Full text of DOI:10.1017/S0317167100001402

ORIGINALARTICLE
Identification of the Temporal Components
of Seizure Onset in the Scalp EEG
Nora S. O’Neill, Manouchehr Javidan, Zoltan J. Koles
ABSTRACT: Background: The identification of the earliest indication of rhythmical oscillations and
paroxysmal events associated with an epileptic seizure is paramount in identifying the location of the
seizure onset in the scalp EEG. In this work, data-dependent filters are designed that can help reveal
obscure activity at the onset of seizures in problematic EEGs. Methods: Data-dependent filters were
designed using temporal patterns common to selected segments from pre-ictal and ictal portions of the
scalp EEG. Temporal patterns that accounted for more variance in the ictal segment than in the pre-ictal
segment of the scalp EEG were used to form the filters. Results: Application of the filters to the scalp
EEG revealed temporal components in the seizure onset in the scalp recording that were not obvious in
the unfiltered EEG. Examination of the filtered EEG enabled the onset of the seizure to be recognized
earlier in the recording. The utility of the filters was confirmed qualitatively by comparing the scalp
recording to the intracranial recording and quantitatively by calculating correlation coefficients between
the scalp and intracranial recordings before and after filtering. Conclusion: The data-dependent
approach to EEG filter design allows automatic detection of the basic frequencies present in the seizure
onset. This approach is more effective than narrow band-pass filtering for eliminating artifactual and
other interference that can obscure the onset of a seizure. Therefore, temporal-pattern filtering facilitates
the identification of seizure onsets in challenging scalp EEGs.
RÉSUMÉ: Identification des composantes temporales du début d’une crise convulsive à l’ÉEG de surface.
I n t r o d u c t Li ’oi dne n: tification, à l’ÉEG de surface, des signes les plus précoces d’oscillations rythmiques et
d’événements paroxystiques associés à une crise épileptique est très importante pour la localisation du site d’origine de
la crise. Dans cette étude, des filtres dépendants des données ont été conçus pour aider à mettre en évidence une activité
masquée au début des crises dans les ÉEG problématiques. M é t h o d De ess:filtres ont été élaborés en utilisant des motifs
temporaux communs à des segments sélectionnés de portions pré-ictales et ictales d’ÉEGs de surface. Des motifs
temporaux qui expliquaient une plus grande part de la variance dans le segment ictal que dans le segment pré-ictal de
l’ÉEG de surface ont été utilisés pour élaborer les filtres. R é s u l t aL t’ as p :p lication des filtres à l’ÉEG de surface a mis
en évidence des composantes temporales du début de la crise, qui n’étaient pas évidentes à l’enregistrement ÉEG non
filtré. L’examen de l’ÉEG filtré a permis de reconnaître plus tôt le début des crises sur l’enregistrement. L’utilité des
filtres a été confirmée qualitativement en comparant l’enregistrement de surface à l’enregistrement intracrânien et
quantitativement en calculant les coefficients de corrélation entre les enregistrements de surface et intracrâniens avec
et sans filtre.C o n c l u s iL o’ a np p: roche à l’élaboration de filtres ÉEG selon les données permet la détection automatique
des fréquences de base présentes au début des crises. Cette approche est plus efficace que le filtrage passe-bande étroit
pour éliminer une interférence due à un artefactuelle ou autre qui peut masquer le début d’une crise. Ce type de filtre
facilite l’identification du début des crises dans les enregistrements ÉEGs problématiques.
Can. J. Neurol. Sci. 2001; 28: 245-253
While the clinical symptoms and/or the scalp EEG are useful
for the diagnosis of epilepsy, it is the knowledge of the location
of the primary seizure focus inside the brain that has the greatest
value for the treatment of the disorder. This knowledge is best
acquired early in the seizure before the recruitment of susceptible
regions. Therefore, the ability to identify the seizure onset at its
earliest manifestation in the EEG is desirable. The difficulty,
however, is that, in some seizure types, the seizure activity inside
the brain begins sometime before it can be discerned in the scalp
EEG. It has been suggested that this interval can be of the order
seizure onset in the scalp EEG is obscured by artifacts, swamped
by normal background activity or is so subtle that the exact time
of seizure onset is difficult to identify. Therefore, the
interpretation of the EEG is initially an exercise in seizure onset
From the Department of Biomedical Engineering, University of Alberta, (NSO, ZJK),
Department of Neurology, University of Alberta Hospital (MJ), Edmonton, Alberta,
Canada.
RECEIVED JUNE 9, 2000. ACCEPTED INFINALFORM MAY 15, 2001.
Reprint requests to: Z.J. Koles, Department of Biomedical Engineering, 1098 Research
Transition Facility, University of Alberta, Edmonton, Alberta, Canada T6G 2V2
1
of several seconds. Furthermore, in many cases, the moment of
THE CANADIAN JOURNAL OF NEUROLOGICAL SCIENCES
245

THE CANADIAN JOURNAL OF NEUROLOGICAL SCIENCES
detection, where the objective is to identify the moment when the
first sustained temporal pattern associated with seizure activity is
recognized amidst the background activity. Identification of the
pattern of the temporal activity leading up to this moment would
also be informative. There is evidence that the temporal
characteristics of the seizure onset, including the frequency
and/or the rhythmical components present, also contain
information related to the location of the responsible source and
the propagation path of the seizure activity.1 For example, both
the hippocampal and neocortical onsets of temporal lobe seizures
seem to be characterized by rhythmical waves or patterns with
different frequency ranges. These findings suggest that there is
value in accurately identifying the onset of ictal activity in the
EEG as well as in identifying and observing the detailed time
course of the temporal patterns that are present during the seizure
onset.
METHODS
The characteristics of the TP-filter are derived from the
temporal patterning of the EEG, at a particular electrode site, in
the selected pre-ictal and ictal segments of the recording. If the
digital samples from these segments are depicted as x (n) and
a
xb(n) respectively, where n = 0 to T-1, then the presence of a
specific temporal pattern can be detected in these segments by
convolving the individual samples with a set of weighting factors
that match that pattern. This process is called matched filtering.
Mathematically, matched filtering can be represented as:
N–1
-11
y (n) = S p(m) x (n+m)
(1a)
a
a
m=0
N–1
y (n) = S p(m) x (n+m)
(1b)
b
b
m=0
so that, for example, y (0) is the weighted sum of x (0) to
In this paper, a data-dependent temporal filter is described
that is very sensitive to temporal changes in the EEG during the
seizure onset. Application of the filter usually enables the onset
of the seizure to be discerned before it becomes obvious in the
raw EEG. In addition, as part of the filter design process, the
evolution of distinct temporal components present during the
seizure onset can be observed. The design of the filter requires
that pre-ictal and ictal segments be selected manually from a
single channel in the EEG recording. To do this, the seizure is
first visually detected in the EEG and/or from the clinical
symptoms of the patient. The ictal segment is then selected early
in the seizure period. This is followed by backtracking through
the recording until a pre-ictal segment that contains what is
judged to be normal background EEG is found. The
characteristics of the filter are then derived from temporal
patterns present in the two segments. Temporal patterns that
account for more variance (signal amplitude) in the ictal segment
than in the pre-ictal segment are used to create what is called the
temporal pattern (TP) filter. The frequency response of this filter
is highly tuned to the seizure and can be considerably more
complex than the responses of band-pass filters traditionally
used to process the EEG.
a
a
x (N-1), y (1) is the weighted sum of x (1) to x (n) and so on.
a
a
a
a
Therefore, from T samples, T-N+1 values of y (n) and y (n) can
a
b
be computed using the N factors in the temporal pattern p(m).
Strictly, y (n) and y (n) are match-filtered versions of x (n) and
a
b
a
xb(n) but they can also be thought of as temporal components of
x (n) and x (n).
a
b
For the TP-filter, the temporal patterns are chosen to
emphasize the seizure onset activity and de-emphasize the
background or artifactual activity. This is accomplished by
selecting these patterns so that the variance of y (n) (derived
b
from the ictal segment) is large compared to variance of y (n)
a
(
derived from the pre-ictal segment). Here, the term variance is
used in the purely statistical sense to mean a measure of the
magnitude variations of the numbers in each of these
components. It can then be said, for example, that the pattern
p(m) accounts for more variance in the ictal segment than in the
pre-ictal segment. There are an infinite number of temporal
patterns that will result in differences of this kind as some of the
patterns will only be marginally different from one another.
Therefore, the temporal patterns are chosen so that they account
for components that are independent.
Two temporal components y (n) and y (n) are said to be
The TP-filter is similar but different from other data-
dependent filters such as the template, matched or inverse filter.
Its design is based on such signal processing methods as
1
2
independent if
T–N+1
autoregressive modeling, autocovariance and eigen analyisis.1
2-21
S y (n) y (n) = 0
(2)
1
2
n=0
Some very promising seizure detection algorithms using these
and related methods have, in fact, been reported in the
so that independent components have a covariance of 0. If
y (n) and y (n) are derived from x (n), they are called
2
2,23
1
2
a
literature.
In contrast to these, the TP-filter is not a seizure
independent temporal components of x (n). It can be shown that
a
detector per se in that the pre-ictal and ictal segments must be
manually selected beforehand. The design of the TP-filter does,
however, involve the automatic decomposition of the onset into
distinct temporal components. The time courses of these
components can be associated with either the seizure onset or
with background or artifactual activity. Since the TP-filter is
specifically tuned to subtle changes in the pre-ictal EEG, it is
probably more sensitive to the seizure onset than conventional
detectors. Therefore, the TP-filter usually enables the onset to be
detected earlier. The spatial patterns in the TP-filtered scalp EEG
at the precise moment of seizure onset can serve as input to
source localization algorithms such as multiple-signal
both x (n) and x (n) can have up to N mutually independent
a
b
components. For the TP-filter, the temporal patterns are chosen
so that they simultaneously account for independent components
in both the selected pre-ictal and ictal segments. They are
therefore called common temporal patterns. Those common
temporal patterns that, in terms of their corresponding
independent temporal components, account for more variance in
the ictal segment than in the pre-ictal segment are used to form
the TP-filter. The impulse response, h(m) of this filter is a
weighted sum of these patterns. That is:
q
2
4
h(m) = S c p (m)
(3)
classification (MUSIC) and low resolution electromagnetic
i
i
i=1
2
5
tomography. (LORETA).
2
46

LE JOURNAL CANADIEN DES SCIENCES NEUROLOGIQUES
where p (m), i = 1 to q, are the q common temporal patterns
that account for more variance in the ictal segment than in the
Figure 2 and the raw EEG containing the seizure is shown in
Figure 3. The recording is referential with respect to the left ear
electrode A1. Figure 3 suggests that the seizure onset occurs at
about 5 or 6 s into the recording and that it originates near depth
electrode a2 and subdural electrode c4. At 7 or 8 s into the
recording, there is evidence of the seizure at the scalp electrodes.
However, this would probably not have been recognized as such
in the absence of the intracranial recordings. To demonstrate the
performance of the TP-filter, four depth electrodes (a2 - a5) in
the left-temporal lobe, two subdural strip electrodes (c4 - c5)
located under the inferior side of the left anterior temporal lobe,
and six scalp electrodes, located at standard 10 - 20 sites over the
left hemisphere, were used.
i
pre-ictal segment and where the c are the associated weighting
i
factors. These weighting factors are determined by finding the
best fit, in a least squares sense, of the q independent temporal
components to the ictal segment. That is:
q
x (n) = S c y (n)
(4)
bf
i
bi
i=1
This fitting procedure leads to x (n), the TP-filtered version
bf
of x (n). In equation (4), y (n) are the q independent components
b
bi
of x (n). Using equation (1b), it can be shown that:
b
N–1
x (n) = S h(m) x (n+m)
(5)
The segment of the EEG selected from Figure 3 for analysis
is shown in Figure 4. In addition, the tracings in Figure 4 have
been band-pass filtered into the range 1-55 Hz. This figure
confirms that the seizure has a left hippocampal onset localized
to the inferomedial depth (a2) and subdural (c4) electrode sites
and suggests that the onset of the seizure actually occurs at about
5.2s.
bf
b
m=0
where the impulse response of the filter, h(m), is given by
equation (3).
Once derived from the selected pre-ictal and ictal segments,
the TP-filter is applied to all channels in the EEG recording. The
details of the method used to obtain the patterns, p (m), and the
i
weightings on these patterns, c , are given in the appendix.
i
To design the TP-filter, the pre-ictal and ictal segments were
selected as shown at the top of Figure 4. Four hundred digital
samples (2 s) from each segment at F9 were used to calculate the
temporal patterns common to the two segments. The independent
components of the tracing from F9 corresponding to the 15
common temporal patterns that account for more variance in the
ictal segment than in the pre-ictal segment are shown in Figure
5. The numbers along the right-hand side of the tracings in
Figure 5 indicate how the total variance in the pre-ictal and ictal
segments is accounted for by each of these patterns. These
numbers are the eigenvalues defined by equation (A5) in the
appendix. For example, Figure 5 shows that, of the total variance
accounted for by the common temporal pattern underlying
component 1, 97% comes from the ictal segment. Therefore,
only 3% comes from the pre-ictal segment. This division of
variance can be seen by comparing the amplitudes of the pre-
ictal (3 - 5 s) and ictal segments (7 - 9 s) in component 1. As the
eigenvalues decrease from 97% to 50%, relatively less and less
of the total variance accounted for by the underlying temporal
pattern comes from the ictal segment.
RESULTS
The tracings in Figure 1, all of which contain a seizure,
demonstrate the performance of the TP-filter. Tracings 1 to 3 are
the raw EEGs while tracings 4 to 6 are their TP-filtered
counterparts. Tracing 1 is a simulated EEG while tracings 2 and
3
are single-channel excerpts from actual clinical EEGs.
The simulated EEG of tracing 1 was constructed from several
distinct sinusoidal components. As background, 0.30, 0.35, 1.0,
and 7 Hz components were superposed and this composite signal
is present throughout the tracing. Frequencies of 4 Hz and 2.5 Hz
were used to simulate the seizure with the former component
starting at about 5.5 s and the latter at about 7.5 s. The TP-filtered
version of this tracing shows that the background activity has
been eliminated, the moment of seizure onset has been sharpened
and that the two frequencies in the seizure are more obvious.
With regard to the clinical EEGs in Figure 1, we note that the
onset of the seizure in tracing 2 occurs at about 5 s while, in
tracing 3, the moment of onset is more obscure. The TP-filtered
versions of these tracings contain far less muscle artifact than the
raw versions. After filtering, it is clear that in the pre-ictal
interval, there are frequency components present that are also
present in the ictal interval. This is particularly obvious in the
filtered version of tracing 3. In this tracing, the seizure activity
appears to grow out of the pre-ictal segment. Using the criterion
that the seizure onset occurs at the moment when persistent ictal
activity begins, the onset, in this filtered EEG, occurs at about
An examination of the components in Figure 5 also indicates
that, in terms of their waveforms, several pairs are very similar.
Each of the pairs contains distinct frequencies with distinctly
modulated amplitudes. For example, components 1 and 2 are
both very similar in terms of their frequencies (14-15 Hz) and
amplitude modulations. A closer examination, however, reveals
that the two components are in fact shifted in phase from one
o
another by 90 . This is true for components 3 and 4 where the
frequency is also about 14-15 Hz, for components 5 and 6 where
the frequency is about 7-8 Hz, for components 7 and 8 where the
frequency is about 18 Hz and so on. Of all the components in
Figure 4, only 9, 10 and 13 appear to stand alone.
Given the distinct frequencies evident in the independent
components of Figure 5, the selected pre-ictal and ictal segments
from F9 and from c4 and a2 were subjected to Fourier analysis.
The results of this analysis are shown in Figure 6. The frequency
spectra show that, in the ictal segment, both intracranial tracings
contain strong components at 7.2 Hz, 14.5 Hz, and 21.5 Hz.
Because the second and third of these frequencies are multiples
3
.5 s. This is a much earlier point in time than that suggested by
the raw tracing.
Figures 2 to 9 illustrate, in some detail, the design and the
performance of the TP-filte r. The results were obtained from a 64
channel EEG consisting of simultaneous scalp, depth and
subdural recordings from a patient with refractory complex
partial seizures. This EEG was selected because the seizure onset
was not obvious in the scalp tracings until some time after it was
clear from the intracranial tracings that the onset had occurred.
The locations of the depth and subdural electrodes are shown in
Volume 28, No. 3 – August 2001
247

THE CANADIAN JOURNAL OF NEUROLOGICAL SCIENCES
Figure 1: Tracings 1 to 3 are the raw EEGs while tracings 4 to 6 are their TP-filtered
counterparts. Tracing 1 is a simulated EEG while tracings 2 and 3 are single-channel
excerpts from actual clinical EEGs. All three tracings contain a seizure.
Figure 2: The depth and subdural electrode
locations with depth-electrode a2 and
subdural-electrode c4 shown. Electrode
locations were obtained from MR images of
the patient.
Figure 3: The raw 64 channel EEG
containing a seizure. All tracings were
obtained with respect to a common electrode
at A1. The top panel shows the tracings from
the depth electrodes, the middle panel from
the subdural strip electrodes and the bottom
panel from the scalp electrodes.
2
48

LE JOURNAL CANADIEN DES SCIENCES NEUROLOGIQUES
Figure 4: The EEG segment used to demonstrate the temporal-pattern
filters. Tracings labeled a2 to a5 are from the depth electrodes, c4 and
c5 are from the subdural strip electrodes and F9, F7, F3, T3, C3 and P3
are the scalp electrodes. Electrode F9 is used to design the TP-filters.
The pre-ictal and ictal segments used for the design of the filters are
shown along the top of the graph.
Figure 5: The temporal waveforms resulting from the decomposition of
the recording from electrode F9 in Figure 4 using the temporal patterns
common to the pre-ictal and ictal segments. Each number along the
right-hand vertical axis of the graph is the percentage of the total
variance in the pre-ictal and ictal segments, accounted for by one of the
common temporal patterns, that comes from the ictal segment.
Figure 6: The frequency spectra of the pre-ictal and ictal segments of the depth (a2), subdural (c4) and scalp
recordings (F9) respectively.
frequency components at 7.2 Hz and 14.5 Hz are in fact present
at F9 and that changes in the amplitudes of these components are
related to the onset of the seizure.
Figure 7 shows the EEG of Figure 4 after the application of
the TP-filter designed from the common temporal patterns
underlying independent components 1 to 4 in Figure 5. As
expected, all of the tracings in the filtered EEG contain a strong
of the first, these components could be harmonically related.
Peaks at these frequencies are not obvious in the spectrum at F9
during the ictal segment nor are they obvious in any of the
spectra in the pre-ictal segment. However, a closer examination
of the spectrum from F9 in the ictal segment shows that there is
in fact a peak at 7.2 Hz and that this peak is actually the second
highest in the spectrum. It would appear, therefore, that the
seizure onset at F9 is swamped by background activity. The
independent components in Figure 5 clearly indicate that
1
4-15 Hz frequency component. Figure 8 shows the EEG of
Figure 4 after the application of the TP-filter designed from the
Volume 28, No. 3 – August 2001
249

THE CANADIAN JOURNAL OF NEUROLOGICAL SCIENCES
common temporal patterns underlying independent components
samples, the upper p < 0.01 point for significant correlation is
about 0.2, absolute values of the numbers in the Table between
0.2 and 0.3 are interpreted as indicating weak correlation and
values above 0.3 as indicating stronger correlation. Very strong
correlation probably does not exist between intracranial and
scalp recordings because the field of view of a scalp electrode
contains many more sources than the field of view of an
intracranial electrode.
5
and 6 in Figure 5. All of the tracings now contain a strong 7-8
Hz frequency component. We note that the seizure onset is
clearer in the scalp tracings of Figure 7 than it is in the scalp
tracings of Figure 8 and that it is most evident at electrode F9,
the scalp electrode site closest to the depth and subdural
electrode sites.
The frequency responses of the TP-filters used to obtain
Figures 7 and 8 from Figure 4 are shown in Figures 9A and B
r e s p e c t i. vT eh el s ye were obtained by Fourier transformation of
the respective impulse responses of the filters. In Figure 9A, the
major peaks in the frequency response are at 12.9 and 15.9 Hz.
Given these frequencies, a more careful examination of the
tracings in Figure 7 indicates that, in fact, frequencies between
about 13 and 16Hz are indeed present at various times. This filter
is therefore labeled the TP-14 filter. In Figure 9B, the three larg e s t
peaks in the frequency response are at about 7.7, 15.8 and 19.4
Hz. The 7.7 Hz peak is, however, more than twice as high as the
An examination of the numbers in the Table indicates that, in
the unfiltered recordings, there is no significant correlation
between F9 on the scalp and the intracranial sites in either of the
pre-ictal or the ictal segments. There is, however, strong
correlation between the intracranial sites in the ictal segment.
Band-pass filtering into the 5-25 Hz band produces no
significant correlation between recordings from F9 and the
intracranial sites although there is now significant correlation
between the intracranial sites in the pre-ictal segment and an
increase in the strength of the correlation between these sites in
the ictal segment. In contrast, TP-filtering results in strong
correlation between all the sites in the ictal segment. TP-8
filtering also results in significant correlation between F9 and c4
in the pre-ictal segment although the correlation between these
sites in the ictal segment is strongest after TP-14 filtering. The
strongest correlation in the ictal segment involving F9 occurs
with a2 after TP-8 filtering. Also, after TP-8 filtering, the almost
complete correlation between the two intracranial sites is
notable. The strength of this correlation is also evident from a
visual comparison of the respective tracings in Figure 8.
A more careful examination of the numbers in the Table
yields a number of interesting results related to the polarities of
the correlations. Firstly, the polarities of the correlations between
F9 and intracranial sites in the ictal segment are reversed after
TP-8 and TP-14 filtering. Secondly, the polarity of the significant
correlation between F9 and c4 seen in the pre-ictal segment is
reversed in the ictal segment. Thirdly, when these filters are
combined to form the TP-8-14 filter, the polarities of the
correlations remain reversed but the strength of the correlations
is reduced to the point where it is now not significant between F9
and c4.
15.8 Hz peak. This is probably why all of the tracings in Figure 8
appear to have only a 7-8 Hz component. This filter is therefore
labeled the TP-8 filter. The composite frequency response of the
TP-8 and TP-14 filters, obtained from the temporal patterns
underlying independent components 1 to 6 in Figure 5, is shown
in Figure 9C and is labeled as the TP-8-14 filter.
To determine whether filters with frequency responses as
elaborate as those shown in Figures 9A and B are required to
detect the onset of the seizure in the scalp EEG, the tracings in
Figure 4 were band-pass (BP) filtered in the frequency range 5 -
2
5 Hz. The result of doing this changed the tracings in Figure 4
only marginally and did not noticeably enhance the onset of the
seizure.
Figures 7 and 8 show that the same frequency components are
present in all of the depth, subdural and scalp tracings. However,
the covariance between these sites, in terms of amplitude
modulation, appears to be low. To examine this covariance
quantitatively, correlation coefficients were calculated between
the scalp recording at F9 and the intracranial recordings at c4 and
a2 in both the pre-ictal and ictal segments. The results of these
calculations are given in the Table. Given that, with n = 400
Table: The correlation coefficients between the raw and filtered recordings from scalp electrode F9 and the subdural-electrode c4 and
the depth-electrode a2 in the pre-ictal (2.4 – 4.4 s) and ictal (6.6 - 8.6 s) segments of the EEG.
Pre-ictal
Ictal
F9-c4
-0.031
0.004
0.045
0.210
0.136
F9-a2
0.152
0.061
-0.041
0.170
0.096
c4-a2
F9-c4
-0.074
-0.089
0.424
F9-a2
0.144
0.187
-0.363
0.572
0.337
c4-a2
Raw EEG
-0.019
-0.236
-0.360
-0.194
-0.258
-0.577
-0.786
-0.790
-0.932
-0.899
BPFilter (5-25 Hz)
TP-14 Filter
TP-8 Filter
-0.386
-0.175
TP-8-14 Filter
The entries in the cells illustrate the effects of the different filter characteristics. BP indicates Band Pass, TP-14 the nominally 14-15 Hz Temporal
Pattern, TP-8 indicates the nominally 7-8Hz Temporal Pattern and TP-8-14 indicates the combined 7-8 and 14-15Hz Temporal Patterns. Based on 400
samples from the respective recordings for the calculations, the upper 0.01 significance point for the correlation coefficients is assumed to be about 0.2.
2
50

LE JOURNAL CANADIEN DES SCIENCES NEUROLOGIQUES
Figure 8: The raw EEG filtered with TP-8. The frequency response of
this filter is shown in Figure 9B.
Figure 7: The raw EEG filtered with TP-14. The frequency response of
this filter is shown in Figure 9A.
Figure 9: A. The frequency response of the filter designated as TP-14. The major peaks in this response are at 12.9 and 15.9 Hz.
B. The frequency response of the filter designated as TP-8. The major peaks in this response are at 7.7, 15.8 and 19.4 Hz.
C. The frequency response of the filter designated as TP-8-14.The major peaks in this response are at 7.7, 12.9, 15.8 and 19.4 Hz.
DISCUSSION AND CONCLUSIONS
first phase of the seizure but is clearly manifested in depth and
subdural EEGs. These results show that judiciously designed TP-
filters can automatically elicit the dominant frequencies of the
seizure onset in the scalp EEG. The dominant onset frequency of
7-8 Hz identified is consistent with the results presented by
The results of this work suggest that data-dependent TP-
filtering can be a useful approach to the identification and
analysis of the components of seizure onset in the EEG. The
results also illustrate that TP-filtering is less effective when the
pre-ictal and ictal segments contain the same frequency
6
Ebersole and Pacia who reported the frequency range of 5-9 Hz
2
6
for hippocampal seizures and with those of Franaszczuk et al
2
1
components. O’Neill has, however, shown that the detection of
the seizure onset can be improved, even when this occurs, by
following TP-filtering with spatial pattern (SP) filtering.
who reported the frequency range of 5.3 – 8.4 Hz. Other studies
report dominant frequencies up to 32 Hz but with a mean
frequency of about 16 Hz.2 The simplest explanation for the
differences between these findings is that the 7-8 Hz component
is the fundamental frequency of a hippocampal seizure recorded
intracranially while the 14-15 Hz component arises as the first
-3
The EEG used to illustrate in detail the design and
performance of the TP-filter is typical of a hippocampal seizure
in which ictal activity is not obvious in the raw scalp EEG in the
Volume 28, No. 3 – August 2001
251

THE CANADIAN JOURNAL OF NEUROLOGICAL SCIENCES
harmonic of the seizure waveform. The higher mean frequency
of intracranial seizure reported by Javidan et al2 may be due to
the inclusion of even more of the higher harmonics. That these
higher harmonics exist in intracranial recordings is clearly
evident from the spectra in Figure 6. As well, it is conceivable
that in some source and/or recording configurations, one or the
other of the fundamental or harmonic frequencies may not be
present.
the onset of many more clinical and computer-simulated
seizures. This experience leads us to suggest that TP-filtering
can, in fact, facilitate the identification of seizure activity in even
the most problematic EEGs. Substantiation of this statement will
require a quantitative analysis of these data. Nevertheless, there
is little doubt that TP-filtering acts to consistently reduce
recording artifact and background activity and to automatically
elicit the specific temporal components present during the
seizure onset. In addition, the filter design process can be
extremely valuable for detecting the precise moment of seizure
onset and for identifying the frequency components involved. An
interesting continuation of this work will undoubtedly be a study
of the utility of TP-filtering in combination with SP-filtering and
source localization.
In conclusion, it would seem that TP-filtering has the
potential of being a useful tool for the diagnosis of epilepsy and,
more importantly, for the planning of resective surgery. The
inclusion of TP-filtering in a review workstation could reduce
the necessity for invasive EEG recording thereby lessening both
the risk and the cost of providing this procedure to the patient.
Early detection of the seizure is critical in locating the
responsible source inside the brain. A developed seizure can
involve large volumes of brain tissue, in effect masking the focus
responsible for the onset. Since most quantitative methods of
source localization assume that the sources are focal and not
distributed, accurate localization of the seizure focus would
almost certainly be facilitated by the early detection of the onset
and the elimination of background sources not related to the
seizure.
-3
The frequency spectra shown in Figure 6 suggest that, in the
ictal segment of the sample EEG, the 7-8 Hz and 14-15 Hz
components in the intracranial recordings are harmonically
related. Therefore, it is interesting to note that in Figures 7 and 8,
in the scalp tracings, the change in the 14-15 Hz component is
more pronounced than the 7-8 Hz component during the seizure
onset. Due to the lack of higher frequency components in the
pre-ictal EEG, the 14-15 Hz component appears to be a more
sensitive indicator of the seizure onset than the latter. Also, a
more careful examination of the independent components in
Figure 5 suggests that sustained 7-8 Hz activity might actually
begin at about 4.75 s in the recording and that this onset actually
leads the onset of the 14-15 Hz activity at 6.75s. These
observations would seem to indicate that, on the scalp, the two
frequency components are, in fact, not harmonically related.
The correlation analysis between the intracranial and scalp
tracings indicates that, during the ictal segment, TP-filtering
results in considerably stronger correlation between the tracings
than does BP filtering. Furthermore, there is stronger correlation
when the two TP-filters are applied separately instead of after
having been combined into one filter. Since the BPfilter is not as
highly tuned to the seizure as the TP-filters, the weaker
correlation is easy to explain in terms of a greater proportion of
uncorrelated activity in the EEG. However, why the correlation
should drop after combining two TP-filters is not so easy to
explain. However, this result supports our previous speculation
that the two components are, in fact, not harmonically related on
the scalp. In any case, we take this finding to conclude that, in
the design of TP-filters, combining independent components that
do not appear to be similar should probably be avoided.
APPENDIX
The characteristics of the TP-filter, specifically the common
temporal patterns, p (m), and c , the weightings on these factors,
i
i
are derived from differences in the temporal patterning of the
EEG at a particular electrode site in the selected pre-ictal and
ictal segments of the recording. If the sets of digital samples
from these segments are depicted as x (n) and x (n) respectively,
a
b
where n = 0 to T-1, then the temporal patterning of the segments
can be quantitatively characterized by their autocovariance
The frequency responses of the TP-filters designed in this
work are such that they are optimal for simultaneously
emphasizing components in the ictal segment of an EEG and de-
emphasizing the components in the pre-ictal segment. The
critical element in the design is the selection of the segments that
best characterize the seizure and background segments of the
recording. The idea that the pre-ictal and ictal segments for the
filter design are chosen by a human expert might be seen as a
major drawback of the method. However, as reported, by Qu and
functions. That is,
T–N–1
R (m) = S x (n) x (n+m)
(A1)
a
a
a
n=0
and similarly for R (m). The values of R (m) (or R (m)) are
b
a
b
computed for m = 0 to N-1, and tend from a maximum at m=0,
toward zero, sometimes in an oscillatory fashion, as m becomes
large. At m=0, R (0) (or R (0)), is simply a measure of the
a
b
2
8
Gotman, seizure onset patterns are highly variable between
patients, the seizures may evolve from relatively subtle changes
in background activity and a particular onset pattern in one
patient may not lead to a seizure in another patient. It is this
variability between seizures and seizure onsets that has slowed
the development of a truly reliable automated seizure detector.
Therefore, until the process by which seizures occur and develop
is better understood, the importance of the expert epileptologist
in the design of the TP-filter should not be underestimated.
To demonstrate the performance of the TP-filter in this paper,
we have used only four seizures. We have, however, applied the
filter to the identification of the temporal components present in
variance of the individual samples in the segment while for other
values of m, it is a measure of the covariance between samples
separated by m intervals. The TP-filter is designed from
differences between the autocovariance functions of the selected
pre-ictal and ictal EEG segments.
To show how this is done, let R and R be the temporal
a
b
autocovariance matrices that correspond to the autocovariance
functions of x (n) and x (n). Each of these matrices consists of
a
b
the respective autocovariance function as its first row. Adjacent
rows are formed by successively shifting the previous row
circularly to the right. In this way, N-1 additional unique rows
can be formed. The autocovariance matrices are, therefore, of
2
52

LE JOURNAL CANADIEN DES SCIENCES NEUROLOGIQUES
dimension N x N ( N T ) where T is the number of digital
samples in each segment.
A composite covariance matrix for the two segments is
formed as:
2. Javidan M, Katz A, Tran T, et al. Frequency characteristics of
neocortical and hippocampal onset seizures. Epilepsia 1992a; 33:
S3-58.
3
.
Javidan M, Katz A, Pacia S, et al. Onset and propagation
frequencies in temporal lobe seizures. Epilepsia 1992b;33:S3-59.
R = R + R
(A2)
4. Spencer SS, Guimaraes P, Katz A, Kim J, Spencer D. Morphological
patterns of seizures recorded intracranially. Epilepsia 1992; 33:
537-545.
5. Ebersole JS, Pacia SV. Temporal neocortical epilepsy syndrome:
scalp EEG identification. Epilepsia 1993; 34: S7-112.
6. Ebersole JS, Pacia SV. Localization of temporal lobe foci by ictal
EEG patterns. Epilepsia 1996; 37: 386-399.
c
a
b
From Bc and
c, the N x N matrices containing the
eigenvectors and eigenvalues of R , a whitening transformation
c
is constructed as:
–
1
⁄
2
W = Bc
c
(A3)
where the column dimension of W is the number of
significant (nonzero) eigenvalues in c. This transformation is
applied to R and R to form two whitened covariance matrices
7. Pacia S, Ebersole JS. The classification of temporal lobe seizures by
scalp EEG. Epilepsia 1992; 33: S3-58.
8
.
Pacia S, Ebersole JS. Temporal neocortical epilepsy syndrome:
intracranial EEG identification. Epilepsia 1993; 34: S6-26.
a
b
S and S :
a
b
9. Spanedda F, Cendes F, Gotman J. Relations between EEG seizure
morphology, interhemispheric spread, and mesial temporal
atrophy in bitemporal epilepsy. Epilepsia 1997; 38: 1300-1314.
S = WЈR W and S = WЈR W
(A4)
a
a
b
b
where Ј is the transpose operator. It can be shown28 that these
1
0. Assaf BA, Ebersole JS. Visual and quantitative ictal EEG predictors
of outcome after temporal lobectomy. Epilepsia 1999; 40: 52-61.
matrices will have common eigenvectors B and eigenvalues
t
a
and
respectively such that:
11. Mizuno-Matsumoto Y, Okazaki K, Kato A, et al. Visualization of
epileptogenic phenomena using cross-correlation analysis:
localization of epileptic foci and propagation of epileptiform
discharges. IEEE Trans Biomed Eng 1999; BME-46: 271-279.
b
+
= I
b
(A5)
a
where ⌱ is the identity matrix. Because of this constraint,
those eigenvectors that account for the largest variance in S will
account for smallest variance in S and vice versa.
1
2. Lopes Da Silva FH, Van Hulten K, Lommen JG, et al. Automatic
detection and localization of epileptic foci. Electroencephalogr
Clin Neurophysiol 1977; 43: 1-13.
a
b
From equations (A2) to (A5), it can be shown that:
13. Pfurtscheller G, Fischer G. A new approach to spike detection using
a combination of inverse and matched filter techniques.
Electroencephalogr Clin Neurophysiol 1978; 44: 243-247.
14. Barlow JS. EEG transient detection by matched inverse digital
filtering. Electroencephalogr Clin Neurophysiol 1980;48:246-248.
15. Barlow JS. Analysis of EEG changes with carotid clamping by
selective analog filtering, matched inverse digital filtering and
PR PЈ =
and PR PЈ =
(A6)
(A7)
a
a
1
BcЈ
b
b
–
⁄
2
where
P = BtЈ
c
is the matrix of temporal patterns common to x (n) and x (n).
The rows of P that correspond to the rows in B account for
a
b
t
maximally different proportions of the variances in the two
automatic adaptive segmentation: a comparative study.
Electroencephalogr Clin Neurophysiol 1984; 58: 193-204.
6. Hjorth B, Rodin E. Extraction of deep components from scalp EEG.
Brain Topogr 1988a; 1: 65-69.
7. Hjorth B, Rodin E. An eigenfunction approach to the inverse
problem of EEG. Brain Topogr 1988b; 1: 79-86.
segments x (n) and x (n). Each of the rows of P is a temporal
a
b
1
1
1
pattern p (m), m = 0 to N-1, that simultaneously decomposes the
i
two segments into independent components. The elements of the
diagonal matrices
and
are related to the variances of
b
a
8. Koles ZJ. The quantitative extraction and topographic mapping of
these components in x (n) and x (n) respectively. Only those
b
a
the abnormal
components in the clinical EEG.
rows of P for which
>
are used for the TP-filter. The
b
a
Electroencephalogr Clin Neurophysiol 1991; 79: 440-447.
9. Liu A, Hahn JS, Heldt GP, Coen RW. Detection of neonatal seizures
through computerized EEG analysis. Electroencephalogr Clin
Neurophysiol 1992; 82: 30-37.
0. O’Neill NS, Javidan M, Koles ZJ. Localization of seizure onset in
the EEG. Proceedings of the 24th meeting of the Canadian
Medical and Biological Engineering Society 1998; 14-15.
21. O’Neill NS. Temporal and spatial pattern filtering of the EEG. M.Sc
Thesis, University of Alberta, Edmonton, Canada 1998.
22. Gotman J. Automatic detection of seizures and spikes. J Clin
Neurophysiol 1999; 16: 130-140.
3. Osorio I, Frei MG, Wilkinson SB. Real-time automated detection
and quantitative analysis of seizures and short-term prediction of
clinical onset. Epilepsia 1998; 39: 615-627.
impulse response of the TP-filter is, as indicated by equation (3),
the weighted sum of these patterns. In matrix terms, this sum is
given by:
1
2
h = [C P ]
(A8)
q
q 1
where C is called the matrix of temporal factors common to
x (n) and x (n) and is given by:
a
b
–
1
⁄
2
C = Bc
c
Bt
(A9)
In equation (A8), the subscript q indicates that only the q r o w s
from P and the corresponding q columns from C for which
are used. The subscript 1 on the square bracket indicates that the
2
>
b
a
first row of the N x N matrix inside is required to form the impulse
response of the filter. The other rows of this matrix are redundant
since they are just circularly rotated versions of the first row.
24. Mosher JC, Lewis PS, Leachy RM. Multiple dipole modelling and
localization from spatio-temporal MEG data. IEEE Trans
Biomed Eng 1992; BME-39: 541-557.
2
5. Pascual-Marqui RD, Michel CM, Lehmann D. Low resolution
electromagnetic tomography: a new method for localizing
electrical activity in the brain. Int J Psychophysiol 1994; 18: 49-65.
6. Franaszczuk PJ, Bergey GK, Durka PJ, Eisenberg HM. Time-
frequency analysis using the matching pursuit algorithm applied
to seizures originating from the mesial temporal lobe.
Electroencephalogr Clin Neurophysiol 1998; 106: 513-521.
7. Qu H, Gotman J. A patient specific algorithm for the detection of
seizure onset in long-term EEG monitoring: possible use as a
warning device. IEEE Trans Biomed Eng 1997;BME-44:11 5 - 1 2 2 .
8. Fukunaga K. Statistical pattern classification. In: Young T,. Fu KS.
eds. Handbook of Pattern Recognition and Image Processing.
Academic Press, 1986.
ACKNOWLEDGEMENTS
2
We acknowledge the Natural Sciences and Engineering Research
Council of Canada (NSERC rgpin 121524) and the Alberta Health
Sciences Research Institute for their financial support and Dr. S. Spencer
of the Yale School of Medicine for the data provided.
2
2
REFERENCES
1
.
Pacia S, Ebersole JS. Intracranial EEG substrates of scalp ictal
patterns from temporal lobe foci. Epilepsia 1997; 38: 642-654.
Volume 28, No. 3 – August 2001
253