Evaluation of an automatic seizure detection method for the newborn EEG

ELSEVIER

Electroencephalographyand clinical Neurophysiology 103 (1997) 363-369

Evaluation of an automatic seizure detection method for the newborn EEG J. G o t m a n a'*, D . F l a n a g a n b, B . R o s e n b l a t t b, A . B y e c, E . M . M i z r a h i d aThe Montreal Neurological Institute and Hospital and McGill University, 3801 University Street, Montreal PQ, H3A 2B4, Canada bMontreal Children's Hospital, Montreal, Canada CSydney Children's Hospital, Sydney, Australia dTexas Children's Hospital, and Baylor College of Medicine, Houston, TX, USA

Accepted for publication: 17 February 1997

Abstract In another publication, we described a set of methods for automatic detection of EEG seizures in the newborn. We describe here the evaluation of these methods using a completely new set of data, which were not used in developing the method. This testing data set consisted of recordings from 54 patients, lasting an average of 4.4 h. Recordings had 8-16 channels and were obtained, in approximately equal numbers, from 3 institutSions in Canada, the USA and Australia. Recording conditions varied from short recordings fully attended by a technologist to overnight recordings largely unattended. The average seizure detection rate was 69% (77%, 53%, 84% in the 3 institutions). False detections occurred at the average rate of 2.3/h (4.1, 1.0, 2.7 in the 3 institutions), with fluctuations that reflected largely the technical quality and level of supervision of the recordings. The results are similar to those obtained in the commonly used method of epilepsy monitoring in adults and allow us to envisage clinical application. © 1997 Elsevier Science Ireland Ltd. Keywords: EEG; Seizure; Detection; Newborn; Evaluation

1. Introduction Before an automated de,tection method is proposed for use in a clinical setting, it is essential that the method be evaluated with real data. Unfortunately this is an expensive and time consuming exercise that is rarely c a r d e d out with acceptable objectivity. Often data used in such evaluations are so severely pre-selected that they do not reflect realistic clinical conditions. It is c o m m o n that the data used may be selected to include only 'prototypical' events and records with little artefact. Sometimes 'synthetic' data are used (i.e. data generated by signal processing methods to look like the real data). Further, some evaluations use the same data set for 'training' and 'testing' the method. The results of these approaches can be highly unrealistic. If a detection method is to be applied in a clinical setting, accurate evaluation is critical. Several basic criteria should be met:

* Corresponding author. Tel.: +l 514 3981953; fax: +1 514 3988106; e-mail: jean @rclvax.medcor.mcg:Lll.ca.

0013-4694/97/$17.00 © 1997 Elsevier Science Ireland Ltd. All rights reserved PII S0013-4694(97)00005-9

•

• • •

There should be 2 data sets, one for training the method and a separate data set for testing the method. If the method is trained automatically, then other partitions of the testing and training data may be considered (Devijver and Kittle, 1982) As large a sample size as is practical should be obtained for both training and testing data. Data should not be pre-selected with respect to 'typicality' or 'quality'. Data from a number o f sources should be used so that methods are not tailored to the features of a single population.

Such an approach may produce lower estimates of detection rates when compared to estimates obtained from preselected data, but the results more accurately reflect clinical realities. In a previous paper we described a method of seizure detection in the newborn (Gotman et al. (1997), accompanying paper). In developing that method we used unselected data from three separate medical centers. In the present paper we provide an extensive evaluation of the

EEG 96178

364

J. Gotman et al. /Electroencephalography and clinical Neurophysiology 103 (1997) 363-369

performance of this method. We have attempted to be as unbiased as possible in data selection and to cover a wide variety of recording conditions so that the results are likely to be representative of real clinical conditions. Although many methods of automatic spike and seizure detection have been developed, these have been specifically designed for the EEGs of adults and older children and few have been extensively validated. Detection of electrographic seizures in the newborn requires different techniques from those employed in adults. To date, only one attempt has been made to address this difficult task in the newborn (Lui et al., 1992). Those authors used EEGs from 14 infants recorded during the first week of life. A total of 58, 30-s 'prototypical seizure' sections and 59, 30-s 'nonseizure' sections were analyzed. The method resulted in a sensitivity of 84% and a specificity of 98%. These results are promising, however, they should be viewed with caution because the method of data selection and evaluation do not fulfill the criteria outlined above. In the present paper we provide a detailed evaluation of our newborn seizure detection method. We believe that no other method of seizure detection has been validated so extensively.

2. Subjects

paper). Recording conditions for the testing data were essentially the same. They are summarized below. 2.3.1. Montreal Children's Hospital We obtained nine recordings lasting an average 5.5 h (range: 0.3-19.5 h). These recordings were most often prolonged recordings made in the Intensive Care Unit with infrequent supervision by the EEG technician. 2.3.2. Sydney Children's Hospital We obtained 25 recordings lasting an average 6 h (range: 5.3-6.8 h). These recordings were most often prolonged recordings made in the intensive care unit with infrequent supervision by the EEG technician. 2.3.3. Texas Children's Hospital We obtained 20 recordings lasting an average 1.8 h (range: 1-2.6 h). These recordings were most often 2-h recordings with continuous supervision by the EEG technician. The testing data set therefore included 54 recordings lasting a total of 237 h, for an average recording duration of 4.4 h. The training data set included a total of 55 recordings lasting a total of 281.5 h, for an average recording duration of 5.1 h. There is no reason to believe these 2 data sets differ in any systematic way.

2.1. Training data set

3. Methods This information has been provided in detail elsewhere (Gotman et al. (1997), accompanying paper). In summary, EEGs were obtained from a total of 55 newborns studied at 3 centers, Montreal Children's Hospital (MCH), Sydney Children's Hospital (SCH) and Texas Children's Hospital (TCH). Consecutive recordings were used and there were no selection criteria other than age.

3.1. Visual seizure identification

Consecutive recordings were used. EEGs containing electrographic seizures were obtained from 9 newborns studied at Montreal Children's Hospital, 14 newborns studied at Sydney Children's Hospital and 18 newborns studied at Texas Children's Hospital. The average postconceptional age at time of study was 39 weeks (range: 28-46 weeks). There were no selection criteria other than age. No records were rejected (regardless of type of seizure or record quality). Thirteen recordings from patients without electrographic seizures were also obtained. No records were rejected (regardless of record quality).

For the purpose of visual assessment by an electroencephalographer (EEGer), an electrographic seizure was defined as a burst of paroxysmal rhythms that was suspected to be of cerebral origin and had a duration of 10 s or more and evolved temporally and spatially (Clancy and Legido, 1987; Hrachouy et al., 1990; Scher et al., 1993; Bye and Flanagan, 1995). For consistency only one EEGer (D.F.) assessed all the raw data. All EEGs were reviewed and start and end time for all seizures were documented. Events shorter than 10 s were not accepted as seizures. Events that the EEGer recognized as electrographic seizures but that were reduced in some aspects (low amplitude, and/or limited electrographic spread, and/or difficult to discern above the surrounding EEG because of high voltage artefact) were noted as being 'minor' seizures. Seizures occurring in close succession (within 90 s of each other) were noted as representing a 'cluster' of seizures (Gotman et al. (1997), accompanying paper).

2.3. Centers

3.2. Record quality

Recording conditions for the training data have been described previously (Gotman et al. (1997), accompanying

Recording quality was assessed by visually estimating the amount of non cerebral activity present in a record: elec-

2.2. Testing data set

J. Gotman et al. /Electroencephalography and clinical Neurophysiology 103 (1997) 363-369 trode noise, movement ~xtefact (patient or externally induced), muscle artefact, ECG artefact. We used the following definition in grading quality: •

Good record quality

Recording quality was defined as good if non-cerebral waveforms were limited to less than 10% of the duration of the total record, independently of how many channels were involved. •

•

3.4.2. Analysis of false detections All false detections were reviewed and classified into one of the following categories: •

•

Marginal record quality

Recording quality was defined as marginal if more than 10% of the duration of the record had some waveforms of non-cerebral origin of any one type (including electrode noise lower than 20 #V).

•

Poor record quality

Recording quality wa:~ defined as poor if more than 10% of the duration of the record had waveforms of non-cerebral origin of two or 'more types, or more than 10% of the record had electrode noise larger than 20 /zV. Non cerebral artefactual activity could be limited to one channel or extend over many channels. It is obvious that this visual assessment of quality is somewhat empirical. It nevertheless provides a general indication of quality. 3.3. Automated detection methods A complete description of the automated detection methods has been given previously (Gotman et al. (1997), accompanying paper). Briefly, 3 methods were used. The first employed spectral analysis to detect paroxysmal rhythmic activity between 0.5 ~xld 10 Hz. To detect arrhythmic runs of spikes a modification of the Gotman et al. (1979) spike detection method was used. To detect seizures having a very low discharge frequency, the EEG was low-pass filtered with a cut-off frequency of 1 Hz. After this filtering, we applied the seizure dete,ction method of Gotman (1990), modified with respect to epoch length. 3.4. Analysis of automated detection results Automatic detections that occurred within 30 s of each other were grouped as a single detection. 3.4.1. Identification of good detections For each record, the results of automated detections were compared to the visual clinical assessment of start and end time of seizures. Any automatic detection that overlapped with the start and end time of a seizure was defined as a 'seizure detection'. If any ,ane seizure within a seizure cluster was detected, then we considered that the seizure cluster was detected. All other detections were defined as false detections.

365

•

'No abnormality' referred to false detections occurring at a time when there were no clear EEG abnormalities and no clear artefact. These usually corresponded to a normal paroxysmal pattern. 'New seizure' referred to detections that identified a genuine seizure that had been missed by the EEGer during the original visual assessment. 'Interesting detection' referred to abnormal EEG waveforms. These may include brief (less than 10 s) ictal phenomena and paroxysmal bursts in records with burst-suppression. It should also be noted that this category may include some events which may be artefactual in nature, but for which there were insufficient data for definite clarification (e.g. suspected pulse artefacts and patting artefacts). 'False detection' refer to detections caused by clear non cerebral events (includes movement, ECG, sucking, muscle and artefact caused by electrode problems).

4. Results

Overall results for the training set have been given in Gotman et al., 1997 (accompanying paper). The results for the testing set will be presented in 3 sections: (1) overall results for the testing set; (2) breakdown according to the institution from which EEGs were recorded, since this is a major factor of variability; (3) analysis of the false detections, since it is important to separate detections of genuine paroxysmal EEG patterns from detections of artefacts or other events of no interest. Since there is a large variability in the number of seizures per patient (the range is from 1 to 54 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. False detections per hour were calculated for each patient and for each center; average false detections per hour were then calculated from those data. 4.1. Results of the testing set The average detection rates for the testing and training data sets are presented in Table 1. As expected, the overall performance was poorer in the testing set compared to the training set. The average detection rate for the testing data

J. Gotman et al. /Electroencephalography and clinical Neurophysiology 103 (1997) 363-369

366

Table 1 Average detection rates in the testing and training data sets for seizures, seizure clusters and false detections Testing data Average percent seizures detected Average percent seizure clusters detected Average false detections/hour

FST0

sam2~U~ 10:2~

Training data

66 69

71 78

2.3

5 sec

11ooj4v

B Fpl-T3

~

Fp2-T4 T3-C3

1.7

Wilca 48126

I lOOvV

5 see

was 66% for seizure detections and 69% for seizure cluster detections. In the training set the average seizure detection rate was 71% and the average seizure cluster detection rate was 78%. The false detection rate for the testing data was 2.3/h, and 1.7/h in the training set. We also noted that some seizures and some artefacts were very 'patient specific', and so there was significant variation in detection rates between individuals (for both seizures and false detections). This contributed to the relatively high standard deviations in the data. This in itself may be a feature of the newborn EEG. Among patients with seizures, we determined that at least one genuine seizure was detected in 85% of patients in the testing set and 97% in the training set. Seizures which were reduced in some of their features and classified as 'minor' by the EEGer represented 16% of the total number of seizures. The average detection rate for 'minor' seizures was 29%, whereas the average detection rate for easily recognized seizures (the seizures not considered 'minor', and labeled 'clear' below) was 72%. Fig. 1 shows examples of two 'clear' seizures that were detected, and Fig. 2 illustrates two 'clear' seizures that were missed; the seizure of Fig. 2A was missed probably because the discharge was so slow (1/3 Hz); that of Fig. 2B was missed because it did not include a 10-s section of rhythmic activity nor a 10-s section with a sufficient number of

Fig. 2. Two 'clear' seizures that were missed. (A) This seizure was missed because the discharge was very slow (1/3 Hz). (B) This seizure did not include 10 s of stable rhythmicity nor 10 s with a sufficient number of spikes.

spikes. Fig. 3 shows 2 'minor' seizures that were detected and Fig. 4 2 'minor' seizures that were missed, probably because there was no single, sustained, dominant rhythm in either. 4.2. Results per institution The detailed training and testing results for each of the 3 institutions are presented in Table 2. Some variation was noted between results from different medical centers and also between testing and training results, but no clear trends were identified. For all institutions, the proportion of patients for whom at least one seizure was recorded remained relatively high. False detections were lowest in the Texas data. This is not surprising since recordings there only lasted 2 h and babies were in a crib and under the constant supervision of a technologist. False detections were high in the Sydney and Montreal data, where recordings were prolonged and were infrequently attended by the technologist. Furthermore, nursing and feeding b y the mother were allowed during some recordings in these institutions.

A

A

P3-01 Fpl-F7 ~ = = ~ . ; ~ F7-T5 ~ T5-01

T5-01

Law

.

.

.

~ A~"23:13

.

.

.

.

t

_ _ ~

_

5 see J=. 09A~ 9:~

B

t

/

s.~

200pV

J

B Fpl-T3 Fp2-T4 T3-C3 C3-Cz

F4-C4 C4-P4 P 4 - 0 2 ~

c32 9:29 i

t

200/AV

1

200tJv

5 sec

5 se¢:

Fig. 1. Two 'clear' seizures that were detected. (A) This seizure has a gradual onset, and was first detected in the 10-s epoch ending at the arrow. (B) This brief, focal seizure was first detected in the 10-s epoch ending at

Fig. 3. Two 'minor' seizures that were detected. (A) This seizure, which was only recorded in the left occipital electrode, was first detected in the 10-s epoch ending at the arrow. (B) This brief frontal seizure was detected

the arrow.

in the 10-s epoch ending at the arrow.

J. Gotman et al.

/Electroencephalography and clinical Neurophysiology 103 (1997) 363-369

367

Table 2 Results from the testing and tradning data set from each institution

Testing data Patients with seizures n Average % of seizures detected Average % seizure clusters detected Number of patients where > 1 seizure was detected Average FDs/hour Patients without seizures n Average FDs/hour Combined FDs/hour

Training data

MCH 9 74.8 77.4 9

SCH 14 81.1 83.9 13

TCH 18 49.3 53.3 13

MCH 11 72.5 83.2 11

SCH 15 63.8 73.3 15

TCH 17 76.2 80 16

4.1

3.2

1.1

2.6

2.0

0.3

11 2.1 2.7

2 0 1.0

0 2.6

10 2.7 2.2

2 3.1 0.6

0 4.1

4.3. Detection analysis

seizure detections, however, were due to physiological or electrode artefacts.

False detection rate was analyzed as a function of the technical quality of the recordings. For the testing set, the rate was 0.8/h for recordings of good quality, 2.9/h for recordings of marginal quality, and 4.9/h for recordings of poor quality. These results confirm that artefacts were the main cause of false detection, although the technical quality of the recordings was no1:judged on the basis of the actual false detections. To obtain an estimate of specificity of the detection method, all detections were analyzed in detail for the test data set for each center (Table 3). For this analysis the total number of detections for all patients from each center were pooled and analyzed. The majority of detections for each center were seizure detections. Between 5 and 18% of all detections were detections of 'interesting events' (such as brief ictal events or paroxysmal bursts of activity in burstsuppressed records). Two examples of 'interesting detections' are shown in Fig. 5. Three real seizures, missed by visual analysis, were also detected. The majority of all non A F4-Cz F

p

l

-

F

3

~

~

I I

Mar 08 A~ 14:37

200/~V

5 BO¢

B F2-04 Fz-Cz Cz-Pz

Ca9:05

_ _ J

IO0#V

Sse¢

Fig. 4. Two 'minor' seizure lJmt were missed. These seizures have no

single sustaineddominant rhythm.

5. Discussion We have presented an extensive evaluation of a method for seizure detection in the newborn. We selected an evaluation procedure in which the testing set is completely separate from the training set. It is established that such a procedure tends to give results that have a pessimistic bias, unbiased results being somewhere between those of the training set and those of the testing set (Devijver and Kittle, 1982). Furthermore, the data was not pre-selected in any way, thus the data reflects 'real' and not 'idealized' data. In total, 109 recordings obtained from three institutions have been used in the development and evaluation of the method. This is a large sample with a broad variety of seizures and false detections and thus provides a good indication of what the performance should be in an average clinical practice. The average seizure detection rate across patients for our testing data was 66% (it was 71% for the training data set). Lui et al. (1992) reported better results for their detection method, however, the data used in that study was preselected for 'typicality' and the same data was used to train and test the method. Such results, while encouraging, should be viewed with caution. Evaluations of seizure detections in adults have also yielded slightly better results than obtained here, however, the approaches used to evaluate the methods varied and direct comparison is difficult. In the validation of Gotman (1990), 76% of seizures were detected. In that study, however, seizures that were not detected and not reported by the patient or clinical staff were not included. In the validation made by Pauri et al. (1992) of the method of Gotman (1990), different detection thresholds were used, yielding different detection and false detection rates. With the lowest sensitivity, the detection rate was 48% and with the most sensitive settings, the rate was 81%. In intracranial electrode recordings in adults,

J. Gotman et al. / Electroencephalography and clinical Neurophysiology 103 (1997) 363-369

368 Table 3

Detection analysis (test data set) MCH

Seizure detection* New seizuret Interesting detection False detection No abnormality Total

SCH

TCH

Combined

n

%Tot detections

n

%Tot detections

n

%Tot detections

n

%Tot detections

120 0 36 21 22 199

60% 0% 18% 11% 11%

443 1 46 327 41 858

52% 0% 5% 38% 5%

99 2 8 24 3 136

73% 1% 6% 18% 2%

662 3 90 372 66 1193

55% 0% 8% 31% 6%

*Seizure detection refers to detections during those events predefined as seizures by the EEGer. tNew seizure refers to seizures missed during the initial visual assessment of the data.

Harding (1993), obtained an 86% seizure detection rate. In this case, however, some detection parameters were adjusted by the operator after the first seizure was detected in each patient. In the present investigation, however, our evaluation was rigorous. The complete record was reviewed, and seizures visually identified prior to automated detection. We also strictly adhered to an accepted definition of electrographic seizures in the newborn. As a result we identified an extraordinary variety of seizure types, including a number of 'minor' seizures (16% of all seizures identified by visual assessment). It is likely that our approach will yield low estimates of seizure detection, but these estimates should be valid for most clinical situations. Furthermore, seizures in the newborn often occur closely spaced in time, and this affects the results of the detection method. For this reason we introduced the concept of seizure clusters to deal with the runs of seizures often encountered. For our analysis, it is more practical to equate the seizure clusters noted in the newborn with single seizures in adults and older children, which are more often isolated in time. The detection rate for seizure clusters was 69%. Possibly of greater significance to the neonatal intensive

~ -C3 ,

v

-

~

Cap 21 Mar 2"[54

- - J l

10OFV

5 sec

B Fp2-F8

$ ...,..,=

Fp2-F4 ~

'

~ • d

"

*

,

=

..~

..,/'V,v.....,~,,.e~h .

~w,,,,v,,'r," , . . ~ ~ . ¥ . . . . . . .

F4-P4 P4-02

v.

.,~,.e.,'%lu.,d~.H~,,~wV

v ' . v ~ , " " • " ." - "

_ ,.,a.2t .~ .:~

I 200/N 5 SOC

F i g . 5. T w o

'interesting

detections'.

(A)

This nm

of abnormal

spikes did

not conform to our seizure definition, but was detected in the 10-s epoch ending in the arrow. (B) This brief frontal rhythm was detected in the 10-s epoch ending at the arrow and may be an example of 'anterior slow dysrhythmia'.

care clinician is the finding that at least one seizure was detected in 85% of those patients having seizures in the testing set (the rate was 97% in the training set). It is therefore unlikely that a false negative diagnosis will be made. Given the difficulty associated with identification of seizures in the newborn, this in itself would provide useful information, and the question of whether or not the patient is having seizures at all could be answered with some confidence with this automatic detection method. The false detection rate of 2.3/h is similar, though slightly higher than that obtained for adults. Gotman (1990), noted a false detection rate of just over 1/h and Pauri et al. (1992) had false detection rates of 2.7/h to 5.4/h. In intracranial electrode recordings in adults, Harding (1993), obtained a false detection rate of almost 2/h. A reason for the slightly higher false detection rate in the present study is that newborn have the potential for a wide variety of confusing artefacts. Many 'at risk' newborn undergo cranial ultrasounds which involve prolonged manipulation of the head. Furthermore, it is common for a newborn who is distressed to be comforted by gentle rocking or patting. A newborn who is not critically ill may even be nursed and fed by the mother. Such events can cause a number of artefacts, some of which can resemble seizures in their electrographic presentation. In the present investigation we noted that some false detection patterns were patient specific, and therefore false detection rates varied considerably between individuals. By analyzing the false detections as a function of recording quality, we established that technical quality was the most important factor in determining the rate of false detections. It is therefore possible to predict with some accuracy the rate of false detections in an environment where the quality is known (e.g. attended vs. unattended recordings). A false detection rate of about 2/h is acceptable. For example over 20 h, the average expected 40 false detections would only represent 20 extra minutes of EEG (assuming that 30 s need to be seen for each detection) aside from whatever true seizure detections occur. Thus while there would be some extra EEG to review, there would still be enormous data reduction for the 20 h of monitoring time.

J. C,otman et al. /Electroencephalography and clinical Neurophysiology 103 (1997) 363-369

It is notable that relatively few detections (6%) occurred when there were no clear electrographic abnormalities. This implies that the vast majority of detections were of some interest. Even the 31% of detections that represent clear physiological or electrical artefact could alert an observer to modify the recording environment, either by removing or remedying the source of e]Lectrical noise, or by temporarily suspending seizure detect:ion when the newborn is being handled (such as during a cranial ultrasound). There is great potential for a reliable automated seizure detection method. Since newborn seizures are notoriously difficult to identify clinically, and can occur sometime (hours to days) after the suspected insult, it would be ideal clinical practice to conduct prolonged EEG monitoring on 'at risk' newborn. Unforttmately this is prohibitively time consuming and expensive. An automated detection method provides a workable alternative. For example, one might normally do only a 2-h recording and review visually the whole recording, whereas an alternative could consist of making the same 2-h recc,rding, review it completely, but then continue monitoring for another 20-h and review only detections. It must be reme,mbered that there is a significant drawback in not recording: if no recording is made during the 20-h period mentioned above, because it is too difficult to record and review such a long recording, the probability of recording a seizure is 0%. One can also view automatic detection as a means of having a first, rapid look at the recording, if no time is available immediately for a complete review: rather than looking at a few samples selected randomly, one can review dellections and have a much higher probability of seeing events of clinical significance. This will not only increase the likelihood seizure detection, but will also provide immediate and important information about response to anticonvulsant treatment. As with seizure detection in adults, automated detection methods do not unequivocally identify seizures in the newborn, but rather identify 'se,izure like' electrographic events, some of which will always be artefacts. Thus it is imperative that an experienced EEGer reviews the detections before clinical decisions are made. In conclusion, we presenLted a method of seizure detection in the newborn and an evaluation that establishes the framework of its performance. Just as seizure detection in older children and adults has found its place in the broad context of the evaluation of patients with epilepsy, we hope that this work will open the door to the use of a similar procedure in

369

newborns, a group of patients in whom seizures are particularly frequent and particularly difficult to identify clinically.

Acknowledgements We are grateful to Ms. Carol Leitner and the other staff of the EEG Laboratory of the Montreal Children's Hospital and to Mr. Burton and Mr. McCully and the EEG Laboratory Staff at the Texas Children's Hospital for their collaboration. 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).

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. Devijver, P.A. and Kittle, J. Pattern Recognition: A Statistical Approach, Prentice-Hall, London, 1982. Gotman, J. Automatic seizure detection: improvements and evaluation. Electroenceph. clin. Neurophysiol. 1990; 76: 317-324. Gotman, J., Flanagan, D., Zhang, J. and Rosenblatt, B. Automatic seizure detection in the newborn: methods and initial evaluation. Electroenceph. clin. Neurophysiol. 1997; 00:000-000 [this issue]. Gotman, J., Ives, J.R. and Gloor, P. Automatic recognition of interictal epileptic activity in prolonged EEG recordings. Electroenceph. clin. Neurophysiol. 1979; 46: 510-520. Harding, G.W. An automated seizure monitoring system for patients with indwelling recording electrodes. Electroenceph. clin. Neurophysiol. 1993; 86: 428-437. Hrachouy, R.A., Mizrahi, E.M. and Kellaway, P. Electroencephalography of the newborn. In: D.D. Daly, T.A. Pedley (Eds.), Current Practice of Clinical Electroencephalography, 2nd edn., Raven Press, New York, 1990, pp. 201-242. 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. Pauri, F., Pierelli, F., Chatrain, G.-E. and Erdly, W.W. Long-term EEGvideo-audio monitoring: computer detection of focal EEG seizure patterns. Electroenceph. clin. Neurophysiol. 1992; 82: 1-9. 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.