- Email: [email protected]
Automatic seizure detection in the newborn: methods and initial evaluation
ELSEVIER
Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
Automatic seizure detection in the newborn: methods and initial evaluation J. Gotman a'*, D. Flanagan b, J. Zhang b, B.
Rosenblatt b
aThe Montreal Neurological Institute and Hospital and McGill University, 380l University Street, Montreal, PQ, H3A 2B4 Canada bThe Montreal Children's Hospital, Montreal, PQ, H3A 2B4 Canada Accepted for publication: 17 February 1997
Abstract Seizures are most common in the newborn period, but at that age seizures can be very difficult to identify by clinical observation. Therefore the EEG plays an even greater role in newborns than in older children and adults. The electrographic features of seizures and EEG background in the newborn are, however, very different to those found in adults. We present a set of methods for the automatic detection of seizures in the newborn. The methods are aimed at detecting a wide range of patterns, including rhythmic paroxysmal discharges at a wide range of frequencies, as well as repetitive spike patterns, even when they are not very rhythmic. The methods were developed using EEGs obtained from 55 newborns, recorded at 3 hospitals that used differing monitoring protocols. A total of 281 h of recordings containing 679 seizures were analyzed. An initial evaluation indicated that 71% of the seizures and 78% of seizure clusters (group of seizures separated by less than 90 s) were detected, with a false detection rate of 1.7/h. The methods were developed so that they can be implemented to operate in real time. © 1997 Elsevier Science Ireland Ltd.
Keywords: Seizure; Detection; Newborn; EEG
I. Introduction Seizures are the most common sign of neurological disorder among newborns, yet such events can be difficult to identify clinically (Dreyfus-Brisac and Monod, 1964; Fenichel and Fitzpatrick, 1984; Mizrahi and Kellaway, 1984; Rose and Lombroso, 1970; Volpe, 1995) and may not occur until some time (up to several days) after the insult (Volpe, 1995). Furthermore, some newborns are pharmacologically paralyzed to improve ventilation, thus making visual identification of seizure behaviors impossible (Tharp and Laboyrie, 1983). At present, prolonged EEG monitoring of these patients provides the best information about presence of seizures, and is also ideal clinical practice for monitoring the response to anticonvulsant treatment. Unfortunately, for most Neonatal Intensive Care Units, prolonged monitoring is prohibitively expensive and impractical. An automated seizure detection method would substantially reduce the amount of data to be reviewed and so improve
* Corresponding author. Tel.: +1 514 398 1953; fax: +1 514 398 8106; e-mail: jean @rclvax.medcor.mcgill.ca.
0013-4694/97/$17.00 © 1997 Elsevier Science Ireland Ltd. All rights reserved PII S0013-4694(97)00003-5
the cost efficiency of prolonged monitoring. Automated seizure detection methods are now routinely used for longterm epilepsy monitoring in adults, but the EEG of the newborn differs significantly from that of the adult. The EEG in the newborn mainly consists of a random mixture of delta and theta frequencies (Scher et al., 1994) and has very few sustained periods of clear rhythmicity. Newborn electrographic seizures however, are characterized by their rhythmicity (Lui et al., 1992). Lui et al. (1992) utilized the rhythmicity of newbom seizures and suggested a method of seizure detection based on the autocorrelation function, in which they looked for multiple peaks in 6-s epochs. They obtained very promising results, although their method of evaluation could be criticized (Gotman et al., 1997 (accompanying paper)). Our experience of seizures in the newborn confirms that most seizures are characteristically rhythmic. Some newborn seizures, however, do not fit neatly into this category. Furthermore, some seizures in the newborn have a very gradual onset when compared to seizures in adults, and so detection methods that rely on sudden onset of seizures may fail. For these reasons, we propose a more comprehensive approach to seizure detection in the newborn.
EEG 96175
Jr Gotman et al. / Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
We present a relatively complex method based on 3 types of analysis of the EEG: spectral analysis to find rhythmic discharges at various frequencies, spike detection to find groups of abnormal spikes that may not be rhythmic, and low-pass digitally filtered EEG to find very slow discharges of 1/2 Hz or even slower. We will describe these methods and their rationale, give examples of their performance, and discuss their drawbacks. A full evaluation of the combined use of the 3 methods with data from 3 hospitals and detailed analysis of correct and incorrect detections is described in Gotman et al., 1997 (accompanying paper).
2. Methods
2.1. Subjects EEGs containing electrographic seizures were obtained from 43 newborns studied at 3 centers, Montreal Children's Hospital (Montreal, Canada), Sydney Children's Hospital (Sydney, Australia; formally known as Prince of Wales Children's Hospital) and Texas Children's Hospital (Houston, TX). The average pest conceptional age at time of study was 40 weeks (range: 25-52 weeks). There were no selection criteria other than age, and consecutive recordings were used. No records were rejected (regardless of type of seizure or record quality). Twelve consecutive recordings from patients without electrographic seizures were also obtained. No records were rejected (regardless of record qu,'dity). 2.2. Centers 2.2.1. Montreal Children's Hospital An 11-electrode montage based on the 10-20 system (the 'little H') was used. We obtained 11 recordings lasting an average 9.8 h (range: (I.5-20 h). These recordings were most often overnight recordings made in the Intensive Care Unit with infrequent supervision by the EEG technician. EEGs were recorded with the Monitor program (Stellate Systems, Montreal), with sampling at 200 Hz and analog filters set at 0.5 and 30 Hz on the amplifiers (Grass Model 12, Quincy, MA) 2.2.2. Sydney Children':r Hospital The montage used was a 12 electrode modification of the 10-20 system. We obtai~Led 25 recordings lasting an average 5.6 h (range: 3.7-6.2 h). These recordings were most often prolonged recordings made in the intensive care unit with infrequent supervision by the EEG technician. Original recordings lasted from 3 to 113.5 h, but only the first 6-h section that included a seizure was used in this study. EEGs were first recorded on the LaMont video-patient monitoring system (Medical Systems International, Sydney, Australia) with filter settings at 70 Hz (high pass) and 0.3 s (time constant). From each re,cording, we digitized the first 6-h
357
section that included a seizure using the Monitor program (Stellate Systems, Montreal). 2.2.3. Texas Children's Hospital An 11-electrode neonatal montage was used (Mizrahi, 1986). We obtained 19 recordings lasting an average 1.8 h (range: 1.6-2.1 h). These recordings were obtained under continuous supervision by the EEG technician. EEGs were recorded on a Nihon Kohden EEG machine (Nihon Kohden, Tokyo, Japan) and digitized with a system developed at Baylor College of Medicine, with a sampling rate of 185 Hz. A program was written so that these files could be analyzed in the same way as the other files. We made the approximation of assuming a 200-Hz sampling rate for these files, thus causing an error of 7.5%. EEG activity would then appear 7.5% faster to our analysis program, an approximation that we considered acceptable in view of the fact that there is no known clinical significance to a difference in frequency between 1 and 1.075 Hz or between 10 and 10.75 Hz in the context of determining if an event is or is not a seizure. 2.3. Automated detections Initial investigations using a detection method based on spectral analysis revealed 2 types of seizure that were not detected by that method. Some clear seizures consisting of sometimes arrhythmic runs of high amplitude spikes were often undetected (Fig. 1, also Fig. 5-37 of Stockard-Pope et al., 1992). Similarly, some seizures with very low frequency discharges (<0.5 Hz) were noted among patients with severely depressed EEGs and were also undetected by the initial method (Fig. 2, also Fig. 5-32 of Stockard-Pope et al., 1992). In an attempt to capture all seizure types, 3 detection methods were therefore used. 2.3.1. Method 1: Detection of rhythmic discharges (0. 5 - 10
nz) This method employed spectral analysis of sequential 10s epochs of the EEG in a sliding window. This window moved along the EEG in 2.5-s steps, thus adjacent epochs overlapped by 75%. This overlap ensured that a 10-s seizure would be almost completely represented in one epoch at least. The algorithm was designed to extract features from each epoch and compare them with the features from 2 previous epochs (the 'background') that were 60 s behind F3-C3 ~ C3-P3
~
~
~"'-'~,~-'--~
~ -="~-'~'~,"'-~~
P3-01 ~ . - , . v - ~ . - - ~ - ~ ' ~ ~ Ca~o21Mat 23:19
I
5 seG
lOO~V
I
Fig. 1. Seizure identified visually by the EEGer and consisting of a run of arrhythmic spikes that were not detected by the spectral analysis method.
358
J. Gotman et al. /Electroencephalography and clinical Neurophysiology 103 (1997) 356-362 •
Fp2-F8 ~
T6-02
"
~
L/""-~
~-'~'
~/,~~
~f ~
~
We computed the power in the full width half-maximum band of the dominant peak and compared it to the power in the same frequency band in the spectrum of the background (Fig. 4). This is to determine if the rhythmic activity that may be present in the current epoch is new compared to the background.
~ 200/~V
sire124Mar10:29 5 sec
Fig. 2. Seizure identified visually by the EEGer and consisting of a very low frequency discharge that was not detected by the spectral analysis method. the current epoch. A constantly updated background was selected to reflect the visual interpretation process; it also has the great advantage over a fixed background that the results are not dependent on the choice of that fixed background. A long gap between background and current epoch was selected to allow the detection of seizures that start very gradually (Fig. 3). The frequency spectrum was computed for each epoch (the FFF length was 2048 points, so that the epoch length was in fact 10.24 s). The spectrum had a frequency range from 0 to 100 Hz, with a resolution of 0.1 Hz. The following variables were computed for each epoch: •
Dominant frequency
Among all the peaks in a spectrum, the peak with the largest average power in its full width half maximum band was defined as the dominant peak. The dominant frequency was defined as the frequency of the dominant peak (Qu and Gotman, 1997). The purpose of this feature was to find the most prominent rhythmic component of the epoch. •
Width of the dominant spectral peak
It was defined as the difference between the frequency corresponding to half the amplitude of the dominant peak in its falling slope and the frequency corresponding to half of this amplitude in the rising slope. If this width is small, the dominant peak represents genuinely 'rhythmic' activity.
Fp2-F8 F8-T6 T6-02 ~ - ;
Power ratio
~ , J \/~/"
....
t
•
Stability of current epoch
We integrated the average amplitude of the signal in each of the four 2.5-s sections of a 10-s epoch, and determined the ratio of the largest of the 4 values to the smallest one. We wanted to include as possible seizure candidates only sections that were relatively stationary, i.e. where this ratio was not too large (indicating that the amplitude of the smallest section was not very different from the amplitude of the largest section). 60-Hz Indicator
•
Poor electrode contact is frequent during long-term monitoring. It is important to detect it since it can lead to spurious slow patterns in the spectrum that can be mistaken for seizures. Although we only looked at frequencies up to 50 Hz in the spectrum, we were able to detect poor electrode contact by measuring activity at 20 Hz, where a sharp peak is clearly present whenever 60 Hz activity is very large. We presume that this peak comes from aliasing of the 3rd harmonic of 60 Hz (180 Hz). Although this activity should be eliminated by our anti-aliasing filter, it is possible that it can sometimes be sufficiently large to remain noticeable. It is also possible that this is a sub-harmonic of 60 Hz. Whatever its origin, it was a useful marker of poor electrode contact. We calculated the ratio of the peak of the spectrum at 20 Hz to the total energy in the spectrum. •
Patient disconnected indicator
Detection and updating of the background was suspended in any channel if the total energy in the spectrum was extremely low. This happens in particular if the patient or one amplifier is disconnected. It is important to eliminate such epochs from the background because they can lead to extremely high power ratios (see above definition) and thus to false Table 1 Combinations of values of dominant frequency, width of dominant peak and power ratio are acceptable for seizure detection in Method 1
e~a 22Maroe:ae
I 200/~V I
5 sec
Fig. 3. Seizure with a very gradual onset, common in newborn seizures. It has an initial low frequencydischarge starting about 40 s before the higher frequency discharge. Seizure first detected in the 10-s epoch ending at the arrow, and subsequently detected 41 times.
Dominant frequency First combination 0.5-1.5 Second combination 1.5-10 Third combination 1.5-10
Width of Power ratio dominantpeak _<0.6 _<0.6 _<1
3-4 2-4 4-80
359
J. Gotman et al. I Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
C
A Fp2-T4 T3-C3 ~
c3-cz ~ ~ ~ ' - - ~ ¢ ' - - ~
20t 1.5 1.0
O.
o.s/,I
-, - " - ~ - - - > ~ . - - ~
ol:
o
i
2
3
4
;
~
7
.
..~ ......~_
f--
z/'/,4
,
,
, ~'-~_
1
2
3
4
o
----_ 6
s
r
Freq(Hz)
Freq (Hz)
E I
Background epochs 20 sec
I
I
Gap 60 sec
I Current epoch lOsec
Fig. 4. Schematic representation of the spectral analysis method. (A) EEG background for comparative analysis. (B) Spectrum of channel T3-C3 for the section shown in A. (C) EEG in the current epoch. (D) Spectrum of channel T3-C3 for the section shown in C. (E) Representation of the relative timing of current epoch and comparative background. The background consists of 2 epochs ending 60 s prior to the current epoch. Dominant frequency and width of dominant spectral peak are calctLlated on spectrum in D. The power ratio is then defined as the ratio of the area hatched in D to the area hatched in B; the location on the frequency axis of the area in B is the same as that in D.
detections. After examination of the training set, the following thresholds were selected for the above variables. The first 3 parameters define a 3-.dimensional space. We defined boundaries in that space that determine where detections belong. These boundaries are given in Table 1. If an epoch was acceptable according to these 3 parameters, then it was considered a de,tection if its Stability was inferior to 3, its 60-Hz indicator inferior to 0.8%, and its Patient disconnect indicator superior to 0.
2.3.2. Method 2: Detection of multiple spikes This method is based on the spike detection procedure of Gotman et al. (1979). The procedure was modified because spikes in newborns tend to be of much longer duration than those in adults. We therefore performed high-pass digital filtering of the EEG prior to spike detection, using an IIR filter of order 3 with a cut-off frequency of 2 Hz. Fig. 5 shows the effect of this filtering on the background EEG and on spikes. A seizure detection was made when 6 or more spikes were detected in a 10-s epoch. 2.3.3. Method 3: Detection of very slow rhythmic discharges Method 1 described above often failed to detect very slow activity (around 0.5 Hz), particularly if the waveforms were complex. When the waveform is particularly far from sinusoidal, it is frequent that the peak of the spectrum is quite broad and that a high hm'monic content is present. This renders difficult the detection of the dominant peak, as defined above. We therefore decided to filter the EEG with a low-pass digital filter of order 2 and with a cut-off frequency of 1 Hz. The effect of such a filter on a very slow
seizure discharge is shown on Fig. 6. After this filtering, we applied the seizure detection method of Gotman (1990), but used the same epoch duration, gap and background duration as used for Method 1 above.
2.4. Visual seizure identification and comparison with automatic detection Prior to automated detection, electrographic seizures were identified visually by the EEGer. An electrographic seizure was defined as a burst of paroxysmal rhythms that were considered to be of cerebral origin and had a duration of 10 s or more and evolved temporally and spatially (Bye
A c3-P3 ~
......
P3-O~
FB3-C3 C3-P3 -
'
P3-01 ~
¢ "
Cap 21 Mar
' "
-
" "
'
~ '
-
~ =
21:34
I IO0/,IV 1 seo
J
Fig. 5. Example of newborn seizure consisting of runs of arrhythmic spikes (sometimes not detected by the spectral analysis method). Spikes in the newborn EEG, however, are relatively slow, and are not detected with the normal spike detection method. (A) Original seizure with relatively long duration of spikes. (B) Seizure filtered with a high pass filter: infinite impulse response (IIR), 3rd order, cutoff 2 Hz. After filtering, the spikes are briefer, sharper, and easier to detect by standard spike detection method.
360
J. Gotman et al. /Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
3. Results
A F4-P4
-
~
~
_
~
'~---~ I ~-,j
~
P4-o2~ - - - - - - - ~ J " V ~ ~/.,~--_.__. ~ B F4-P4 P4-02
~
'
~
-
-
~
Sam24 Mat 10:10
I 200/~V 2 SOC
Fig. 6. Example of newborn seizure consisting of a very slow frequency discharge. The waveforms are often complex and such seizures are sometimes not detected by the spectral analysis method. (A) Original seizure. (B) Seizure after application of a low pass filter: IIR, 2nd order, cutoff 1 Hz.
and Flanagan, 1995; Clancy and Legido, 1987; Scher et al., 1993). When several seizures follow each other closely, it is possible that the second or third seizure in a group is not detected due to our definition of 'background': if a seizure is itself the 'background' for the next seizure, this latter seizure may not be detected because it is then not different from the 'background'. Since the purpose of the detection method is to alert to the presence of seizures, we assume that if one seizure of a group is detected, the user will most often look before and after the seizure and will therefore see possible neighboring seizures. W e therefore defined the concept of 'cluster of seizures' as follows: a set of 2 or more seizures for which the interval between the end o f one is less than 90 s from the beginning of the next one. All EEGs were reviewed, and start and end time for all seizures were documented. Automated detections that occurred within 30 s of each other were grouped as a single detection. Automatic detections were then compared to visual identification: if an automatic detection occurred any time between the beginning and end of a seizure, the detection was considered as a valid seizure detection. Similarly, if a seizure within a cluster of seizures was detected, there was a valid cluster detection. All other detections were counted as false detections (FDs).
Table 2
W e will first give some examples of the seizures detected by each method, as well as of the false detections specific to each detection method. These are of course only examples and it is impossible to give a comprehensive view of the variety of seizure patterns and false detections. W e will then give results of each method on the data set described above. A detailed analysis of the results by institution, technical quality of recording, type and duration of seizure, for the training and testing data sets, is given in the accompanying publication (Gotman et al. (1997), accompanying paper). Since there is a large variability in the number of seizures per patient (the range is from 1 to 70 seizures and 1 to 49 clusters), we assessed the results by calculating for each patient the percentage of seizures detected and then averaging these percentages over all patients. If we just counted the percentage of seizures detected, then patients with a large number of seizures would completely overshadow those with one or 2 seizures. Since the type of seizure varies greatly from patient to patient, it is important to have all seizure types evenly represented. A total of 281.5 h of E E G from 55 patients were analyzed. Clinical assessment identified 679 seizures, and 228 of those occurred within 90 s of another seizure and so 451 'seizure clusters' were identified. 3.1. Results o f a u t o m a t e d detection
Table 2 reveals the average percentage of seizures and seizure clusters detected by each o f the 3 methods. While Method 1 detected the highest average percentage of seizures, the other 2 methods contributed to the overall detection success. Two detection methods sometimes detected the same seizure: for example a seizure detected by Method 1 may slow down at the end of its evolution and then be detected by Method 3. In this way the detection methods could sometimes reinforce each other. False detection rates were similar for each detection method, however each method tended to falsely detect differing events. As a result, the combined false detection rate is higher. Method 1 occasionally detected movement artifact that occurred when the mother was nursing the baby, and possibly rocking or patting it (Fig. 7). Method 2 occasionally detected excessive electrode noise, along with high
T~-o~~ ~ ~ , , ~ ¢ ~ ~ ~ ' ~ .
Results of seizure detection on the testing set Method 1 Method 2 Method 3 Methods 1-3 combined Average % seizures detected Average % clusters detected Average FD/hour
64
35
13
71
71
40
15
78
0.8
0.8
0.2
1.7
°
5 sec
I
Fig. 7. False detection occurring in the 10-s epoch ending at the arrow. This event was detected by the spectral analysis method in channels T6-O2 and P4-O2 and was caused by the mother patting the baby while nursing (visible on the video recording).
J. Gotman et aL /Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
361
voltage EMG artifacts, and sucking artifacts (Fig. 8). Method 3 also tended to occasionally detect slow movement artifacts, such as when the baby was moving or distressed (Fig. 9).
I200~V
Aus 20 Apt 13:40
4. Discussion
5 SOC
The newborn EEG does not contain the sustained rhythms that are characteristic of the EEG of older children, such as alpha, mu, central rhythms and the so called posterior slow waves of youth. Newborn seizures, however, are often rhythmic and this is most often their defining characteristic. An awareness of these features leads to the realization that spectral analysis may provide a means of seizure detection in the newborn. We designed an algorithm that extracted important features about rhythmicity, power and stability from the spectrum of a current epoch and compared those features with those of an earlier epoch in the background. In this way we were able to identify clear periods of paroxysmal rhythmicity (and therefore likely seizures). Some of the principles of this method are quite similar to the seizure detection method for adults developed by Gotman (Gotman, 1982; Gotman, 1990). Important differences include a much longer detection epoch (10 instead of 2 s) and a much longer background, reflecting the very low frequencies of the newborn EEG, as well as a spectral analysis method rather than wave decomposition in oTder to detect rhythmic discharges even in the presence of complex waveforms. Similar principles were used by Lui et al. (1992), who used autocorrelation rather than spectra to detect rhythmicity. We found that our spectral analysis method identified a large proportion of seizures, but did miss some events that were clearly seizures, but were either of very low frequency (<1 Hz) or runs of arrhythmic spike,;. For this reason we introduce 2 separate methods, specifically designed to identify both of these types of events. While these methods were successful, they identified relatively few seizures. Our experience leads us to believe the reason for this is that while very low frequency seizures and arrhythmic spike seizures are very easily identified visually, they are relatively uncommon among the newborn. Furthermore, each detection method also contributed a good number of false detections. As long as they are not too frequent, false detections do not represent a significant burden for a seizure detection method because they only represent extra EEG that has to Fp2-Fa~ . ~ ~ f . - - ~ , , ~ , ~ ~
~
T6-O2 ~
,
/
~
.
,
~
-
~
12oo~
Ang 23 Mar 1~42 5 see
Fig. 8. False detection occurring in the ]0-s epoch ending at the arrow. Tl~s event was detected in channels Fp2-F8 by the multiple spike detection method and was caused by a sucking artifact while the newborn was being fed.
Fig. 9. False detection occurring in the 10-s epoch ending at the arrow. This event was detected in channels F4-Cz and F3-P3 by the method used to identify slow rhythmic activity. It was caused by movement of the newborn while distressed and crying.
be reviewed. A false detection rate of 2/h only represents 40-60 s of EEG for each hour of recording that has to be unnecessarily reviewed. It is much more important not to miss seizures, than it is have a low false detection rate. Nonetheless, newborns in an intensive care situation have the potential of having a large variety and number of events that are clearly not seizures, but fulfill our detection criteria. For example, critically ill newborn often undergo cranial ultrasounds which can cause a number of electrographic artifacts. Less ill patients may be gently rocked or patted if distressed, and some newborn may even be held and fed by the mother. All these events can cause different false detections (for example the high EMG content of sucking may trigger the spike detection method). The method of Qu and Gotman (1993) allowed an important reduction in false detection by learning about the false detections particular to a patient during the first day of monitoring and then avoiding similar events in subsequent days. This principle is less applicable to newborn monitoring, which usually does not extend over more than a couple of days. The 3 methods used here were not combined to run simultaneously in real time in this study, but the methods were designed so that it is be possible and the combined method could run on line on a standard processor (we estimate that a 100-MHz PC would be sufficient). This raises the question as to whether the gain achieved in broadening the types of seizure detected by using 3 separate methods is worth the cost of an increase in false detections.
Acknowledgements We are grateful to Carol Leitner and the other staff of the EEG Laboratory of the Montreal Children's Hospital for their collaboration. We are also grateful to Dr. E. Mizrahi, Baylor College of Medicine and Dr. A. Bye, Sydney Children' s Hospital, for providing many of the recordings used in this study. This work was supported in part by a grant from the Medical Research Council of Canada and Stellate Systems through the University-Industry program of the Medical Research Council (Grant MRC-UI-11514).
362
J. Gotman et al. /Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
References Bye, A.M.E. and Flanagan, D. Spatial and temporal characteristics of neonatal seizures. Epilepsia 1995; 36: 1009-1016. Clancy, R.R. and Legido, A. The exact ictal and interictal duration of electroencephalographic neonatal seizures. Epilepsia 1987; 28: 537541. Dreyfus-Brisac, C. and Monod, N. Electroclinical studies of status epilepticus and convulsions in the newborn. In: P. Kellaway and I. Petersen (Eds.), Neurological and Electroencephalographic Correlative Studies in Infancy. New York: Grune and Stratton, 1964, pp. 250-272. Fenichel, G.M. and Fitzpatrick, J.E. Difficulty in clinical identification of neonatal seizures: an EEG monitoring study. Electroenceph. clin. Neurophysiol.1984; 58: 33. Gotman, J. Automatic seizure detection: Improvements and evaluation. Electroenceph. clin. Neurophysiol. 1990; 76: 317-324. Gotman, J. Automatic recognition of epileptic seizures in the EEG. Electroenceph, clin. Neurophysiol. 1982; 54: 530-540. Gotman, J., Ives, J.R. and Gloor, P. Automatic recognition of iuterictal epileptic activity in prolonged EEG recordings. Electroenceph. clin. Neurophysiol. 1979; 46:.5 ! 0-520. Gotman, J., Flanagan, D., Rosenblatt, B., Bye, A. and Mizrahi, E.M. Evaluation of an automatic seizure detection method for the newbom. Electroenceph. clin. Neurophysiol. 1997; 103: 363-369. Lui, A., Hahn, J.S., Heldt, G.P. and Coen, R.W. Detection of neonatal seizures through computerized EEG analysis. Electroenceph. clin. Neurophysiol. 1992; 82: 30-37. Mizrahi, E.M. Neonatal electroencephalography. Clinical features of the newborn, techniques of recording and characteristics of the normal EEG. Am. J. EEG Technol. 1986; 26: 81-103.
Mizrahi, E.M. and Kellaway, P. Clinical, electroencephalographic, therapeutic, and pathophysiologic studies of neonatal seizures. In: C.G. Wasterlain and P. Vert (Eds.), Neonatal Seizures. Raven Press, New York, 1990, pp. 1-12. Mizrahi, E.M. and Kellaway, P. Characterization and classification of neonatal seizures. Neurology 1987; 37: 1837-1844. Mizrahi, E.M. and Kellaway, P. Characterization of seizures in neonates and young infants by time-synchronized EEG polygraphic/video monitoring [abstract]. Ann. Neurol. 1984; 16: 383. Qu, H. and Gotman, J. A self-adapting algorithm for the detection of seizure onset in long-term EEG monitoring: possible use as a warning device. IEEE Trans. Biomed. Eng. 1997; 00: in press. Qu, H. and Gotman, J. Improvement in seizure detection performance by automatic adaptation to the EEG of each patient. Electroenceph. clin. Neurophysiol. 1993; 86: 79-87. Rose, A.L. and Lombroso, C.T. Neonatal seizure states. A study of clinical, pathological, and electroencephalograpic features in 137 full-term babies with a long-term follow-up. Pediatrics 1970, 45: 404-425. Scher, M.S., Hamid, M.Y., Steppe, D.A., Beggarly, M.E. and Painter, M.J. Ictal and interictal electrographic seizure durations in preterm and term neonates. Epilepsia 1993; 34: 284-288. Scher, M.S., Sun, M., Steppe, D.A., Guthrie, R.D. and Sclabassi, R.J. Comparison of EEG spectral and correlation measures between healthy term and preterm infants. Pediatr. Neurol. 1994; 10: 104-108. Stockard-Pope, J.E., Werner, S.S. and Bickford, R.G. Atlas of Neonatal Electroencephalography. New York: Raven Press, 1992. Tharp, B.R. and Laboyrie, P.M. The incidence of EEG abnormalities and outcome in infants paralyzed with neuromuscular blocking agents. Crit. Care Med. 1983; 11: 926-929. Volpe, J.J. Neonatal seizures. In: Neurology of the Newborn. W.B. Saunders, Philadelphia, 1995, pp. 172-207.
Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
Automatic seizure detection in the newborn: methods and initial evaluation J. Gotman a'*, D. Flanagan b, J. Zhang b, B.
Rosenblatt b
aThe Montreal Neurological Institute and Hospital and McGill University, 380l University Street, Montreal, PQ, H3A 2B4 Canada bThe Montreal Children's Hospital, Montreal, PQ, H3A 2B4 Canada Accepted for publication: 17 February 1997
Abstract Seizures are most common in the newborn period, but at that age seizures can be very difficult to identify by clinical observation. Therefore the EEG plays an even greater role in newborns than in older children and adults. The electrographic features of seizures and EEG background in the newborn are, however, very different to those found in adults. We present a set of methods for the automatic detection of seizures in the newborn. The methods are aimed at detecting a wide range of patterns, including rhythmic paroxysmal discharges at a wide range of frequencies, as well as repetitive spike patterns, even when they are not very rhythmic. The methods were developed using EEGs obtained from 55 newborns, recorded at 3 hospitals that used differing monitoring protocols. A total of 281 h of recordings containing 679 seizures were analyzed. An initial evaluation indicated that 71% of the seizures and 78% of seizure clusters (group of seizures separated by less than 90 s) were detected, with a false detection rate of 1.7/h. The methods were developed so that they can be implemented to operate in real time. © 1997 Elsevier Science Ireland Ltd.
Keywords: Seizure; Detection; Newborn; EEG
I. Introduction Seizures are the most common sign of neurological disorder among newborns, yet such events can be difficult to identify clinically (Dreyfus-Brisac and Monod, 1964; Fenichel and Fitzpatrick, 1984; Mizrahi and Kellaway, 1984; Rose and Lombroso, 1970; Volpe, 1995) and may not occur until some time (up to several days) after the insult (Volpe, 1995). Furthermore, some newborns are pharmacologically paralyzed to improve ventilation, thus making visual identification of seizure behaviors impossible (Tharp and Laboyrie, 1983). At present, prolonged EEG monitoring of these patients provides the best information about presence of seizures, and is also ideal clinical practice for monitoring the response to anticonvulsant treatment. Unfortunately, for most Neonatal Intensive Care Units, prolonged monitoring is prohibitively expensive and impractical. An automated seizure detection method would substantially reduce the amount of data to be reviewed and so improve
* Corresponding author. Tel.: +1 514 398 1953; fax: +1 514 398 8106; e-mail: jean @rclvax.medcor.mcgill.ca.
0013-4694/97/$17.00 © 1997 Elsevier Science Ireland Ltd. All rights reserved PII S0013-4694(97)00003-5
the cost efficiency of prolonged monitoring. Automated seizure detection methods are now routinely used for longterm epilepsy monitoring in adults, but the EEG of the newborn differs significantly from that of the adult. The EEG in the newborn mainly consists of a random mixture of delta and theta frequencies (Scher et al., 1994) and has very few sustained periods of clear rhythmicity. Newborn electrographic seizures however, are characterized by their rhythmicity (Lui et al., 1992). Lui et al. (1992) utilized the rhythmicity of newbom seizures and suggested a method of seizure detection based on the autocorrelation function, in which they looked for multiple peaks in 6-s epochs. They obtained very promising results, although their method of evaluation could be criticized (Gotman et al., 1997 (accompanying paper)). Our experience of seizures in the newborn confirms that most seizures are characteristically rhythmic. Some newborn seizures, however, do not fit neatly into this category. Furthermore, some seizures in the newborn have a very gradual onset when compared to seizures in adults, and so detection methods that rely on sudden onset of seizures may fail. For these reasons, we propose a more comprehensive approach to seizure detection in the newborn.
EEG 96175
Jr Gotman et al. / Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
We present a relatively complex method based on 3 types of analysis of the EEG: spectral analysis to find rhythmic discharges at various frequencies, spike detection to find groups of abnormal spikes that may not be rhythmic, and low-pass digitally filtered EEG to find very slow discharges of 1/2 Hz or even slower. We will describe these methods and their rationale, give examples of their performance, and discuss their drawbacks. A full evaluation of the combined use of the 3 methods with data from 3 hospitals and detailed analysis of correct and incorrect detections is described in Gotman et al., 1997 (accompanying paper).
2. Methods
2.1. Subjects EEGs containing electrographic seizures were obtained from 43 newborns studied at 3 centers, Montreal Children's Hospital (Montreal, Canada), Sydney Children's Hospital (Sydney, Australia; formally known as Prince of Wales Children's Hospital) and Texas Children's Hospital (Houston, TX). The average pest conceptional age at time of study was 40 weeks (range: 25-52 weeks). There were no selection criteria other than age, and consecutive recordings were used. No records were rejected (regardless of type of seizure or record quality). Twelve consecutive recordings from patients without electrographic seizures were also obtained. No records were rejected (regardless of record qu,'dity). 2.2. Centers 2.2.1. Montreal Children's Hospital An 11-electrode montage based on the 10-20 system (the 'little H') was used. We obtained 11 recordings lasting an average 9.8 h (range: (I.5-20 h). These recordings were most often overnight recordings made in the Intensive Care Unit with infrequent supervision by the EEG technician. EEGs were recorded with the Monitor program (Stellate Systems, Montreal), with sampling at 200 Hz and analog filters set at 0.5 and 30 Hz on the amplifiers (Grass Model 12, Quincy, MA) 2.2.2. Sydney Children':r Hospital The montage used was a 12 electrode modification of the 10-20 system. We obtai~Led 25 recordings lasting an average 5.6 h (range: 3.7-6.2 h). These recordings were most often prolonged recordings made in the intensive care unit with infrequent supervision by the EEG technician. Original recordings lasted from 3 to 113.5 h, but only the first 6-h section that included a seizure was used in this study. EEGs were first recorded on the LaMont video-patient monitoring system (Medical Systems International, Sydney, Australia) with filter settings at 70 Hz (high pass) and 0.3 s (time constant). From each re,cording, we digitized the first 6-h
357
section that included a seizure using the Monitor program (Stellate Systems, Montreal). 2.2.3. Texas Children's Hospital An 11-electrode neonatal montage was used (Mizrahi, 1986). We obtained 19 recordings lasting an average 1.8 h (range: 1.6-2.1 h). These recordings were obtained under continuous supervision by the EEG technician. EEGs were recorded on a Nihon Kohden EEG machine (Nihon Kohden, Tokyo, Japan) and digitized with a system developed at Baylor College of Medicine, with a sampling rate of 185 Hz. A program was written so that these files could be analyzed in the same way as the other files. We made the approximation of assuming a 200-Hz sampling rate for these files, thus causing an error of 7.5%. EEG activity would then appear 7.5% faster to our analysis program, an approximation that we considered acceptable in view of the fact that there is no known clinical significance to a difference in frequency between 1 and 1.075 Hz or between 10 and 10.75 Hz in the context of determining if an event is or is not a seizure. 2.3. Automated detections Initial investigations using a detection method based on spectral analysis revealed 2 types of seizure that were not detected by that method. Some clear seizures consisting of sometimes arrhythmic runs of high amplitude spikes were often undetected (Fig. 1, also Fig. 5-37 of Stockard-Pope et al., 1992). Similarly, some seizures with very low frequency discharges (<0.5 Hz) were noted among patients with severely depressed EEGs and were also undetected by the initial method (Fig. 2, also Fig. 5-32 of Stockard-Pope et al., 1992). In an attempt to capture all seizure types, 3 detection methods were therefore used. 2.3.1. Method 1: Detection of rhythmic discharges (0. 5 - 10
nz) This method employed spectral analysis of sequential 10s epochs of the EEG in a sliding window. This window moved along the EEG in 2.5-s steps, thus adjacent epochs overlapped by 75%. This overlap ensured that a 10-s seizure would be almost completely represented in one epoch at least. The algorithm was designed to extract features from each epoch and compare them with the features from 2 previous epochs (the 'background') that were 60 s behind F3-C3 ~ C3-P3
~
~
~"'-'~,~-'--~
~ -="~-'~'~,"'-~~
P3-01 ~ . - , . v - ~ . - - ~ - ~ ' ~ ~ Ca~o21Mat 23:19
I
5 seG
lOO~V
I
Fig. 1. Seizure identified visually by the EEGer and consisting of a run of arrhythmic spikes that were not detected by the spectral analysis method.
358
J. Gotman et al. /Electroencephalography and clinical Neurophysiology 103 (1997) 356-362 •
Fp2-F8 ~
T6-02
"
~
L/""-~
~-'~'
~/,~~
~f ~
~
We computed the power in the full width half-maximum band of the dominant peak and compared it to the power in the same frequency band in the spectrum of the background (Fig. 4). This is to determine if the rhythmic activity that may be present in the current epoch is new compared to the background.
~ 200/~V
sire124Mar10:29 5 sec
Fig. 2. Seizure identified visually by the EEGer and consisting of a very low frequency discharge that was not detected by the spectral analysis method. the current epoch. A constantly updated background was selected to reflect the visual interpretation process; it also has the great advantage over a fixed background that the results are not dependent on the choice of that fixed background. A long gap between background and current epoch was selected to allow the detection of seizures that start very gradually (Fig. 3). The frequency spectrum was computed for each epoch (the FFF length was 2048 points, so that the epoch length was in fact 10.24 s). The spectrum had a frequency range from 0 to 100 Hz, with a resolution of 0.1 Hz. The following variables were computed for each epoch: •
Dominant frequency
Among all the peaks in a spectrum, the peak with the largest average power in its full width half maximum band was defined as the dominant peak. The dominant frequency was defined as the frequency of the dominant peak (Qu and Gotman, 1997). The purpose of this feature was to find the most prominent rhythmic component of the epoch. •
Width of the dominant spectral peak
It was defined as the difference between the frequency corresponding to half the amplitude of the dominant peak in its falling slope and the frequency corresponding to half of this amplitude in the rising slope. If this width is small, the dominant peak represents genuinely 'rhythmic' activity.
Fp2-F8 F8-T6 T6-02 ~ - ;
Power ratio
~ , J \/~/"
....
t
•
Stability of current epoch
We integrated the average amplitude of the signal in each of the four 2.5-s sections of a 10-s epoch, and determined the ratio of the largest of the 4 values to the smallest one. We wanted to include as possible seizure candidates only sections that were relatively stationary, i.e. where this ratio was not too large (indicating that the amplitude of the smallest section was not very different from the amplitude of the largest section). 60-Hz Indicator
•
Poor electrode contact is frequent during long-term monitoring. It is important to detect it since it can lead to spurious slow patterns in the spectrum that can be mistaken for seizures. Although we only looked at frequencies up to 50 Hz in the spectrum, we were able to detect poor electrode contact by measuring activity at 20 Hz, where a sharp peak is clearly present whenever 60 Hz activity is very large. We presume that this peak comes from aliasing of the 3rd harmonic of 60 Hz (180 Hz). Although this activity should be eliminated by our anti-aliasing filter, it is possible that it can sometimes be sufficiently large to remain noticeable. It is also possible that this is a sub-harmonic of 60 Hz. Whatever its origin, it was a useful marker of poor electrode contact. We calculated the ratio of the peak of the spectrum at 20 Hz to the total energy in the spectrum. •
Patient disconnected indicator
Detection and updating of the background was suspended in any channel if the total energy in the spectrum was extremely low. This happens in particular if the patient or one amplifier is disconnected. It is important to eliminate such epochs from the background because they can lead to extremely high power ratios (see above definition) and thus to false Table 1 Combinations of values of dominant frequency, width of dominant peak and power ratio are acceptable for seizure detection in Method 1
e~a 22Maroe:ae
I 200/~V I
5 sec
Fig. 3. Seizure with a very gradual onset, common in newborn seizures. It has an initial low frequencydischarge starting about 40 s before the higher frequency discharge. Seizure first detected in the 10-s epoch ending at the arrow, and subsequently detected 41 times.
Dominant frequency First combination 0.5-1.5 Second combination 1.5-10 Third combination 1.5-10
Width of Power ratio dominantpeak _<0.6 _<0.6 _<1
3-4 2-4 4-80
359
J. Gotman et al. I Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
C
A Fp2-T4 T3-C3 ~
c3-cz ~ ~ ~ ' - - ~ ¢ ' - - ~
20t 1.5 1.0
O.
o.s/,I
-, - " - ~ - - - > ~ . - - ~
ol:
o
i
2
3
4
;
~
7
.
..~ ......~_
f--
z/'/,4
,
,
, ~'-~_
1
2
3
4
o
----_ 6
s
r
Freq(Hz)
Freq (Hz)
E I
Background epochs 20 sec
I
I
Gap 60 sec
I Current epoch lOsec
Fig. 4. Schematic representation of the spectral analysis method. (A) EEG background for comparative analysis. (B) Spectrum of channel T3-C3 for the section shown in A. (C) EEG in the current epoch. (D) Spectrum of channel T3-C3 for the section shown in C. (E) Representation of the relative timing of current epoch and comparative background. The background consists of 2 epochs ending 60 s prior to the current epoch. Dominant frequency and width of dominant spectral peak are calctLlated on spectrum in D. The power ratio is then defined as the ratio of the area hatched in D to the area hatched in B; the location on the frequency axis of the area in B is the same as that in D.
detections. After examination of the training set, the following thresholds were selected for the above variables. The first 3 parameters define a 3-.dimensional space. We defined boundaries in that space that determine where detections belong. These boundaries are given in Table 1. If an epoch was acceptable according to these 3 parameters, then it was considered a de,tection if its Stability was inferior to 3, its 60-Hz indicator inferior to 0.8%, and its Patient disconnect indicator superior to 0.
2.3.2. Method 2: Detection of multiple spikes This method is based on the spike detection procedure of Gotman et al. (1979). The procedure was modified because spikes in newborns tend to be of much longer duration than those in adults. We therefore performed high-pass digital filtering of the EEG prior to spike detection, using an IIR filter of order 3 with a cut-off frequency of 2 Hz. Fig. 5 shows the effect of this filtering on the background EEG and on spikes. A seizure detection was made when 6 or more spikes were detected in a 10-s epoch. 2.3.3. Method 3: Detection of very slow rhythmic discharges Method 1 described above often failed to detect very slow activity (around 0.5 Hz), particularly if the waveforms were complex. When the waveform is particularly far from sinusoidal, it is frequent that the peak of the spectrum is quite broad and that a high hm'monic content is present. This renders difficult the detection of the dominant peak, as defined above. We therefore decided to filter the EEG with a low-pass digital filter of order 2 and with a cut-off frequency of 1 Hz. The effect of such a filter on a very slow
seizure discharge is shown on Fig. 6. After this filtering, we applied the seizure detection method of Gotman (1990), but used the same epoch duration, gap and background duration as used for Method 1 above.
2.4. Visual seizure identification and comparison with automatic detection Prior to automated detection, electrographic seizures were identified visually by the EEGer. An electrographic seizure was defined as a burst of paroxysmal rhythms that were considered to be of cerebral origin and had a duration of 10 s or more and evolved temporally and spatially (Bye
A c3-P3 ~
......
P3-O~
FB3-C3 C3-P3 -
'
P3-01 ~
¢ "
Cap 21 Mar
' "
-
" "
'
~ '
-
~ =
21:34
I IO0/,IV 1 seo
J
Fig. 5. Example of newborn seizure consisting of runs of arrhythmic spikes (sometimes not detected by the spectral analysis method). Spikes in the newborn EEG, however, are relatively slow, and are not detected with the normal spike detection method. (A) Original seizure with relatively long duration of spikes. (B) Seizure filtered with a high pass filter: infinite impulse response (IIR), 3rd order, cutoff 2 Hz. After filtering, the spikes are briefer, sharper, and easier to detect by standard spike detection method.
360
J. Gotman et al. /Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
3. Results
A F4-P4
-
~
~
_
~
'~---~ I ~-,j
~
P4-o2~ - - - - - - - ~ J " V ~ ~/.,~--_.__. ~ B F4-P4 P4-02
~
'
~
-
-
~
Sam24 Mat 10:10
I 200/~V 2 SOC
Fig. 6. Example of newborn seizure consisting of a very slow frequency discharge. The waveforms are often complex and such seizures are sometimes not detected by the spectral analysis method. (A) Original seizure. (B) Seizure after application of a low pass filter: IIR, 2nd order, cutoff 1 Hz.
and Flanagan, 1995; Clancy and Legido, 1987; Scher et al., 1993). When several seizures follow each other closely, it is possible that the second or third seizure in a group is not detected due to our definition of 'background': if a seizure is itself the 'background' for the next seizure, this latter seizure may not be detected because it is then not different from the 'background'. Since the purpose of the detection method is to alert to the presence of seizures, we assume that if one seizure of a group is detected, the user will most often look before and after the seizure and will therefore see possible neighboring seizures. W e therefore defined the concept of 'cluster of seizures' as follows: a set of 2 or more seizures for which the interval between the end o f one is less than 90 s from the beginning of the next one. All EEGs were reviewed, and start and end time for all seizures were documented. Automated detections that occurred within 30 s of each other were grouped as a single detection. Automatic detections were then compared to visual identification: if an automatic detection occurred any time between the beginning and end of a seizure, the detection was considered as a valid seizure detection. Similarly, if a seizure within a cluster of seizures was detected, there was a valid cluster detection. All other detections were counted as false detections (FDs).
Table 2
W e will first give some examples of the seizures detected by each method, as well as of the false detections specific to each detection method. These are of course only examples and it is impossible to give a comprehensive view of the variety of seizure patterns and false detections. W e will then give results of each method on the data set described above. A detailed analysis of the results by institution, technical quality of recording, type and duration of seizure, for the training and testing data sets, is given in the accompanying publication (Gotman et al. (1997), accompanying paper). Since there is a large variability in the number of seizures per patient (the range is from 1 to 70 seizures and 1 to 49 clusters), we assessed the results by calculating for each patient the percentage of seizures detected and then averaging these percentages over all patients. If we just counted the percentage of seizures detected, then patients with a large number of seizures would completely overshadow those with one or 2 seizures. Since the type of seizure varies greatly from patient to patient, it is important to have all seizure types evenly represented. A total of 281.5 h of E E G from 55 patients were analyzed. Clinical assessment identified 679 seizures, and 228 of those occurred within 90 s of another seizure and so 451 'seizure clusters' were identified. 3.1. Results o f a u t o m a t e d detection
Table 2 reveals the average percentage of seizures and seizure clusters detected by each o f the 3 methods. While Method 1 detected the highest average percentage of seizures, the other 2 methods contributed to the overall detection success. Two detection methods sometimes detected the same seizure: for example a seizure detected by Method 1 may slow down at the end of its evolution and then be detected by Method 3. In this way the detection methods could sometimes reinforce each other. False detection rates were similar for each detection method, however each method tended to falsely detect differing events. As a result, the combined false detection rate is higher. Method 1 occasionally detected movement artifact that occurred when the mother was nursing the baby, and possibly rocking or patting it (Fig. 7). Method 2 occasionally detected excessive electrode noise, along with high
T~-o~~ ~ ~ , , ~ ¢ ~ ~ ~ ' ~ .
Results of seizure detection on the testing set Method 1 Method 2 Method 3 Methods 1-3 combined Average % seizures detected Average % clusters detected Average FD/hour
64
35
13
71
71
40
15
78
0.8
0.8
0.2
1.7
°
5 sec
I
Fig. 7. False detection occurring in the 10-s epoch ending at the arrow. This event was detected by the spectral analysis method in channels T6-O2 and P4-O2 and was caused by the mother patting the baby while nursing (visible on the video recording).
J. Gotman et aL /Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
361
voltage EMG artifacts, and sucking artifacts (Fig. 8). Method 3 also tended to occasionally detect slow movement artifacts, such as when the baby was moving or distressed (Fig. 9).
I200~V
Aus 20 Apt 13:40
4. Discussion
5 SOC
The newborn EEG does not contain the sustained rhythms that are characteristic of the EEG of older children, such as alpha, mu, central rhythms and the so called posterior slow waves of youth. Newborn seizures, however, are often rhythmic and this is most often their defining characteristic. An awareness of these features leads to the realization that spectral analysis may provide a means of seizure detection in the newborn. We designed an algorithm that extracted important features about rhythmicity, power and stability from the spectrum of a current epoch and compared those features with those of an earlier epoch in the background. In this way we were able to identify clear periods of paroxysmal rhythmicity (and therefore likely seizures). Some of the principles of this method are quite similar to the seizure detection method for adults developed by Gotman (Gotman, 1982; Gotman, 1990). Important differences include a much longer detection epoch (10 instead of 2 s) and a much longer background, reflecting the very low frequencies of the newborn EEG, as well as a spectral analysis method rather than wave decomposition in oTder to detect rhythmic discharges even in the presence of complex waveforms. Similar principles were used by Lui et al. (1992), who used autocorrelation rather than spectra to detect rhythmicity. We found that our spectral analysis method identified a large proportion of seizures, but did miss some events that were clearly seizures, but were either of very low frequency (<1 Hz) or runs of arrhythmic spike,;. For this reason we introduce 2 separate methods, specifically designed to identify both of these types of events. While these methods were successful, they identified relatively few seizures. Our experience leads us to believe the reason for this is that while very low frequency seizures and arrhythmic spike seizures are very easily identified visually, they are relatively uncommon among the newborn. Furthermore, each detection method also contributed a good number of false detections. As long as they are not too frequent, false detections do not represent a significant burden for a seizure detection method because they only represent extra EEG that has to Fp2-Fa~ . ~ ~ f . - - ~ , , ~ , ~ ~
~
T6-O2 ~
,
/
~
.
,
~
-
~
12oo~
Ang 23 Mar 1~42 5 see
Fig. 8. False detection occurring in the ]0-s epoch ending at the arrow. Tl~s event was detected in channels Fp2-F8 by the multiple spike detection method and was caused by a sucking artifact while the newborn was being fed.
Fig. 9. False detection occurring in the 10-s epoch ending at the arrow. This event was detected in channels F4-Cz and F3-P3 by the method used to identify slow rhythmic activity. It was caused by movement of the newborn while distressed and crying.
be reviewed. A false detection rate of 2/h only represents 40-60 s of EEG for each hour of recording that has to be unnecessarily reviewed. It is much more important not to miss seizures, than it is have a low false detection rate. Nonetheless, newborns in an intensive care situation have the potential of having a large variety and number of events that are clearly not seizures, but fulfill our detection criteria. For example, critically ill newborn often undergo cranial ultrasounds which can cause a number of electrographic artifacts. Less ill patients may be gently rocked or patted if distressed, and some newborn may even be held and fed by the mother. All these events can cause different false detections (for example the high EMG content of sucking may trigger the spike detection method). The method of Qu and Gotman (1993) allowed an important reduction in false detection by learning about the false detections particular to a patient during the first day of monitoring and then avoiding similar events in subsequent days. This principle is less applicable to newborn monitoring, which usually does not extend over more than a couple of days. The 3 methods used here were not combined to run simultaneously in real time in this study, but the methods were designed so that it is be possible and the combined method could run on line on a standard processor (we estimate that a 100-MHz PC would be sufficient). This raises the question as to whether the gain achieved in broadening the types of seizure detected by using 3 separate methods is worth the cost of an increase in false detections.
Acknowledgements We are grateful to Carol Leitner and the other staff of the EEG Laboratory of the Montreal Children's Hospital for their collaboration. We are also grateful to Dr. E. Mizrahi, Baylor College of Medicine and Dr. A. Bye, Sydney Children' s Hospital, for providing many of the recordings used in this study. This work was supported in part by a grant from the Medical Research Council of Canada and Stellate Systems through the University-Industry program of the Medical Research Council (Grant MRC-UI-11514).
362
J. Gotman et al. /Electroencephalography and clinical Neurophysiology 103 (1997) 356-362
References Bye, A.M.E. and Flanagan, D. Spatial and temporal characteristics of neonatal seizures. Epilepsia 1995; 36: 1009-1016. Clancy, R.R. and Legido, A. The exact ictal and interictal duration of electroencephalographic neonatal seizures. Epilepsia 1987; 28: 537541. Dreyfus-Brisac, C. and Monod, N. Electroclinical studies of status epilepticus and convulsions in the newborn. In: P. Kellaway and I. Petersen (Eds.), Neurological and Electroencephalographic Correlative Studies in Infancy. New York: Grune and Stratton, 1964, pp. 250-272. Fenichel, G.M. and Fitzpatrick, J.E. Difficulty in clinical identification of neonatal seizures: an EEG monitoring study. Electroenceph. clin. Neurophysiol.1984; 58: 33. Gotman, J. Automatic seizure detection: Improvements and evaluation. Electroenceph. clin. Neurophysiol. 1990; 76: 317-324. Gotman, J. Automatic recognition of epileptic seizures in the EEG. Electroenceph, clin. Neurophysiol. 1982; 54: 530-540. Gotman, J., Ives, J.R. and Gloor, P. Automatic recognition of iuterictal epileptic activity in prolonged EEG recordings. Electroenceph. clin. Neurophysiol. 1979; 46:.5 ! 0-520. Gotman, J., Flanagan, D., Rosenblatt, B., Bye, A. and Mizrahi, E.M. Evaluation of an automatic seizure detection method for the newbom. Electroenceph. clin. Neurophysiol. 1997; 103: 363-369. Lui, A., Hahn, J.S., Heldt, G.P. and Coen, R.W. Detection of neonatal seizures through computerized EEG analysis. Electroenceph. clin. Neurophysiol. 1992; 82: 30-37. Mizrahi, E.M. Neonatal electroencephalography. Clinical features of the newborn, techniques of recording and characteristics of the normal EEG. Am. J. EEG Technol. 1986; 26: 81-103.
Mizrahi, E.M. and Kellaway, P. Clinical, electroencephalographic, therapeutic, and pathophysiologic studies of neonatal seizures. In: C.G. Wasterlain and P. Vert (Eds.), Neonatal Seizures. Raven Press, New York, 1990, pp. 1-12. Mizrahi, E.M. and Kellaway, P. Characterization and classification of neonatal seizures. Neurology 1987; 37: 1837-1844. Mizrahi, E.M. and Kellaway, P. Characterization of seizures in neonates and young infants by time-synchronized EEG polygraphic/video monitoring [abstract]. Ann. Neurol. 1984; 16: 383. Qu, H. and Gotman, J. A self-adapting algorithm for the detection of seizure onset in long-term EEG monitoring: possible use as a warning device. IEEE Trans. Biomed. Eng. 1997; 00: in press. Qu, H. and Gotman, J. Improvement in seizure detection performance by automatic adaptation to the EEG of each patient. Electroenceph. clin. Neurophysiol. 1993; 86: 79-87. Rose, A.L. and Lombroso, C.T. Neonatal seizure states. A study of clinical, pathological, and electroencephalograpic features in 137 full-term babies with a long-term follow-up. Pediatrics 1970, 45: 404-425. Scher, M.S., Hamid, M.Y., Steppe, D.A., Beggarly, M.E. and Painter, M.J. Ictal and interictal electrographic seizure durations in preterm and term neonates. Epilepsia 1993; 34: 284-288. Scher, M.S., Sun, M., Steppe, D.A., Guthrie, R.D. and Sclabassi, R.J. Comparison of EEG spectral and correlation measures between healthy term and preterm infants. Pediatr. Neurol. 1994; 10: 104-108. Stockard-Pope, J.E., Werner, S.S. and Bickford, R.G. Atlas of Neonatal Electroencephalography. New York: Raven Press, 1992. Tharp, B.R. and Laboyrie, P.M. The incidence of EEG abnormalities and outcome in infants paralyzed with neuromuscular blocking agents. Crit. Care Med. 1983; 11: 926-929. Volpe, J.J. Neonatal seizures. In: Neurology of the Newborn. W.B. Saunders, Philadelphia, 1995, pp. 172-207.