Clinical Neurophysiology 114 (2003) 898–908

www.elsevier.com/locate/clinph

Wavelet based automatic seizure detection in intracerebral

electroencephalogram
Y.U. Khan, J. Gotman*
Montreal Neurological Institute and Department of Neurology and Neurosurgery, McGill University, Room 767, 3801 University Street,
Montreal, Quebec, Canada H3A 2B4

Accepted 30 January 2003

Abstract
Background: Automatic seizure detection is often used during long-term monitoring, and is particularly important during intracerebral
investigations. Existing methods make many false detections, particularly in intracerebral electroencephalogram (EEG) because of frequent
large amplitude rhythmic activity bursts that are non-epileptiform.
Objective: To develop a seizure detection method for intracerebral monitoring that is as sensitive as existing methods but has fewer false
detections.
Methods: To capture the rhythmic nature of seizure discharges, we developed a wavelet-based method, examining how different
frequency ranges fluctuate compared to the background. In particular, the system remembers rhythmic bursts occurring commonly in the
background to avoid detecting them as seizures.
Results: The method was evaluated on test data from 11 patients, including 229 h and 66 seizures, and its performance compared to the
method of Gotman (Electroencephalogr clin Neurophysiol 76 (1990) 317). Detection sensitivity was unchanged at close to 90%, but false
detections were reduced from 2.4 to 0.3/h.
Conclusions: Perfect sensitivity is unlikely because the morphology of seizure discharges is so variable. Nevertheless, the 87% sensitivity
obtained in the combined training and testing data is quite high. We reduced the average false alarm rate to one per 3 h of recording, or 6 per
24-h period. Given how rapidly one can decide visually that a detection is erroneous, false detections should not cause any burden to the
reviewer.
Significance: In intracerebral EEG it is possible to detect seizures automatically with high sensitivity and high specificity.
q 2003 International Federation of Clinical Neurophysiology. Published by Elsevier Science Ireland Ltd. All rights reserved.
Keywords: Epilepsy; Automatic seizure detection; Electroencephalogram; Discrete wavelet transform

1. Introduction particularly critical when patients have intracerebral

Automatic seizure detection is an important aspect of
long-term epilepsy monitoring. It is well known that many
patients have seizures for which they have no warning, for
which they are not able to press an alarm button when they
feel a warning, or for which they have no awareness at all
(Blum et al., 1996). It is very tedious, expensive and
impractical to have a person constantly observe every
patient and every electroencephalogram (EEG) being
monitored. It is also tedious and difficult to review a 24-h
continuous EEG recording, the problem becoming worse as
the number of channels increases. The situation is
* Corresponding author. Tel.: þ1-514-398-1953; fax: þ 1-514-398-8106.
E-mail address: jean.gotman@mcgill.ca (J. Gotman).
1388-2457/03/$30.00 q 2003 International Federation of Clinical Neurophysiology. Published by Elsevier Science Ireland Ltd. All rights reserved.
doi:10.1016/S1388-2457(03)00035-X
electrodes, when their EEG often has 100 or more channels.
Seizures can be limited to a very small number of channels,
sometimes to a single electrode, and it is therefore necessary
to observe or review all channels. Automatic detection,
particularly if it is performed on-line, can be of great
assistance in identifying sections of EEG likely to have
seizures. Even if false alarms are relatively frequent
compared to true detections, automatic detection can be
very useful because all detections are always reviewed by
qualified personnel; false detections can be dismissed
relatively rapidly.
Apart from early attempts at seizure detection from
simple calculations of amplitude, the first widely applicable
method of seizure detection was that of Gotman (1982),
later modified and subject to an extensive evaluation
CLINPH 2002114
 Y.U. Khan, J. Gotman / Clinical Neurophysiology 114 (2003) 898–908 899

(Gotman, 1990). This method has been in relatively
widespread clinical use, as it has been integrated to several
commercial devices. It was subjected to independent
evaluation on large patient groups in diverse clinical
environments (Pauri et al., 1992; Salinsky, 1997). Taken
altogether, these evaluations indicated that the method
detected 70 –80% of seizures, that 20 –40% of seizures were
not noticed by the nursing staff nor the patient and were
detected by the computer only, and that the false detection
rate was of one to 3 false detections per hour of monitoring.
Other methods have used a different approach based on
artificial neural networks. With this approach, the algorithm
is trained by being presented with many examples of
seizures and many examples of ‘non-seizure’ EEG sections;
the artificial neural network learns to discriminate between
these patterns. It is not necessary to define formally what a
seizure consists of, but it is necessary to define EEG features
that are likely to represent seizure activity. Webber et al.
(1996) use this approach but have not validated their
method. Gabor et al. (1996) also reported such a method and
later performed an extensive evaluation in 65 patients and
4500 h of scalp recording (Gabor, 1998). They found that
92.8% of seizures were detected and there was an average
1.35 false detections per hour. It may be noted that seizures
were defined as events for which there were behavioral
manifestations noted by the patient or observer, or visible on
fast video review.
Apart from the evaluation of Gotman (1990) that
included 5 patients with intracerebral electrodes, all above
evaluations were performed in scalp recordings. Intracer-
ebral recordings have the clear advantage of presenting
fewer artefacts that scalp recordings (although technical
problems from malfunctioning electrodes are not rare).
They also have the disadvantage of presenting a larger
variety of EEG patterns, with larger fluctuations in
amplitude and frequency than scalp EEGs. Harding (1993)
developed a method specifically aimed at detecting seizures
in intracerebral EEG (often called stereoelectroencephalo-
grams or SEEG), but it is difficult to evaluate its
performance because detection criteria were changed for
each patient after a first seizure was recorded. Osorio et al.
(1998) also presented a method specifically aimed at SEEG

Fig. 1. A section of EEG with seizure activity. The signal ‘s’ is shown decomposed into 5 levels of detail coefficients: d1, d2, d3, d4, d5 and the residue signal, a5
(approximation coefficients). [s ¼ a5 þ d5 þ d4 þ d3 þ d2 þ d1]. The x axis is labeled in samples (sampling is at 200 Hz). The seizure discharge is mostly
visible in levels 3 and 4 of the wavelet transform.
900 Y.U. Khan, J. Gotman / Clinical Neurophysiology 114 (2003) 898–908
recordings. They evaluated the method on 125 seizure
segments and over 200 non-seizure segments, but the total
amount of data analyzed only totaled about 54 h; recordings
had 16 or 8 channels. They obtained, extraordinarily, 100%
sensitivity and had no false detections. It is important to
note, however, that the algorithm was specifically tuned to
the data set being investigated (no independent test set was
evaluated).
We present here a new method of seizure detection
primarily aimed at SEEG. The method is based on wavelet
analysis and makes use of a wide temporal context. It is
evaluated on a set of unselected data totally independent
from the set on which it was developed. Its performance is
also compared to that of the method of Gotman.
2. Methods
The wavelet transform has developed into an important
tool in signal analysis and feature extraction (Kalayci and
Ozdamar, 1995; Williams et al., 1995). It has the ability of
providing a representation of the signal in both the time and
frequency domains. In contrast to the Fourier transform,
which provides the description of the overall regularity of
signals, the wavelet transform identifies the temporal
evolution of various frequencies. This property suits the
EEG signal, which is not stationary by its nature, and has a
time varying frequency content and contains paroxysmal
events. Many books and numerous papers have been
published describing wavelets (Daubechies, 1992; Burrus
et al., 1998). It is important to select a wavelet that matches
the shape or frequency characteristics of seizures. The use of
Daubechies wavelet in the analysis of non-stationary signals
such as EEG is well documented (Kalayci and Ozdamar,
1995; Gabor et al., 1996; Park et al., 1997; Gabor, 1998).
We have selected Daubechies-4 wavelet after testing other
conventional wavelets and finding only minor differences in
detection performance. The discrete wavelet coefficients
(DWT) for different scales are obtained using a simple
recursive digital filter scheme. The output of the high pass
filter gives the set of DWT coefficients (also called detail
coefficients) for that scale and captures the high frequency
energy in the signal. The output of the low pass filter
encompasses the lower frequency energy of the signal and is
associated with the set of companion functions called
scaling functions (or approximation coefficients). The
scaling function is therefore a residual function that is
representative of all the lower frequency information in the
signal after the detail function is computed.
A signal sampled at 200 Hz is band limited to 100 Hz
(Nyquist criteria). Therefore, the recursive filtering gives
the detail coefficients representing 50 –100 Hz (scale 1),
25 –50 Hz (scale 2), 12 –25 Hz (scale 3), 6 –12 Hz (scale 4),
3 –6 Hz (scale 5) and approximation coefficients represent
the rest (lower frequency information). Seizures in depth
recorded EEGs commonly occur between 3 and 25 Hz.
Therefore, this frequency range will be represented in
wavelet scales 3, 4 and 5. Fig. 1 shows a small section of
EEG (‘s’) with seizure activity that is decomposed into 5
detail functions (‘d’) and the approximation function (‘a’).
For this EEG the coefficients have maximum values in
scales 3 and 4. Wavelets can therefore be viewed as a set of
functions that can be used to represent natural, highly
transient phenomena such as spikes and seizures in the EEG.
In the following section, we describe how the coefficients
from the scales of interest are used to evaluate the features.
2.1. Features
2.1.1. Energy
In the interictal period, the EEG is most often spread
across most of the scales (i.e. includes a broad range of
frequencies). In contrast, when the seizure occurs, the EEG
shows a rhythmic behavior, i.e. a repetition of same
waveform over a period of time, thus exhibiting most of
its energy in limited scales of the multi-resolution frame-
work. For the Daubechies wavelet, which forms an
orthogonal basis in the time-frequency plane, the sum of
square of coefficients of the wavelet series is the energy of
the signal. Energy for the DWT for level (or scale) ‘l’ is
given as:
X
n
eðlÞ ¼ D2i p Dt=n
i¼1
where n is the number of DWT coefficients (D) present in
the level ‘l’ and ‘Dt’ is the sampling interval. The relative
energy (er(l)) for the level ‘l’ is evaluated as:
X
n
er ðlÞ ¼ eðlÞ= eðiÞ
i¼1
Therefore, for the part of a recording where a seizure occurs,
the relative energy of the DWT coefficients for a particular
level (scale) rises to a high value and can be detected by
setting an appropriate threshold (ETH).
2.1.2. Coefficient of variation
When using EEG amplitude analysis, features often used
are the mean (m), the standard deviation (s) and the
coefficient of variation (s2/m2). These are computed from
the original signal or the rectified signal. In the present
work, the coefficient of variation (c) is computed on the
DWT coefficients for the selected level ‘l’. For this, the
waveform decomposition method based on Gotman’s
approach (Gotman, 1982; Gotman and Gloor, 1976) in
which the waveform is broken into segments is utilized. A
segment is defined as a section of the waveform between a
minimum and the following maximum or vice-versa. Each
segment is characterized by its duration, amplitude and
direction. The mean and the standard deviation of the
amplitudes of the segments is computed, and from these we
compute the coefficient of variation (c(l)) of the amplitudes
 Y.U. Khan, J. Gotman / Clinical Neurophysiology 114 (2003) 898–908
Table 1
Ten possible detection levels for 5-s block
of the segments. The coefficient of variation is likely to give
smaller values for signals exhibiting rhythmic behavior of
regular amplitude, such as what is seen during many
seizures. Thus a pre-set threshold (CTH) is employed and
this coefficient of variation used as a second feature for
seizure analysis.
2.1.3. Relative amplitude
At some point in the course of a seizure, the amplitude of
the rhythmic seizure component usually becomes greater
than it was in the pre-seizure background. The average
amplitude is computed as indicated above, as the mean of
the amplitudes of the segments. To measure the amplitude
relative to the background, each EEG epoch is normalized
by the amplitude in the corresponding level in the
background. Based on past experience, we have selected a
background of 15 s after leaving a gap of 20 s (Gotman,
1982). The gap is chosen in order to allow detection of
seizures having a very gradual onset. The epoch size is
varied with the selected level and will be described later.
The relative amplitude is then compared with the pre-set
threshold (ATH).
2.2. Epoch length
The features specified above are computed for scales 3, 4
and 5. The wavelet transformation allows the use of epoch
lengths that are adapted to the frequency range, with a
shorter epoch better adapted to higher frequencies than to
lower frequencies. Hence, we selected an epoch of 1.25 s
(250 samples for a sampling frequency of 200) for level 3
(12 – 25 Hz) and an epoch of 2.5 s (500 samples) for levels 4
(6 – 12 Hz) and 5 (3 –6 Hz). Epochs of 2.5-s were half
901
overlapped to render more similar the representations of all
levels in a 5-s block. Table 1 shows a distribution of the 10
possible epochs involving the 3 levels for a 5 s block of
EEG. Whenever the features for one epoch pass all the
thresholds given above, then a preliminary detection is said
to have taken place.
2.3. Rhythmic non-seizure activity
One of the major causes of false seizure detection is the
occurrence of paroxysmal rhythmic discharges that are part
of the ‘normal’ background. This occurs in the scalp EEG
with alpha activity. It is also frequent in intracerebral
recordings, with the frequent occurrence of bursts of alpha
activity in posterior head regions, or theta activity in
anterior head regions and of various other paroxysmal bursts
of uncertain origin and significance. There is a need for the
algorithm to learn such paroxysmal non-seizure bursts as it
analyzes the EEG. After some time, just like the human
observer, it should be able to ignore such non-seizure
patterns. To identify and reject such rhythms we have
followed a procedure that mimics the human behavior in a
sense that the human decision is based on the past content of
the EEG.
Preliminary detections, as defined above, are an indi-
cation of the presence of paroxysmal rhythmic activity. For
each channel, we retain in a running background array the
mean frequency of the signal for each preliminary detection.
When a new preliminary detection is made, we compare its
frequency to those in the background array. If preliminary
detections with a similar frequency have occurred often in
the past, that detection probably originates from rhythmic
activity that is occurring frequently in this channel. It is
probably not a seizure. The details of this procedure follow.
The running background array of average frequencies
retains at any time the last 100 preliminary detections. At
the time of a seizure there may be a large number of
preliminary detections. In order to avoid filling the array
with many adjacent preliminary detections corresponding to
the same burst or the same seizure, a new preliminary
detection must be separated from the last one entered in the
array by at least 5 epochs, or must have frequencies differing
by a preset threshold. This adaptive array is then utilized to
decide if the preliminary detection is a part of rhythmic
discharge or a candidate to a seizure. For this, let f be the
average frequency in the preliminary detection. Then, the
number of times f is represented in the adaptive background
array in the range f 2 Df and f þ Df is calculated. If this
occurrence is for less than pre-defined times (10% in this
work), the preliminary detection is confirmed.
In order for a channel detection to take place in a 5-s
epoch of EEG, there must be at least two confirmed
preliminary detections (out of the possible 10, see Tables 1
and 2). For a seizure detection to take place, there must be at
least two channel detections separated by less than 15 s in
the same channel.
902 Y.U. Khan, J. Gotman / Clinical Neurophysiology 114 (2003) 898–908

Table 2 of 32 channels (range of 14 – 49). All recordings were

Block diagram of the wavelet method reviewed by experienced EEG technologists who marked
seizures, and subsequently re-reviewed by one of the
authors for confirmation of seizure identification. All
EEGs were reviewed prior to any analysis being performed.
No EEG was rejected as a result of this review.
The first 4-h section including at least one seizure, in the
first 11 patients, were used for developing the method. They
constitute, strictly speaking, the training data set. We then
evaluated the method on two data set: the training set
included all recordings of the first 11 patients; it should have
included only the first recording of each patient, but we
included the other recordings because we felt that these
could not be part of the test set since they belonged to the
same patients as the recordings of the training set. The test
set included all recordings of the last 11 patients.

2.5. Statistical measures

The statistical measures used to assess the performance

2.4. Subjects
EEGs were recorded at our institution in accordance with
the rules of the Research Ethics Board and patients gave
their informed consent. Data were collected from 22
consecutive patients with intracerebral electrodes in whom
we could collect at least the following data: 3 sections of 4 h,
each section including at least one seizure, one seizure-free
4-h section recorded during the day when the subject was
mostly awake, and one seizure-free 4-h section recorded at
night when the subject was mostly asleep. Thus there were 5
4-h sections for each patient. No other selection criteria
were used. Patients had a variety of electrode configur-
ations, including mostly strereotactically implanted depth
electrodes in combination with epidural peg electrodes.
Data were sampled at 200 Hz after antialiasing filtering at
65 Hz, using the Harmonie system from Stellate (Montreal,
Canada). Bipolar montages were analyzed, with an average
Table 3
Comparative results of the two methods on the training data. [‘TP’ are true-positives, ‘FP’ are false-positives, ‘FN’ are false-negatives]
were:
† True positives (TP): the number of seizures identified by
the automated system and by the EEG experts.
† False positives (FP): the events identified as seizures by
the automated system but not by the EEG experts.
† False negatives (FN): the events identified as seizures by
the experts but missed by the automated system.
† Sensitivity: the ratio of the number of true positives to the
total number of seizures marked by the experts.
† Number of false detections/hour.
3. Results
3.1. Training data
The thresholds ETH, CTH and ATH were optimized with
the training data. Table 3 provides a detailed analysis of the

Training data Wavelet method Traditional method

Subject Seizures Hours TP FP FN Sensitivity TP FP FN Sensitivity

1 10 20 10 4 0 100 10 33 0 100
2 29 20 14 18 15 48 11 105 18 38
3 6 20 6 9 0 100 6 58 0 100
4 14 20 13 9 1 93 14 36 0 100
5 7 20 6 2 1 86 7 43 0 100
6 5 20 5 14 0 100 4 102 1 80
7 8 22 7 6 1 88 4 49 4 50
8 3 20 3 23 0 100 3 110 0 100
9 8 20 4 18 4 50 4 40 4 50
10 4 20 4 1 0 100 4 30 0 100
11 3 20 3 9 0 100 3 38 0 100
SUM: 97 222 75 113 22 m ¼ 87.7 70 644 27 m ¼ 83.4
 Y.U. Khan, J. Gotman / Clinical Neurophysiology 114 (2003) 898–908
performance of the wavelet-based method and the method
of Gotman on the training data set. Column 2 gives the
number of seizures as seen by the EEG experts and column
3 the total duration of EEG analyzed for each subject (some
recordings lasted more than 4 h). The average sensitivity
was calculated by averaging the sensitivities for each
subject in order to avoid biasing the results by the patients
who happen to have many seizures (for instance patient 2 in
the training set).
To have a better insight in to the missed seizures by each
method, consider one such seizure of this subject as
illustrated in Fig. 2. It is a short seizure lasting approxi-
mately 20 s on a single channel, RH1 –RH3. This seizure
missed the attention of both the human observers (as no
seizure was reported) and the traditional method. The
wavelet-based method, however, successfully detects this
seizure. Another seizure of similar nature at a later stage in
the recording gave the same results. A seizure that was
missed by the wavelet method is illustrated in Fig. 3. It is a
short seizure with rhythmic bursts and lasts approximately
10 s mainly on channels RH1 – RH3 and RC1 –RC3. Given
the short duration of the seizure and the methodology
developed in our work, such short seizures are likely to be
missed.
The cumulative performance for the subjects represent-
ing the training set can be summarized as follows. The
Fig. 2. Example of brief seizure on RH1-RH3 detected by the wavelet method and missed by the traditional method, in one of the files from the training data.
Detections are shown with ‘ # ’. The wavelet method detects that there is a sudden increase in amplitude at high frequencies, even if there is no overall increase
in amplitude.
903
average sensitivity of the wavelet method was 87.7% and
the average sensitivity of the traditional method was 83.4%.
The primary benefit, however, is the reduction in false
detections from 644 for the traditional method to 113 for the
wavelet method, i.e. a reduction of around 80%. The rate of
false detection fell from 2.9 to 0.5/h.
3.2. Test data
The test set comes from subjects whose EEGs have not in
any way taken part in the development of the wavelet-based
method. Table 4 illustrates the results in a fashion similar to
Table 3. Here again, the False detection rate per hour for the
wavelet method was only 0.3 as compared to 2.4 for the
existing system, thereby reducing the false alarm rate by
approximately 88%. The test database consist 66 seizures in
229.5 h of EEG recording. The proposed method was able to
identify 57 of these seizures (average sensitivity of 86%).
The traditional approach, however, identified 59 of the
seizures (average sensitivity of 90%).
As an example to the detection by two methods, consider
subject 6 who has a seizure of approximately 20 s mainly on
two channels (Fig. 4). This was successfully detected by the
wavelet method whilst the traditional method missed this
seizure. Subject 11 had a short seizure on multiple channels
(Fig. 5). This was detected by the traditional method but it
904 Y.U. Khan, J. Gotman / Clinical Neurophysiology 114 (2003) 898–908

Fig. 3. Example from the training data set of a short seizure with rhythmic bursts on RH1 –RH3 and RC1–RC3. The wavelet method missed this seizure whilst
the traditional method had one detection on RC1 –RC3 (shown with ‘ # ’).

was missed by the wavelet method. The reason again is the but there is a considerable reduction in the rate of false
short duration of the seizure. detections with the wavelet method.

3.3. Combined data sets 4. Discussion

Combining the two sets of data, it was found that the
false detection rate was 0.4/h for the wavelet method and
2.6/h for the traditional method. Average sensitivity was
87% for both methods. The results therefore clearly show
that the two methods have the same seizure detection rate
In the development of this method the emphasis has been
on the fact that there is a large number of false detections
interspersed between the correct detections with most
methods that have been published. Care has also been
taken that the system does not sacrifice seizure detection

Table 4
Comparative results of the two methods on the test data

Test data Wavelet method Traditional method

Subject Seizures Hours TP FP FN Sensitivity TP FP FN Sensitivity

1 5 13.0 5 14 0 100 5 63 0 100
2 6 20.0 5 2 1 83 6 22 0 100
3 3 20.0 1 1 2 33 2 10 1 67
4 4 21.0 4 0 0 100 4 42 0 100
5 11 28.0 11 1 0 100 11 11 0 100
6 12 20.0 7 16 5 58 9 67 3 75
7 6 25.8 6 5 0 100 3 105 3 50
8 9 20.0 9 6 0 100 9 48 0 100
9 3 20.3 3 13 0 100 3 55 0 100
10 4 27.3 4 13 0 100 4 116 0 100
11 3 14.2 2 1 1 67 3 8 0 100
SUM: 66 229.5 57 72 9 m ¼ 85.6 59 547 7 m ¼ 90.2
 Y.U. Khan, J. Gotman / Clinical Neurophysiology 114 (2003) 898–908 905

Fig. 4. A short seizure from the test data set on RH1 –RH3 and RC1–RC3. The wavelet method detects this seizure that was missed by the traditional method.
As in the example of Fig. 2, the sudden increase of amplitude at high frequency is the cause of detection.

Fig. 5. Another example of a short seizure in a subject from the test set where the traditional method detects a seizure whilst wavelet method missed it because
of its short duration.
906 Y.U. Khan, J. Gotman / Clinical Neurophysiology 114 (2003) 898–908
performance in achieving the goal of reducing false
detections. However, it may be mentioned here that it is
nearly impossible for any automated system to achieve a
perfect seizure detection rate with no false positives. This is
because of the great variability in the morphology and the
mode of occurrence of seizures, and the great variability of
non-ictal EEG patterns in SEEG.
Although it was not our purpose, we have in fact
performed a validation of the traditional Gotman method on
a reasonably large intracerebral data set including 22
patients and over 150 seizures, a validation that has not
been done before. We can look at the combined training and
testing data sets, since both were testing sets for this method.
Detection sensitivity was 87%, higher than the 70– 80%
sensitivity that has been reported in validation studies with
scalp data or mixed scalp and intracerebral data. The false
alarm rate was relatively high (2.6/h), a bit higher than the
one or two false alarms per hour seen in scalp recordings.
The wavelet based method resulted in a very similar
sensitivity but a much reduced false alarm rate. It may be
noticed that results did not deteriorate significantly from the
training data set to the testing data set, reflecting in part the
fact that the training set in fact included much data that were
not used in training, but also indicating that the method was
not tightly tailored to the training data and could generalize
well, without losing efficiency.
Intracerebral recordings present a special challenge in
seizure detection for several reasons. Seizures can be
extremely focal, sometimes present in just one or two
channels. It is difficult to rely on spatial features in this
Fig. 6. A short seizure from the test data missed by wavelet method but detected by traditional method.
context (for instance presence of seizure activity in several
spatially congruent channels). Many electrodes are placed
and therefore many channels must be analyzed. As the
number of channels increases, the number of false
detections increases. This increase may not be linear
because of similar activity often appearing in several
channels, but it is nonetheless very important. If we were,
for instance, to select for each patient the 4 or 8 channels in
which the seizure is known to occur, the detection rate
would probably not change, but false detections would most
likely be considerably reduced.
Given that false alarms are quite rare, we turn our
attention to missed seizures, obviously an important issue.
In order to give an insight to the false negatives by the
wavelet method, it should be re-emphasized that in order to
have a preliminary detection on a channel, it is required to
have at least two candidate detections in a 5 s block and that
there should be a repetition of the two candidate detections
on the same channel within 15 s. The assumption is that
seizures will have these detections in a short span of time
while shorter burst of paroxysmal non-seizure activity will
cause candidate detections that are fairly separated. We
have indeed verified that this approach greatly reduced false
detections by eliminating those caused by short rhythmic
bursts. The problem, however, is that seizures lasting less
than 10 s are likely to be missed. This is one of the main
reasons for false negatives in the wavelet method. To
highlight this, consider a seizure from subject 3 of the test
data shown in Fig. 6. This seizure was successfully
identified by the traditional method on channels LE7 –
 Y.U. Khan, J. Gotman / Clinical Neurophysiology 114 (2003) 898–908
Fig. 7. An example of a seizure in the subject 3 of test set where both the methods missed detection. The fast rhythmic activity is of very low amplitude and the
high amplitude rhythmic activity is irregular and of very low frequency.
LE8, LE8 – LE9. In contrast, the wavelet method had two
candidate detections each on LE7 – LE8 and LE8 – LE9 for
one 5-s block. However, there was no repetition of
candidate detections on any of the two channels and the
seizure was not detected. If we include only seizures lasting
15 s or more, the average sensitivity of the wavelet method
in the test set goes up to 92.2%.
On the other hand, as shown before, there were also
instances where the seizures were missed by the traditional
method but were identified by the wavelet method (Figs. 2
and 4). As a last example we consider a case where both
automated methods and the observer missed a small seizure.
Fig. 7 illustrates this seizure for subject 3 in the test data. This
seizure, on channels LE7 –LE8, LE8 – LE9 and LE11– LE12,
exhibits rhythmic activity that is difficult to recognize
visually and it is difficult to imagine that it could reliably
be detected by an automatic method.
We have only discussed seizure detection in the context
of marking long-term EEG recordings in order to facilitate
review. A seizure detection system, if it operates in real
time, can also be used to attract the attention of an observer.
Such an alarm system, which can be called a seizure-
907
warning system, was first presented by Qu and Gotman
(1995, 1997). They noted that the standard seizure detection
system could not be used for that purpose because its false
alarms were too frequent (an average false alarm rate of two
to 3 per hour is obviously not tolerable for nursing staff).
However, with a false alarm rate of 0.3/h or roughly one
false alarm per 3 h, it becomes possible to consider the use
of the wavelet method as an alarm system. The method,
however, was not designed specifically to detect seizures
early in their development. Such an improvement could be
considered in future developments.
We have not mentioned the application of this method to
scalp recordings. It is likely that it would also perform well,
but works remains to be done to deal with the artefacts
common in scalp recordings (EMG, eye movements) that
are absent from intracranial EEG.
Acknowledgements
This work was supported by grant FRN-10189 of the
Canadian Institutes of Health Research.
908 Y.U. Khan, J. Gotman / Clinical Neurophysiology 114 (2003) 898–908
References
Blum DE, Eskola J, Bortz JJ, Fisher RS. Patient awareness of seizures.
Neurology 1996;47:260–4.
Burrus CS, Gopinath RA, Gao H. Introduction to wavelets and wavelet
transforms: a primer. Englewood Cliffs, NJ: Prentice Hall, 1998.
Daubechies I. Ten lectures on wavelets. CBMS – NSF regional conference
series in applied mathematics. Philadelphia: SIAM, 1992.
Gabor AJ. Seizure detection using a self-organizing neural network:
validation and comparison with other detection strategies. Electro-
enceph clin Neurophysiol 1998;107:27– 32.
Gabor AJ, Leach RR, Dowla FU. Automatic seizure detection using a self-
organizing neural network. Electroenceph clin Neurophysiol 1996;99:
257– 66.
Gotman J. Automatic recognition of epileptic seizures in the EEG.
Electroenceph clin Neurophysiol 1982;54:530–40.
Gotman J. Automatic seizure detection: improvements and evaluation.
Electroenceph clin Neurophysiol 1990;76:317–24.
Gotman J, Gloor P. Automatic recognition and quantification of interictal
epileptic activity in the human scalp EEG. Electroenceph clin
Neurophysiol 1976;49:513 –29.
Harding GW. An automated seizure monitoring system for patients with
indwelling recording electrodes. Electroenceph clin Neurophysiol
1993;86:428–37.
Kalayci T, Ozdamar O. Wavelet preprocessing for automated neural
network detection of EEG spikes. IEEE Eng Med Biol Mag 1995;14:
160 –6.
Osorio I, Frei MG, Wilkinson SB. Real-time automated detection and
quantitative analysis of seizures and short-term prediction of clinical
onset. Epilepsia 1998;39:615–27.
Park HS, Lee YH, Lee DS, Kim SI. Detection of epileptiform activity using
wavelet and neural network. Proceedings 19th International Conference
IEEE/EMBS 1997; 1194–1197.
Pauri F, Pierelli F, Chartrian GE, Erdly WW. Long term EEG-video-audio
monitoring: computer detection of focal EEG seizure patterns.
Electroenceph clin Neurophysiol 1992;82:1–9.
Qu H, Gotman J. A seizure warning system for long-term epilepsy
monitoring. Neurology 1995;45:2250 –4.
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;44:115– 22.
Salinsky MS. A practical analysis of computer based seizure detection
during continuous video-EEG monitoring. Electroenceph clin Neuro-
physiol 1997;103:445 –9.
Webber WR, Lesser RP, Richardson RT, Wilson K. An approach to seizure
detection using an artificial neural network (ANN). Electroenceph clin
Neurophysiol 1996;98:250–72.
Williams WJ, Zaveri HP, Sackellares JC. Time-frequency analysis of
electrophysiology signals in epilepsy. IEEE Eng Med Biol Mag 1995;
14:123–43.