ParametrizationHjorth activity, f-o-complexity, power in theta- and beta-band

"parameter time course diagram", fig. 3a,b

III. Classificationcosine classifier

"class occurrence diagram ", fig. 4a, b

IV. Evaluation of class occurrence densitynumber of classes that occur within 10 epochs"class occurrence density diagram ", fig. 5a-d

V. Evaluation of class transition coincidencetable 1

Figure 1, Steps of analysis: after (step I) segmentation,(step II) parametrization for parameters Hjorth activity,mobility, first-oder-complexity, and power in the theta (4-8Hz), and beta-band (14-35 Hz) is made. Normalized tomin-max, parameters form a four-dimensional vectorspace with one element for every epoch. Classification(step III) is with the cosine-classifier. The number of differentclasses that occur within 10 epochs gives the (step IV) classoccurrence density. Transitions between classes may hap-pen independently in several channels. This is tested withthe (step V) class transition coincidence.

class occurrence density all the time, as due to the"rigidity" (Wang et al. 1994) of the "clockwork focus",the EEG recorded from the epileptogenic focus will onlychange little. Post-ictally the class occurrence density isa measure of the involvement of a brain area in theseizure, whether it is seizure onset zone or only in-volved in secondary generalization. Ictally the valueswill be low in all channels where a seizure equivalenttakes place, as during the seizure in most cases only onerhythmical component is dominant at a time. Couplingof possibly far away areas of the brain can be wellidentified with the class transition coincidence (figure1, step V). In case of functional coupling, as happensduring the epileptic seizure, a shift from one class toanother in different channels will happen in coincidenc,even if the explicit recorded graphoelements differ.Data from a subdural recording from a patient withpharmacoresistant epilepsy of left mesial temporal ori-gin is shown (figure 2a-d).

Methods

Electrodes used are stripe and foramen ovale elec-trodes (AD-Tech Corp., U.S.A.). Data aquisition is witha 64 -channel Video-EEG Neurosys(tm) (Glonner Elec-tronic, Germany). Filter settings are 0.5 - 70 Hz, amplifi-cation is set to 20uV/mm. Digitization rate is 200 Hz,running MONITOR 7.0(um) (Stellate Systems, Canada)that is used routinely for spike detection. The data of thecase example are from the epilepsy monitoring unit of theneurological clinic of the university.

The algorithms used in our study for parametriza-tion and classification are all written in C and are anoptional part "Neomic" of an EEG visualization andanalysis tool XEBS (X_Window Extensible Biosignal Sys-tem) developed by the Biocybernetics working group atthe Institute for Physiology und experimental Patho-physiology. These programs run under Unix both on PC(Linux) and Sun Spare and will be implemented underWindows NT(tm) soon. Analysis of 10 minutes of EEGtakes about 8 sec / channel (mainly necessary for spec-trum analysis) with a fast Intel Pentium under Linux andX_Window. Data are first converted from the MONITOR7.0(tm) file format to a freely available data format EBS(Extensible Biosignal Format) developed at our institu-tion (Hellmann et al. 1996).

The steps of analysis are shown in figure 1. AfterEEG segmentation (figure 1, step I), spectral analysis ofall the EEG-epochs and evaluation of all parameters isdone (figure 1, step II). Parameters used in this studyare Hjorth activity, Hjorth first-order-complexity(Hjorth 1970; Chavance et al. 1976), and power in thetheta (4-8 Hz), and beta-band (14-35 Hz), all calculationsbeing made in the frequency domain, but can easily beextended to the time domain. Motivation for the EEGfrequency bands selected and Hjorth parameters comesfrom clinical observation. Theta is in most cases gener-ated from limbic structures that are very often involvedin the generation of epileptic seizures (Blume et al.1984). Beta activity is regarded as the "high frequency"activity of the brain. Concerning seizure development,it is not known whether this activity is an ictal sign ofthe a mesiobasal focus or is still pre-ictal desynchroni-zation (Wieser et al. 1994). For evaluaton of Hjorthparameters in the frequency domain (Chavance et al.1976), the spectrum and its first and second derivativeare taken as input. Let Al be the sum of all squaredspectral values during an epoch of 0.64 sec duration, letA2 be the same for the first derivative of the signal andA3 for second derivative. Then, by definition of Hjorththe parameter activity is square root of Al, and is ameasure of amplitude variation. This will fluctuateduring the awake state as there is low-amplitude activ-ity during mental task and high amplitude (sign of

Computerized Seizure-EEG Classification

Figure 2a, EEG printout (amplitudes set to arbituary units): 9h 16min 40 sec is epoch 125: Inter- / pre-ictal EEG.

Figure 2b. EEG printout: 9h 17min 20sec is epoch 312: Seizure onset. First bioelectric changes are in foramen ovale left(channel 17), some low amplitude beta in left temporal-anterior and left temporal-posterior leads (i.e., channel 25).

123

Hofmann and Spreng

Figure 2c. EEG printout: 9h 17min 45sec is epoch 351: Seizure generalization.

Figure 2d. EEG printout: 9h 18min 47sec is epoch 448: End of bioelectric seizure signs in channels from foramen ovale left(i.e., channel 17), left temporal-anterior and -posterior (i.e., channel 25) electrodes. In channels from right hemisphere(i.e., channel 1, right hippocampus) and fronto-lateral left (channel 40) rhythmical activity still continues for 12 sec.

124

Computerized Seizure-EEG Classification 125

Figure 3a. Time course of parameter Hjorth activity inchannel 17 foramen ovale left, normalized to extremalvalues. Seizure onset is after 180 seconds, seizue ends after300 seconds.

Figure 3b. Time course of parameter beta-power in chan-nel 17 foramen ovale left. Seizure onset after 180 secondsIs not marked that clear as in figure 3a.

synchronization) during rest-awake state and also dur-ing the late phase of a seizure. Hjorth parameter first-order-complexity is square root ((A3/A2)-(A2/A1)),regarded as a measure of spectral purity of the EEG andis reduced in case of polymorphic activity as recordedabove a cortical lesion. By evaluating these many pa-rameters we hope to obey the warning that besideschanges of spike density there is a great variation ofseizure initiaton patterns (Spencer 1994). All parame-ters are normalized to the interval [0,1] according totheir minimum and maximum values. Taken together,the parameters form a four-dimensional vector space.From this, "parameter time course diagrams" can beextracted as for instance shown for Hjorth activity andbeta-power (figure 3a,b). They give a first impressionof the significance a single parameter has for the de-scription of the seizure course.

Next cosine classification (figure 1, step III) with a"class occurrence diagram" is made (figure 4a,b). Withthe cosine classifier, the cosine or scalar-product of twoepochs or elements (a reference- and a comparison-vec-tor) gives a measure of similarity. If a pre-defined crite-rion (i.e., 0.8 for normalized vectors) is fulfilled, theybelong to one class. Then, a new reference-vector is builtas centre-of-mass of all vectors belonging to this class.Otherwise, a new class is created. This approach is de-pendent on initial conditions, as the direction of the ref-erence vector depends on the vectors from which it iscreated. It was decided to use cosine classifier instead ofthe well known greatest-nearest-neighbour classifier, asthe implicit learning ability (resulting from dependanceon initial conditions) is of use when describing the

bioelectric seizure structure and development.As every epoch belongs to exactly one class, the

number of different classes (that a predefined number of10 epochs are assigned to) is shown as a "class occurrencedensity diagram" (figure 1, step IV; figure 5a-d). Consid-ering a moving window of 10 epochs and a pre-set maxi-mum of 6 classes for all recording time, there are from 1to 6 different classes within the time slice. The maximumnumber of classes must not exceed the number of epochswithin the time slice but still be above the number ofassumed bioelectrical defined brain states and epochswith artifacts.

Finally (figure 1, step V), we test the null hypothesesthat the class transitions in two channels happen inde-pendent by, thus indicating a possible functional cou-pling between channels. We follow an approach used forspike-detection in multi-channel recordings (Guedes deOleveira et al. 1980):

Let the number of class transitions in channel i be Ni,in channel j Nj and the number of class transitions occur-ing simultaneously in channel i, j be Nij. Let k be numberof epochs. Then we define a classs transition coincidencechi3 = k*(abs(k*Nij Ni*Nj)-0.5*k)2 / Ni*Nj*(k-Ni)*(k-Nj)with k>=40 which has a chi2 distribution for k largeenough.

The null hypotheses is that there is no time coinci-dence between class transitions. This algorithm onlycompares pairs of channels. If groups (triads, tetrads) areto be compared, this must be made in an iterative mannerthat is not commutative, such that ((a,b),c) != ((a,c),b) insome cases.

126

Figure 4a. Class occurrence diagram channel 17 foramen ovale left. Number of classes is 22. Number of epochs wherea class occurs is given in brackets (i.e., 52 epochs are assigned to class 1). Seizure onset is after 180 seconds.

Figure 4b, Class occurrence diagram channel 1 foramen ovale right. Number of classes is 23. Seizure onset is after 225seconds.

Hofmann and Spreng


Illustrative Case Example

Data are from a 39-year-old female patient with com-plex partial seizures since age twenty. The interictal EEG(scalp and chronic subdural recording) shows a bitemporalindependent delta-theta focus with right temporal sharpwaves, left temporal sharp waves and a few spike-wavecomplexes. CT showed a left temporal space occupyinglesion of unclear type, MRI showed left temporal parahip-pocampal signal enhancement and two SPECT scansshowed inter-ictal minor hypoperfusion and ictal hyper-perfusion in the left temporal. The electrode montage issymmetrical with the two foramen ovale electrodes forregistration from the hippocampal area and three fourcontact strip electrodes at the temporal-basal-temporal-an-terior, -posterior and fronto-lateral on both sides. Note theforamen ovale left electrode (channel 17, which is seizureonset zone) and foramen ovale right (channel 1, for com-parison and as indicator of functional hippo-hippocampalconnection), temporal-posterior left (Channel 25, to showinvolvement of temporal lobe in this seizure of mesio-tem-poral origin) and fronto-lateral left (channel 40, to indicateinfluence upon frontal lobe) leads.

The electroencephalographic seizure description ofa 475 seconds recording:09:14:00 or epoch 0 (of 0.64 sec duration): maximum

delta/theta-activity is seen by pre-ictal in foramenovale left (channel 17) and temporal-posterior left(channel 25). There is also some spike-bursting inforamen ovale left of 2 sec duration, (figure 2a, 9h16min 40 sec, epoch 125).

09:17:07 / epoch 292 after 187 sec later: seizure starts with2/sec spikes in foramen ovale left (channel 17).

09:17:20 / epoch 312 after 200 sec: continous 20/sec beta-activity in foramen ovale left (channel 17), some low-amplitude beta in temporal-anterior and -posteriorleft (channel 25) (figure 2b).

09:17:30 / epoch 328 after 210 sec: continous theta in fora-men ovale left (channel 17), generalized high-fre-quency in bilateral fronto-lateral electrodes (channel40).

09:17:45 / epoch 351 after 225 sec: higher amplitudebeta-activity now emerges also in foramen ovaleright (channel 1) and theta waves fronto-lateral right,followed by generalized rhythmical theta (figure 2c).

09:18:47 / epoch 448 after 287 sec: abrupt end of seizure,first in foramen ovale left (channel 17), temporal-an-terior and -posterior left (channel 25) (figure 2d).

09:18:59 / epoch 467 after 299 sec: abrupt end of seizure inelectrodes from right hippocampus (channel 1), righttemporal and right frontal lobe; rhythmical activity infronto-lateral left (channel 40) stops only now.

09:20:59 / epoch 654 after 419 sec: post-ictal delta activityis seen in both fronto-lateral leads (channel 40). Thereis also high frequency beta above 30/sec in both fora-men ovale (channel 1,17) superimposed on slow delta.

09:21:55 / epoch 743 after 475 sec: end of recording.

From this EEG seizure description we define a pre-ictal phase with epoch 1 - 292, a localized ictal phase fromepoch 293 - 351, a generalized ictal phase from epoch 352- 467 and a post-ictal phase with epoch 468 - 743. This isan operational decision and maybe at times somewhatproblematic, as bioelectric seizure equivalents may startand stop in different electrodes at very different times,but without this simplification comparison of class tran-sition coincidences would be impossible.

ResultsIn some cases transitions between bioelectric de-

fined brain states can be seen in the time course ofparameters themselves, but are difficult to interpret.Hjorth activity in electrode 17 from foramen ovale left(figure 3a) increases continously following seizure onsetafter 180 sec / epoch 281 and stays high post-ictally after285 sec / epoch 448, though the variation of this parame-ter increases, too. First-order complexity decreasesshortly before and in the early phase of this seizure.Changes of theta are seen as a reduction of relativetheta-power ictal and increasing values again post-ictal.Beta-power behaves inversely (figure 3b) and changesare most obvious at seizure onset with initial beta-rhyth-micity after 200 sec / epoch 312 and post- ictal reduction.Extreme minimal values around 260 sec result from anintra-ictal electrode artifact. The time course of one sin-gle parameter alone may mark seizure onset and endingvery well, but this may vary as seizures may be differentaccording to the graphoelements generated even whencomparing several seizures from the same patient oreven single channels from the same seizure only. Also,the visual comparison of the time courses of many dif-ferent parameters for a great number of channels is avery time consuming task.

Taking together all time courses of all parameters, theclass occurrence diagram is the synopsis of the parametercourses. The number of classes should be set higher thanthe number of assumed (patho-)physiological brain statesas shown with EEG background activity and artifacts.Here, we assume that at least epochs with inter-ictal back-ground, pre-seizure low-amplitude and high-frequencyactivity, rhythmical theta, EEG-signs of clonk and tonicphase, post-ictal depression without and with delta, andepochs showing only technical artifact may occur. If thenumber of classes is set very high (i.e., 24 classes for 6 to 8assumed brain-states), 2 to 4 classes will mark more or lessthe same brain state. It should be kept in mind that therecording time is long enough so that number of epochsexceeds about tenfold the number of classes.

In foramen ovale left channel 17 of the case examplethe class occurrence diagram for 24 classes is shown (fig-ure 4a). The seizure starting after about 180 sec / epoch

Figure 5b. Class occurrence density diagram channel 1 from right hippocampus.

281 recording time is marked most obviously by vanish-ing of classes 1-8 and 12. No new classes occur at once,but as the seizures evolves, the occurrence density ofmany classes changes. Class 1 - 8 are markers of theseizure, even indicating the generalization with class 1,5and 7. Class 9 indicates the still partial bioelectric seizuresigns. Classes 15 - 21 also indicate the seizure. Theseclasses vanish after the end of the seizure. If only very fewepochs belong to a class, this is hard to interpret and

therefore those classes should not be taken into furtherconsideration. The post-ictal state is most impressive withclass 4 and 7 after 285 seconds / epoch 448.

In foramen ovale right electrode channel 1, that isonly later influenced by seizure propagation according tovisual inspection of the EEG, first class changes can beobserved after 225 sec / epoch 351 with a stop of occur-rence of classes 1-7 (figure 3b). Post-ictal signs are seenin classes 6,9 and 22 occurring again, and an almost total

Figure 5a. Class occurrence density diagram channel 17 from left hippocampus. Only 6 classes (instead of a maximumof 24) are pre-set, as this number must not exceed number of epochs of the time slice taken for evaluation.

Hofmann and Spreng128


Figure 5c. Class occurrence density diagram channel 25 from left basal posterior temporal lobe.

Figure 5d. Class occurrence density diagram channel 40 from left frontal lobe.

cessation of classes 17 and 19 only then. The post-ictalbrain state here is only reached after 300 sec. Though notseizure onset zone, the right temporal lobe is affectedstrongly by the seizure originating from the left mesial-temporal structures so it cannot reach pre-ktal statesuntil the end of recording time.

Every epoch is assigned to exactly one class. Theclass occurrence density is defined as the number ofdifferent classes that occur within 10 epochs (figure 1,

step IV). The class occurrence density changes during aseizure, and are visualized as a class occurrence densitydiagram (figure 5a-d). We set the number of classes toonly 6 instead of 24 here. So only very few class transi-tions will mark a change of bioelectric defined brain state.Taking 24 classes as with the class occurrence diagram,classes would show a fine structure with more or lesssimilar graphoelements. Several epochs would be neces-sary to indicate the transition between bioelectric defined


brain states as according to the algorithm only one classtransition can occur within one epoch.

Regarding channel 17 that is representing activityfrom left hippocampus most often 4 classes are reachedwithin 10 epochs before seizure onset after about 290epochs or 180 seconds. During the seizure with its manydifferent graphoelements (i.e., rhythmical activity, spikeand slow-wave activity, movement artefacts) occuring,this class occurrence density value is even reduced from3-4 down to 2 - 3 after 225 seconds / 350 epochs,indicating generaliziation of the seizure. Lowest valuesof 1 - 2 are only reached within the post-ictal epochs after320 seconds / 500 epochs. The class occurrence densityvalues remain low, showing that this brain area is hardlyable to react to external or internal stimuli (figure 5a).Comparing the two foramen ovale electrodes, channel 17from foramen ovale left (figure 5a) and channel 1 fromforamen ovale right (figure 5b) with the class occurrencedensity diagram, the class occurrence density values areabout the same in right hippocampus channel 1. Whereaspre-ictally the differences are minor, the reduction of theclass occurrence density happens somewhat later in fora-men ovale right after 250 seconds / 400 epochs, markingthe generalized activity at the end of the still localizedinitial part of the seizure.

The ictal phase starts later in channel 1 and has anabrupt end after 500 epochs. This is in accordance withthe class occurrence diagrams (figure 4a,b). Post-ictallythe class occurrence density values are higher for theright hippocampus than for the left (right 2 - 3 vs. left 1 -2). As the area below foramen ovale left is the seizureonset area, this means that the left hippocampus is func-tionally insufficient and cannot express a large amountof different EEG graphoelements. This is in accordancewith the imaging data techniques for this patient markingparahippocampus on left side as abnormal. Left tempo-ral-posterior channel 25 (figure 5c) shows a relatively

high value of 3 - 4 for the pre-ictal epochs. It is not theseizure onset zone and only affected between 255 - 305seconds or during epochs 400 - 480, which is the general-ized phase of the bioelectric seizure signs. It reaches anindex of 3 again rather exactly at the end of the seizure(299 seconds / epoch 467).

According to these results, the left temporal neo-cortex has a stronger functional connection to the lefthippocampus (channel 17) than there is between bothhippocampi. In the leftneocortical temporal lobe, seizureonset is markedly clearer than in right hippocampus, butstarts later than in left mesial temporal areas and theneocortex is not that much affected by the seizure activ-ity, as seen with the post-ictal class occurrence densityvalues that reach pre-ictal values there earlier. Leftfronto-lateral channel 40 (figure 3d) has pre-ictal highclass occurrence density indexes not very much differentfrom the other electrodes. This area, too, is only affectedduring seizure generalization after 225 seconds / 350epochs. The class occurrence density index remainshigher ictally than in the other channels, but pre-ictalvalues are not reached again in the post-ictal phase. Theability to restore the pre-ictal brain state depends only onseizure involvement of this area and is not an indicatorof the seizure onset zone.

Classification of EEG background activity gives ahint to quantify functional sufficiency of underlyingbrain tissue. Low values of the class occurrence densitypre-ictally mark functionally insufficient areas. Post-ic-tal reduction of class occurrence density is a result ofseizure involvement that need not correlate to seizureonset. Bioelectric seizure signs result in a greater uni-formity of the classes, though many different gra-phoelements may occur.

Finally, the class transition coincidence is evaluated(figure 1, step V; table I). We compare a pre-ictal phase(up to 187 sec / epoch 1 - 292), an early ictal phase where

Table 1. significant class transition coincidences for combinations of channels with channel 17 from foramen ovaleelectrode near left hippocampus, channel 1 right hippocampus, channel 25 left posterior basal temporal, channel 40 left-basal frontal lobe, for total recording time and pre-ictal (epoch 1 - 292), localized ictal (epoch 293 - 351), generalizedictal (epoch 352 - 467) and post-ictal phase (epoch 468 - 743).

17&1

19,60

10,92

1,17

3,12

2,31

2,80

25&1

0,04

0,03

0,68

1,62

1,29

10,02

40&1

0,32

0,87

0,82

0,05

0,02

12,69

17&25

8,07

0,01

2,98

6,23

15,95

5,96

17&40

7,18

0,03

0,08

1,19

0,00

8,21

25&40

0,97

1,30

0,09

2,86

0,40

19,64

(17&1)&25

3,08

0,00

1,91

0,00

2,62

0,24

(17&1)&40

5,02

0,00

0,00

0,05

0,90

1,81

((17&1)&25) &40

2,16

0,15

0,00

0,01

0,70

1,18

(17&25)&40

2,60

0,38

0,12

2,40

0,03

9,35

(17&25)&40

8,47

3,05

0,00

7,73

2,86

4,14

channels (single and combined)

total recording time

pre-ictal (< epoch 293)

local ictal (epoch 293-351)

generalized ictal (epoch 352-467)

total ictal time (epoch 293-467)

post-ictal (epoch 458-743)


rhythmical activity is still localized (epoch 293 - 351 / upto 225 sec), a generalized seizure phase (epoch 352 - 467/ up to 299 seconds after start of registration) and finallythe post-ktal phase (epoch 467 - 743 / 475 sec). It must bekept in mind that a secondary generalized seizure has nofixed time for transition to the ictal phase for all channelsas some channels are only engaged late in the time course.Also, there need not be a fixed endpoint for all electrodepositions, so the definition of bioelectric phases of a sei-zure is somewhat arbitrary in secondary generalized epi-lepsy. Nevertheless, the number of transitions is ratherthe same for all channels and all phases of this recording.

There is a highly significant coupling of both hippo-campi (channels 17 and 1) pre-ictally. The coupling be-tween left hippocampus and left temporal lobe (channel 17and 25) then is very low. There is also a low couplingbetween left temporal (hippocampal) and left frontal leads(channels 17 and 40). Concerning triads, for all recordingtime except pre-ictal, the transitions in both hippocampiand left frontal channels ((channels 17 and 1) and 40)behave more dependently than both hippocampi and lefttemporal ((channels 17 and 1) and 25). In the early phaseof the seizure with activity most prominent in left foramenovale only, there is still a class transition coincidence be-tween both hippocampi (channels 17and 1), but lower thanpre-ictal. But as the left temporal neocortex is engaged inthe course of the seizure, coupling between temporalmesial and neocortical is now higher (channels 17 and 25)than pre-ictal. In the generalized ictal phase, couplingbetween left hippocampus and all the other channels be-comes significant, but also coupling of left hippocampaland left temporal electrodes (channels 17 and 25) and alsothe triads have higher values. Whereas the triad ((channels17 and 25) and 1) which indicates functional connectionbetween left temporal lobe as a whole and right hippocam-pus is significant, ((17 and 1) and 25) indicating couplingof both hippocampi to left neocortex is low during thegeneralized seizure. Post-ictal values are high for all pairsof combinations considered, showing the involvent of thewhole brain in this secondary generalized seizure.

DiscussionThe human EEG interpreter draws attention to the

waxing and waning of graphoelements of different types,occurrence of transients (i.e., spikes) and changes of back-ground activity. He continously adapts his expectationconcerning their chronological order. "Frequency and,within certain limits, amplitude are not as critical and,even though form and organization of certain transientsare rather suggestive and almost pathognomonic of epi-lepsy, most epileptic rhythms may be distinguished onlyby their paroxysmal character." (Ajmone Marsan 1961).The time course of the bioelectric seizure signs may be

different in neighbouring intracranial electrodes evenduring the same seizure (Blume et al. 1984), so a descrip-tion of channel dependencies is necessary. But the epi-leptic spike is not necessarily an indicator of aseizure-onset zone (Spencer 1994). If the electrode is faraway from the assumed generator, even a reduction ofthe interictal spike rate might be the first sign of seizureonset, marking desynchronization as a pre-ictal sign (Wi-eser et al. 1994). Maps of spectral power density markinter-ictal background activity only, as the epochs mustbe of several seconds duration and/or averaging of sev-eral epochs of short duration (Lehmann 1984; Pfurtschel-ler et al. 1986) is necessary.

Several authors draw attention to the change of ex-pression of a single parameter as the spectral edge fre-quency (Wang et al. 1994) or to specific parts of thespectrum, which is a valuable tool when comparing simi-lar brain areas from both hemispheres, but the existenceof a certain spectral peak does not have any physiologicalsignificance (Hilfiker et al. 1992). Coherence analysis in-cluding event related coherence (Andrew et al. 1996) alsorelies on a specific frequency as indicator of coactivation,though "ictal discharges in each area (e.g. left hippocam-pus, right hippocampus, left temporal neocortex, righttemporal neocortex...) are not necessarily synchronizedand differ in frequency." (Sperling et al. 1994).

There is a gap concerning the quantification of rela-tively fast events in the EEG that are not transient activityand also do not fulfill stationarity criterion to make powermaps possible. Taking four parameters for multi-dimen-sional classification and evaluation helps to get rid of thesignificance of any special graphoelements. The classifi-cation itself is done in a fast, automatic and unsuparvisedmanner as there are no predefined classes, although thetime course of a parameter might often mark a seizurestrikingly clear. In addition, there is no methodologicalproblem to do an implementation of further parameters,i.e., from nonlinear dynamics. Failure of one specificparameter for seizure detection - that might neverthelessbe of importance in another patient, another seizure ofthe same patient or even another channel of the sameregistration - is thus of minor influence upon separationefficacy as a whole. With the class transition coincidencewe are able to mark coupling between channels in thecourse of a seizure. This goes far beyond a simple seizuredetection algorithm and allows analysis of the totalbioelectric seizure structure.

A first limitation of the method is that the seizureepochs to be evaluated are selected as a result of visualEEG inspection or even from a clinical seizure descrip-tion. This might not necessarily correlate with the bioelec-tric seizure equivalent in single channels. Also, only pairsor iterative combinations of channels can be tested ac-cording to their dependency. But in subdural recordings


the number of channels is often very large (50). A pre-se-lection of channels for analysis is somewhat arbitrary.Therefore, the approach presented here can only serve ashypotheses testing of the clinician's assumptions. Whilein this study we focus on macroanatomic distinct brainareas, in focal epilepsy comparison of nearby electrodesfrom subdural electrodes is of great diagnostic impor-tance. It might be possible to analyze very local EEGchanges in the area surrounding a seizure-onset zone oran epileptogenic lesion with the method presented here.

ReferencesAjmone Marsan, C. Electrographic aspects of epileptic neuro-

nal aggregates, Epilepsia, 1961,2: 22-38.Andrew, C. and Pfurtscheller, G. Event-related coherence as a

tool for studying dynamic interactions of brain regions.Electroenceph. Clin. Neurophysiol., 1996,98:144-148.

Blume, W.T., Young, G.B. and Lemieux, J.F. EEG morphologyof partial epileptic seizures. Electroenceph. Clin. Neuro-physiol., 1984,57: 295-302.

Chavance, M. and Samson-Dollfus, D. Analyse spectrale deI' EEG de I' enfant normal entre 6 et 116 ans: choix et validationdes parametres le plus informationnels. Electroenceph.Clin. Neurophysiol., 1976,45: 767-776.

Guedes de Oliveira, P.H.H. and Lopes da Silva, F.H. A topog-raphical display of epileptiform transients based on a statis-tical approach. Electroenceph. Clin. Neurophysiol., 1980,48:710-713.

Hellmann, G., Kuhn, M., Prosch, M., and Spreng, M. Extensiblebiosignal (EBS) file format - simple method for EEG dataexchange. Electroenceph. Clin. Neurophysiol., 1996,99:426-431.

Hilfiker, P. and Egli, M. Detection and evolution of rhythmiccomponents in ictal EEG using short segment spectra and

discriminant analysis. Electroenceph. Clin. Neurophysiol.,1992, 82: 255-265.

Hjorth, Bo. EEG analysis based on time domain properties.Electroenceph. Clin. Neurophysiol., 1970, 29: 306-310.

Lehmann, D. EEG assessment of brain activity: spatial aspects,segmentation and imaging. Int. J. of Psychophysiol. 1984,1:267-276.

Pfurtscheller, G., Lindinger, G. and Klinesch, W. DynamischesEEG-mapping - bildgebendes Verfahren fur die Unter-suchung perzeptiver, motorischer und kognitiver Hirnleis-tung. Z. EEG-EMG, 1986,17:113-116.

Spencer, S.S. Evolving indications and applications of epilepsysurgery. Clin. Neuroscience, 1994,2(1): 3-9.

Sperling, M.R. and O'Connor, M.J. Electrographic correlates ofspontaneous seizures. Clin. Neuroscience, 1994,2(1): 17-46.

Spreng, M., Hellmann, G., Kuhn, M., Reinartz, K.-D. and Stefan,H. Bearbeitung von evozierten Potentialen und Epilepsie-EEG/MEG mit unuberwacht lernenden Klassifikatoren.Biomedizinische Technik, 1993, Band 38.

Spreng, M. Problems of different stages of vigilance in diagno-sis of central nervous system function, Proc. 2nd Int. Work-shop and Symposion on Evoked Potentials, EdizioneTecniche Amplifon, Milano, 1984.

Spreng, M., Hellmann, G., Kuhn, M., Reinartz, K.-D. and Stefan,H. Unsupervised EEG/ECoG/MEG classification in epi-lepsy using array processors. In: M. Eiselt, U. Zwiener, H.Witte (eds.): Quantitative and Topological EEG and MEGAnalysis, Universitatsverlag Jena, Jena, 1995:133-136.

Wang, J. and Wieser, H.G. Regional 'rigidity' of backgroundEEG activity in the epileptogenic zone. Epilepsia, 1994,35(3): 495-403.

Wieser, H.G., Bjelac, M., Khan, N, Muller, S., Siegel, A.M., Yonek-awa, Y.: Interplay of interictal discharges and seizures: someclinical and EEG findings. In: P. Wolf: Epileptic Seizures andSyndromes, John Libbey and Comp., London 1994:563-578.

Documents

Unsupervised Classification of EEG from Subdural Seizure Recordings