Electroencepht~lography and clinical Neurophysiology, 82 (1992) 1-9
1
© 1992 Elsevier Scientific Publishers Ireland, Ltd. 0013-4649/92/$05.00
EEG 91142
L o n g - t e r m E E G - v i d e o - a u d i o m o n i t o r i n g : c o m p u t e r d e t e c t i o n o f focal E E G seizure patterns * Flavia Pauri a, Francesco Pierelli a, Gian-Emilio Chatrian a,b and William W. Erdly c '~ Section of EEG and Clinical Neurophysiology, Division of Neurology, Department of Medicine, h Department of Neurological Surgery, University of Washington, School of Medicine, and c Department of Psychology, University of Washington, School of Arts and Sciences, Seattle, WA (U.S.A.) (Accepted for publication: 16 September 1991)
Summary Twelve individuals with medically refractory partial seizures had undergone EEG-video-audio (EVA) monitoring over 1-15 (mean 10.5) days. We selectively reexamined available 15-channel EEGs (video-cassettes) totaling 461 h and containing 253 EEG focal seizures. Computer analysis (CA) of these bipolar records was performed using a mimetic method of seizure detection at 6 successive computer settings. We determined the computer parameters at which this method correctly detected a reasonably large percentage of seizures (81.42%) while generating an acceptable rate of false positive results (5.38/h). These parameters were adopted as the default setting for identifying focal EEG seizure patterns in all subsequent long-term bipolar scalp and sphenoidal recordings. Factors hindering or facilitating automatic seizure identification are discussed. It is concluded that on-line computer detection of focal EEG seizure patterns by this method offers a satisfactory alternative to and represents a distinct improvement over the extremely time consuming and fatiguing off-line fast visual review (FVR). Combining CA with seizure signaling (SS) by the patients and other observers increased the correct detections to 85.38~. CA is best used in conjunction with SS. Key words: EEG-video-audio monitoring; Focal EEG seizures; Partial seizures; Computer seizure detection; Visual seizure detection; Seizure signaling
Long-term EEG-video-audio (EVA) monitoring is an established method for the evaluation and treatment of selected patients suffering from intractable epileptic seizures (Gotman et al. 1985; Gumnit 1987). The success of this technique depends in large measure on the method followed to detect the patients' seizures. This may consist of one or more of the following: continuous or intermittent, direct or indirect, patient observation and E E G recording by trained personnel; seizure signaling (SS) by the patients and other observers; fast visual review (FVR) of all EEGs obtained during monitoring; and computer analysis (CA) of these EEGs. In our initial practice of EVA monitoring, we strived to recognize our patients' seizures as completely as possible by FVR of all their E E G records aided by SS by the patients and other observers (Pierelli et al. 1989). The present work reports on the results of
* Presented in part at the 5th European Congress of Clinical Neurophysiology, Paris, France, September 1990.
Correspondence to: Dr. G.E. Chatrian, Section of EEG and Clinical Neurophysiology, NN 282A, University of Washington Medical Center, RC-97, Seattle, WA 98195 (U.S.A.). Tel.: (206) 548-6465; Fax: (206) 548-4102.
CA of the EEGs of the same patients included in our previous investigation.
Patients and methods
We studied 12 patients who suffered from medically intractable partial seizures (Pierelli et al. 1989). All subjects had undergone 24 h EVA monitoring over 1-15 days (mean 10.5, S.D. 3.86, median 11) at the University of Washington Medical Center in Seattle between July 1986 and April 1988. These individuals ranged in age from 10 to 59 years (mean 29.1, S.D. 15.1, median 29.0). Eight were females and 4 males. Fifteen channels of E E G recorded bipolarly with scalp and, in 5 patients, sphenoidal electrodes and 1 channel of ECG were amplified and multiplexed (Telefactor TM 100-16 CTE), transmitted by cable telemetry to the E E G laboratory and stored on video cassette together with the video and the audio signals (Pierelli et al. 1989). At the time they were monitored, 216 focal seizure patterns had been detected on these individuals by combined FVR and SS. In the present investigation, we selectively reexamined those EEGs (video cassettes) that contained one or more focal seizures recognized earlier by FVR and SS (total: 461 h). These
2
F. P A U R I ET'AL.
records were: decoded (Telefactor TM 101-16D), lowpass filtered ( - 3 dB at 50 Hz) by a 6-pole switch capacitor filter (Biomedical Monitoring Systems MF32), digitized at the rate of 200 s a m p l e s / s e c / c h a n n e l by a 32-bit word length processor personal computer (Compaq Deskpro 386/20) and stored on a 130 Mb fixed disk drive with 135 Mb tape backup. The digitized EEGs were analyzed using a mimetic seizure detection method (Gotman 1982, 1985, 1986, 1990; Gotman et al. 1985) based on modifications of the the algorithm described by Gotman and Gloor (1976). The computer program (Stellate System Monitor 3.0) broke individual E E G waves into half-waves and determined the following 3 measures for each 2 sec E E G epoch: (1) the average amplitude of the half-waves relative to that of a sample of background EEG activity consisting of a 16 sec epoch ending 20 sec prior to the beginning of the current epoch. This parameter established whether or not the E E G epoch under scrutiny had "paroxysmal" features; (2) the average duration of the half-waves as a measure of frequency; and (3) the coefficient of variation of half-wave duration (S.D./mean) expressing degree of rhythmicity. A seizure was detected in any given channel when the computer recognized in 2 successive 2 sec epochs a series of waves fulfilling certain established criteria, i.e., having a sufficiently large relative amplitude, a frequency comprised within a range and a coefficient of variation below a maximum. Specific values of these parameters used in our study are given in Table I. A seizure was also recognized when the E E G demonstrated a sudden increase in frequency, even if the amplitude was unchanged from, or smaller than, that of the background activities. Seizure detections so determined in individual channels were rated from 1 to 4 according to their likelihood of being genuine. Detection ratings were then summed for all channels and a minimum sum of 4 was required to indicate a detection. The sensitivity of detection was adjusted by using a detection threshold. A threshold of 0 corresponded to the minimum sum of 4 with higher thresholds requiring correspondingly higher sums. Computations TABLE I Computer settings for detection of E E G focal seizures. Computer setting a
Amplitude threshold
Detection threshold
Frequency range (Hz)
Maximum coefficient of variation (%)
A B C D E F
2.7 2.5 2.7 2.5 2.3 2.1
1.0 1.0 0.0 0.0 0.0 0.0
3.0-20.0 3.0-20.0 3.0-20.0 3.0-20.0 3.0-20.0 3.0-20.0
40.0 40.0 40.0 40.0 40.0 40.0
a Monitor 3.0 program, Stellate Systems Inc., Westmount, Quebec, Canada.
were facilitated by the use of a math co-processor (Compaq 80387). In our study, the digitized EEGs were successively analyzed at each of 4 computer settings corresponding to increasing seizure detection sensitivity (Table I A - D ) . Each setting combined a seizure "detection threshold" of 1.0 or 0.0 with an "amplitude threshold" of 2.7 or 2.5. Values of 3.0-20.0 for "seizure frequency range" and 40.0 for "maximum coefficient of variation" were utilized at all settings. Some records (video-cassettes) containing one or more seizure patterns undetected at setting D (total: 154 h) were again analyzed with amplitude thresholds of 2.3 and 2.1 respectively, without altering the remaining detection parameters. After every E E G was analyzed at each computer setting, a printout was generated (IBM Proprinter II) which listed the seizure detections and their time of occurrence. This made it possible to retrieve and display on the computer monitor (Compaq 295 videographic monitor with 297 video-graphic controller board) the E E G samples recognized by the computer as seizures and to establish by visual analysis whether the computer determination had been right or wrong. An especially useful feature was that the channel numbers in which a detection had occurred were highlighted on the monitor. Whenever multiple successive seizure detections were made by the computer in the course of the same paroxysm, they were scored as a single detection. "Correct" and "missed" detections were expressed as percentages of total, and false detections were characterized by their rate of occurrence per hour. For each pattern recognized as a seizure we determined: the time of occurrence; the estimated duration; the approximate time elapsed from onset to full development (maximal amplitude and spread); the characteristic electrographic features; the patient's state of consciousness prior to seizure onset; the clinical manifestations (Commission on Classification and Terminology of the International l e a g u e Against Epilepsy 1985); and whether or not they had been signaled by the patient or other observers. As in our previous study (Pierelli et al. 1989), we arbitrarily regarded as E E G seizures the ictal patterns lasting 10 sec or longer. Eight partial seizures unaccompanied by E E G changes were excluded from consideration. "False positive" seizure detections were classified as presumably caused by: artifacts (muscular, ocular, mechanical, or mixed); normal E E G patterns (including beta, alpha and mu rhythms, vertex sharp waves and other sleep transients, spindles and theta and delta activity during sleep or in children); and, abnormal patterns (including theta and delta activities in waking adults and interictal spikes, sharp waves and their variants). Attempts were made to determine the causes of failure of CA to detect E E G seizure patterns unam-
C O M P U T E R DETECTION OF FOCAL S E I Z U R E S
3
biguously identified by combined FVR and SS. These included analyzing all E E G parameters of these paroxysms, searching for potentially confounding artifacts, and repeating several times the CA of these episodes.
Results
Computer analysis, seizure signaling, and visual analysis Our data base consisted of 253 E E G focal seizures. Their estimated mean duration was 44.87 sec (S.D. 3.32, median 35 sec) and the approximate mean time elapsed from their onset to their full development was 8.42 sec (S.D. 8.15, median 7 sec). These paroxysms occurred during both wakefulness (N = 132 or 52%) and sleep (N = 121 or 47.8%). In 175 instances (69.17%), they were associated with clinical manifestations of partial seizures, either complex (N = 158) or simple (N = 17). Symptoms and signs of these attacks were included in an earlier description (Pierelli et al. 1989). The remaining E E G seizure patterns were unaccompanied by detectable clinical changes. Table II summarizes the results of CA of the EEGs of our patients. Changing successively the parameters of CA from setting A through D yielded a progressive increase in correct seizure detections from 121 (47.82%) to 206 (81.42%). The increase in percentage of correct seizure detections achieved by ehanging computer parameters from A to D was paralleled by an increment in false positive detections from 2.70 to 5.38/h. Some of the tapes containing seizures missed by CA at setting D were re-analyzed using amplitude threshold of 2.3 and 2.1 (Table IE and F) with otherwise unchanged detection parameters. Correct seizure detections increased by 5.5% at E and 7.83% at F. However, this improvement was paralleled by further increase of false detections to 7.42/h at E and 11.60/h at F. Thus, no further attempt was made to quantify the effects of CA at these extreme settings. CA detected a much larger number of seizures than had been identified by SS at the time EVA monitoring
was performed (69 or 27.27%). This difference was statistically significant at all computer settings ( P < 0.001), with the X 2 ranging from 22.79 at A to 258.36 at D. However, SS recognized some seizures that CA had missed. Thus, combining CA at setting D with SS augmented the number of correct detections from 206 to 216 (85.38%). FVR by skilled technologists of all EEGs recorded on our patients during each 24 h EVA monitoring period (Pierelli et al. 1989) demanded from 2 to over 6 h of unrelenting visual effort, depending on the presence or absence and the number and features of seizures. A realistic approach to this labor intensive and fatiguing examination required that, when large numbers of highly stereotyped focal seizures occurred in a given 24 h period, some be disregarded to alleviate the burden of this analysis. These paroxysms were not logged in our monitoring records. In addition, some focal seizures were missed by FVR. These primarily consisted of ictal patterns, which barely exceeded 10 sec in duration, were characterized by relatively subtle E E G changes, or were partly obscured by artifacts. FVR performed under these constraints with the aid of SS resulted in the logging in our monitoring records of 186 of 253 E E G focal seizures (73.52%). This result was in contrast to the recognition of 216 seizure patterns by joint CA at setting D and SS. This difference was statistically significant (X 2 = 10.89, 1 dr, P = 0.001). However, 37 focal ictal patterns identified by FVR were missed by CA at setting D. Hence, combining FVR, SS and CA (D) resulted in the detection of a total of 253 seizures (100%) which represented our data base.
EEG seizures correctly detected or missed by computer analysis Our study demonstrated no apparent relationship between the electrographic features of the E E G seizure pattern and the computer's ability to detect it (Table liD. Similarly, no statistically significant difference existed between the number of seizures identified by CA
TABLE II Detection of 253 E E G focal seizures by computer analysis, seizure signaling, and both. Method of detection Computer analysis
Seizure signaling Combined computer analysis (D) and seizure signaling a Specified in Table I.
Computer setting a A B C D -
Detections Correct N(%)
Missed N (%)
False positive (N/h)
121 (47.82) 146 (57.70) 185 (73.12) 206 (81.42) 69 (27.27)
132 (52.18) 107 (42.30) 68 (26.88) 47 (18.58) 184 (72.73)
2.70 3.30 5.01 5.38 0.00
216 (85.38)
37 (14.62)
5.38
4
F. PAURI ET AL.
TABLE II1 EEG focal seizure patterns correctly detected or missed by computer analysis ~. Seizure pattern
Correctly detected Missed N
Repetitive spikes, sharp waves and their variants Paroxysmal alpha, theta, and delta Electrodecremental Combinations of the above Paroxysmal alpha, theta and delta, and artifacts Total number of seizures
(patients)
N
(patients)
17
(7)
1
( 1)
133 2 54
(8) (2) (6)
17 0 11
(4) (0) (2)
0
(0)
18
(1)
2()6
(12)
47
(5)
~' Setting D, Table 1.
d u r i n g wakefulness (n = 1 l l ) a n d sleep ( n = 9 5 ) , respectively (X 2 = 1.30, 1 dr, n.s.). In contrast, the n u m ber of c o m p u t e r - d e t e c t e d seizures was significantly larger when the E E G recordings utilized scalp and
s p h e n o i d a l electrodes (n = 124) c o m p a r e d to scalp electrodes alone (n = 82; X 2 = 7.09, 1 dr, P < 0.01; Fig. 1). T h e estimated d u r a t i o n of the E E G focal seizures detected by C A at settings A - D was significantly longer than that of seizures missed (Table IV). Moreover, when the c o m p u t e r settings were c h a n g e d from A through D, correct detections occurred for seizure patterns characterized by increasingly short time elapsed from onset to full d e v e l o p m e n t (Table V). T h e existence of a statistically significant difference as regards this last p a r a m e t e r b e t w e e n seizures detected and missed at settings A a n d C may reflect the use of a higher a m p l i t u d e threshold (2.7) than at B a n d D (2.5). Forty-seven of 253 E E G focal seizures were unrecognized by C A at setting D. T h e p r e s u m e d causes for these repeatedly verified false negative results i n c l u d e d o v e r w h e l m i n g E M G artifact in 18 seizures (38.30%) or the p r e s e n c e of E E G seizure activity in both the epoch u n d e r scrutiny a n d that utilized for comparatively assessing b a c k g r o u n d activity, in 15 (31.91%). In some
F8-Fp2 Fpl-F7 F8-Sp2 Sp2-Spl Spl-F7 F8-F4 F4-Fz Fz-F3 F3-F7 J
T4-C4 C4-Cz Cz-C3 C3-T3 T 6 - P4
.
.
.
.
.
.
P3-T5 01-02 TT 606741U
no clinical signs
chewing
iso~vT
1 sec
,
Fig. 1. Thirty-two-year-oldwoman with complex partial seizures characterized by automatisms, and generalized tonic-clonic attacks. This EEG seizure pattern occurred during wakefulness at 22:54 h. It began with a 5 Hz rhythmic discharge best developed at the left sphenoidal electrode and was computer detected within 8.5 sec of onset. The paroxysm subsequently involved the temporal and occipital areas of the scalp and lasted a total of 2 min and 16 sec. Neither the patient nor other observers signaled this episode. In this and subsequent figures, black arrow indicates time of first computer detection.
5
COMPUTER DETECTION OF FOCAL SEIZURES
Fp2-Pz F8-Pz Sp2-Pz T2-Pz T4-Pz T6-Pz Fz-Pz Cz-Pz Oz-Pz
E
,
l
_ ;.._L _ J i ~ ' : ' . ~ ~.; ; ~ , "
.aa
I
--
.,=,,.k
~__--~.r~Y. !--'.'--; . . . . . . . . . .
Ja.JtJ.
m
ill--
l,...=.Jil~uJIl=d.
"t"-,----lr.m~--
--r~
~ I
-,---, . rr,r
t
~
.......
~
=
L,~
. ; q . ;L. ~
tl
-,r~,'~
O=
.
a . ~ * l . i
; ~ ' ~ : " IR ~ ' r - -
-.
~tl
=
~
t
. . . .
d
I . a l l
~"
.....
...IJl
"
"
Fpl -Pz F7-Pz Spl-Pz T1 -Pz T3-Pz T5-Pz
w.,,.J~,.-,~ ..l,=.,~.~, . . . . . . . . . . . . . . . ,~a,.,. ~.=.~.~,=,=,u,JL. ,.aJ.,WL~,l,~lt,,,,,.a,.i_..~.j ,Li~,l,.a..a,a.L,../,,i.£6~..=~
LA 454563U
20o .vT i 1 sec
tonic posturing
Fig. 2. Twenty-four-year-old woman with partial seizures characterized by tonic posturing of one or both upper extremities, followed by automatisms, and infrequent generalized tonic-clonic attacks. During this episode, which occurred during sleep at 13:59 h, prominent E M G activity became apparent in the E E G and was followed within 18 sec by a 3 - 4 Hz rhythmic discharge. This was computer detected at Fz-Pz and Cz-Pz were E M G contamination was less evident. Total duration of the paroxysm was 40 sec. This seizure was not signaled by the patient or other observers.
puzzling
instances
ble explanation recognize
(14 seizures,
was found
certain
or 29.79%),
no plausi-
for the computer's
failure to
unambiguous
seizure patterns
ActiL'ities falsely detected as seizures by computer analysis
identi-
In
f i e d b y v i s u a l a n a l y s i s ( F i g . 3).
normal
some or
instances,
artifacts
abnormal
EEG
and,
less
patterns
commonly,
demonstrated
T A B L E IV Estimated duration of E E G focal seizures correctly detected or missed by computer analysis. Computer
Seizures detected
Seizures missed
setting a
Time (sec)
Time (sec)
Mean
(S.D.)
Median
Mean
(S.D.)
Median
A B C D
55.52 51.50 48.68 46.85
(38.50) (36.55) (34.29) (33.12)
41 40 35 35
35.10 35.81 34.50 36.15
(18.10) (18.98) (17.60) (19.69)
30 30 30 30
a Specified in Table I. t test, 2-tailed, df = 251.
Mean difference (sec)
Statistical comparison b
20.42 15.69 14.18 10.70
t t t t
= = = =
5.47, 4.06, 3.25, 2.13,
P P P P
< < = =
0.0001 0.0001 0.001 0.034
6
F. PAURI ET AL.
TABLE V Estimated time from onset to full development of EEG focal seizures correctly detected or missed by computer analysis. Computer setting a
Seizures detected N Time (sec)
A B C D
121 146 185 206
Seizures missed N Time (sec)
Mean
(S.D.)
Median
10.34 9.21 9.17 8.79
(8.83) (8.57) (8.77) (8.47)
8 7 7 7
132 107 68 47
Mean
(S.D.)
Median
6.66 7.34 6.38 6.79
(7.05) (7.44) (5.69) (6.34)
5 6 5 5
Mean difference (sec)
Statistical comparison
3.68 1.87 2.79 2.00
t t t t
= = = =
3.58, P 1.78, P 2.43, P 1.53, P
< 0.0001 < 0.07 c < 0.016 < 0.128 c
Specified in Table I. b t test, 2-tailed, df = 251. c Not significant.
p a r o x y s m a l f e a t u r e s , c h a n g e s in f r e q u e n c y , r h y t h m i c i t y , or combinations of these features which mimicked those c h a r a c t e r i z i n g E E G s e i z u r e p a t t e r n s . N o t surprisingly, s o m e o f t h e s e a c t i v i t i e s w e r e e r r o n e o u s l y d e t e c t e d as s e i z u r e s by t h e c o m p u t e r . I n t h e E E G s e x a m i n e d , t h e t o t a l n u m b e r o f false p o s i t i v e d e t e c t i o n s p r o g r e s s i v e l y i n c r e a s e d f r o m 1203 at s e t t i n g A to 2483 at D ( T a b l e
VI). A r t i f a c t s c a u s i n g 2030 false s e i z u r e d e t e c t i o n s at D c o n s i s t e d o f p o t e n t i a l s o f p r i m a r i l y m e c h a n i c a l (553 o r 2 7 . 2 4 % ) , m u s c u l a r (173 o r 8 . 5 2 % ) o r o c u l a r (112 o r 5 . 5 2 % ) origins. C o m b i n a t i o n s o f t h e s e e x t r a c e r e b r a l p o t e n t i a l s (Fig. 4) o r a r t i f a c t s f r o m o t h e r s o u r c e s acc o u n t e d f o r 1192 a d d i t i o n a l false p o s i t i v e r e s u l t s ( 5 8 . 7 2 % ) . N o statistically s i g n i f i c a n t r e l a t i o n was
F8-Fp2 Fpl-F7 F8-Sp2 Sp2-Spl Spl-F7 F8-F4 F4-Fz Fz-F3 F3-F7 T4-C4 C4-Cz Cz-C3 C3-T3 T6-P4 P3-T5 01 -O2
100~vT MT 404954U
paresthesias
1 sec
Fig. 3. Sixty-year-old man with simple partial seizures characterized by paresthesias, feeling of unreality and inability to move. This seizure occurred during wakefulness at 19:09 h. It was characterized by a rhythmic discharge which began at 5 Hz, involved the right sphenoidal lead and the adjacent inferior frontal and mid-temporal areas of the scalp, and subsequently decreased in frequency while increasing in amplitude. Total duration of the paroxysm was 16 sec. This seizure was signaled by the patient after its termination, but was not detected by the computer.
,
COMPUTER DETECTION OF FOCAL SEIZURES TABLE VI Presumed causes of false EEG seizure detections by computer analysis. Computer setting a A B C D
Artifacts
Normal EEG patterns
Abnormal EEG patterns
Total
N
(%)
N
(%)
N
(%)
N
(%)
964 1172 1872 2030
(80.13) (79.57) (83.83) (81.76)
110 144 206 260
(9.14) (9.78) (9.23) (10.47)
129 157 155 193
(10.72) (10.66) (6.94) (7.77)
1203 1473 2233 2483
(100) (100) (100) (100)
" Specified in Table 1.
d e m o n s t r a t e d b e t w e e n false d e t e c t i o n s c a u s e d by artifacts a n d time of day or night. N o r m a l E E G activities e r r o n e o u s l y d e t e c t e d as s e i z u r e p a t t e r n s in 260 i n s t a n c e s at setting D i n c l u d e d : p r o l o n g e d b u r s t s o f p r o m i n e n t p o t e n t i a l s o f b e t a (15 or 5.76%), a l p h a (46 o r 17.70%), t h e t a (11 o r 4.23%) or d e l t a (98 or 37.70%) frequency; long-lasting s l e e p spindles of large a m p l i t u d e (49 o r 18.84%); a n d r e p e t i t i v e V waves (8 or 3.07%) o r K c o m p l e x e s (33 or 12.70%). A m o n g the causes o f false d e t e c t i o n s in w a k i n g p a -
tients, o f special i n t e r e s t was the t r a n s i t i o n following eye c l o s u r e from a mixed f r e q u e n c y E E G p a t t e r n o f small v o l t a g e to a well s y n c h r o n i z e d a l p h a r h y t h m o f l a r g e r a m p l i t u d e (Fig. 5). A b n o r m a l E E G p a t t e r n s r e s p o n s i b l e for 193 s p u r i o u s s e i z u r e d e t e c t i o n s (7.7% at setting D) c o n s i s t e d of b u r s t s o f d e l t a waves (122 or 63.21%) o r r e p e t i t i v e spikes, s h a r p waves a n d t h e i r v a r i a n t s which lasted < 10 sec a n d w e r e t h e r e f o r e a r b i t r a r i l y c a t e g o r i z e d as interictal by the criteria of this study (71 o r 36.78%).
Fp2-Pz F8-Pz Sp2-Pz T2-Pz T4-Pz T6-Pz Fz-Pz Cz-Pz Oz-Pz Fpl -Pz F7-Pz Spl-Pz T1 -Pz T3-Pz T5-Pz
TT606741U
15o scratching head
vT 1 sec
Fig. 4. Same patient as in Fig. 1. Repeated scratching of the scalp caused rhythmic mechanical and muscular artifacts which mimicked an EEG seizure pattern and were detected by the computer as a seizure.
,
F. PAURI ET AL.
Fp2-F8 F8-T4 T4-T6 F4-C4 C4-P4 P4-O2 Fz-Cz Cz-Pz Pz-Oz F3-C3 C3-P3 P3-O1 Fpl-F7 F7-T3 T3-T5 ECG DM 608077U
i
J
I
/X eye closure
50.vT , 1 sec
Fig. 5. Thirty-five-year-oldwoman with complexpartial seizures. Eye closure was followedby the appearance of a prominent alpha rhythm which was computer detected as a seizure. Discussion
During the first 3 years of operation of our long-term monitoring unit, the identification of focal E E G seizure patterns in the entire 24 h E E G recording depended on combined FVR by skilled personnel and SS by the patients and other observers. This approach yielded a number of seizure detections that we felt at the time to be adequate and perhaps optimal for diagnostic purposes. However, the "fast" visual review proved very time consuming and fatiguing (Pierelli et al. 1989). Thus, we searched for alternative methods of seizure recognition and chose to test the computer detection method described by Gotman (1982, 1985, 1986, 1990) and Gotman et al. (1985). One of the attractive features of this technique was that it closely mimicked the approach followed by clinical electroencephalographers in visually identifying ictal E E G events. To our knowledge, no quantitative study comparable to ours exists in the available literature. The results of off-line CA of the tapes containing 253 E E G focal seizures varied according to the computer setting utilized. Changing computer parameters
from setting A through D (Table I) resulted in an increase of correct seizure detections from 47.82 to 81.42%. This improved detection efficiency was achieved at the price of an increase of false positive detections from 2.70 to 5.38/h. At settings E and F, further increments in correct seizure detections were associated with major increases in false positive results. The false positive detections generated by CA at setting D (5.38/h) proved far less burdensome than we had expected. Only 5-20 min were required for experienced technologists visually to review and categorize as correct or false detections all E E G samples identified and stored as seizures by the computer over each 24 h period. The relative brevity of this process was explained by the finding that patterns falsely recognized as seizures by CA mostly consisted of gross artifacts or normal or abnormal E E G activities which were readily identified by skilled personnel. By contrast, the assessment of substantially larger numbers of false positive detections at computer settings E and F was sufficiently time consuming to partly offset the benefits of the attending increase in correct detections. We concluded that CA at setting D detected a reasonably
COMPUTER DETECTION OF FOCAL SEIZURES
large proportion of focal EEG seizures while generating an acceptable rate of false positive results. Combining SS and CA raised the percentage of correct seizure detections to 85.38%. Our monitoring records indicated that the percentage of EEG focal seizures logged as detected by FVR (with the aid of SS) was 73.52%. This disappointingly small figure should not be construed as indicative of the potential efficiency of FVR in identifying EEG focal seizures. The paroxysms not logged in our records (26.48%) included, not only truly missed seizures but also ictal patterns deliberately rejected as redundant in recordings containing large numbers of highly stereotyped seizures. Thus, it would be misleading to compare the results of the earlier FVR with those of the present CA without taking into consideration the limitations of FVR performed during EVA monitoring. A precise assessment of the efficiency of this last method in detecting focal EEG seizures in our patients would have required repeating the visual analysis of all EEGs under scrutiny, an extremely time consuming effort which was beyond the scope of this study. However, we believe that our findings are revealing of the level of efficiency that FVR may realistically be expected to attain in most EVA monitoring units operating, like ours, under certain time constraints. Some of the findings of our study were not surprising. These included the observations that the detection of focal EEG seizure patterns by CA was facilitated by the use of sphenoidal electrodes, longer seizure durations, and shorter times from seizure onset to full development. By contrast, the lack of apparent relationships to seizure recognition of both the patient's state of consciousness and the time of occurrence of the paroxysms were somewhat unexpected. We had assumed that seizures occurring during sleep, hence mostly at night, would be more readily recognized by CA due to the manifestation of fewer artifacts. Upon completion of our investigation, computer setting D (Table I), was adopted as the default setting for detecting focal EEG seizure patterns in all subsequent long-term bipolar scalp and sphenoidal recordings. This setting proved satisfactory over a 2 year period, thus making it unnecessary to attempt to optimize detection parameters by varying them in individual patients according to the specific features of the seizures occurring during the early phases of monitoring. Our study indicates that CA of bipolarly recorded scalp and sphenoidal EEGs by the method of Gotman (1982, 1985, 1986, 1990) and Gotman et al. (1985) using appropriate detection parameters achieves a reason-
9
ably large proportion of correct detections of seizure patterns, while generating a tolerable rate of false positive detections. When this analysis is performed on-line, results are immediately available upon termination of and at any time during each 24 h monitoring period. Hence, we regard this method as a satisfactory alternative to and a distinct improvement over the extremely time consuming and fatiguing off-line FVR of all E E G records. Best results are obtained by combining this type of CA with SS by the patients and other observers which is insufficiently reliable when used alone (Pierelli et al. 1989). We express appreciation to Dr. Jean Gotman, Montreal Neurological Institute and Department of Neurology and Neurosurgery, McGill University, Montreal, Quebec, Canada for his suggestions and criticisms. This research was supported in part by grants of the Dottorato di Ricerca in Neurofisiopatologia e Neuroriabilitazione IV Ciclo, Universit~ degli Studi "La Sapienza," Rome, Italy to Dr. Flavia Pauri, and No. 106701/00/8902386 from the Italian Consiglio Nazionale delle Ricerche to Dr. Francesco Pierelli.
References Commission on Classification and Terminology of the International League Against Epilepsy. Proposal for classification of epilepsy and epileptic syndromes. Epilepsia, 1985, 26: 268-278. Gotman, J. Automatic recognition of epileptic seizures in the EEG. Electroenceph. clin. Neurophysiol., 1982, 54: 530-540. Gotman, J. Seizure recognition and analysis. In: J. Gotman, J.R. Ives and P. Gloor (Eds.), Long-Term Monitoring in Epilepsy. Electroenceph, clin. Neurophysiol., Suppl. 37. Elsevier, Amsterdam, 1985: 133-145. Gotman, J. Computer analysis during intensive monitoring of epileptic patients. In: R.J. Gumnit (Ed.), Intensive Neurodiagnostic Monitoring. Advances in Neurology, Vol. 46. Raven Press, New York, 1986: 249-269. Gotman, J. Automatic seizure detection: improvements and evaluation. Electroenceph. clin. Neurophysiol., 1990, 76: 317-324. Gotman, J. and Gloor, P. Automatic recognition and quantification of interictal epileptic activity in the human scalp EEG. Electroenceph. clin. Neurophysiol., 1976, 41: 513-529. Gotman, J., Ives, J.R. and Gloor, P. (Eds.). Long-Term Monitoring in Epilepsy. Electroenceph. clin. Neurophysiol., Suppl. 37. Elsevier, Amsterdam, 1985. Gumnit, R.J. (Ed.). Intensive Neurodiagnostic Monitoring. Advances in Neurology, Vol. 46. Raven Press, New York, 1987. Pauri, F., Pierelli, F. and Chatrian, G.-E. Automatic detection of partial epileptic seizures: comparison with fast visual review. Neurophysiol. Clin., 1990, 20 (Suppl.): 95S. Pierelli, F., Chatrian, G.-E., Erdly, W.W. et al. Long-term EEGvideo-audio monitoring: detection of partial epileptic seizures and psychogenic episodes by 24 h EEG record review. Epilepsia, 1989, 30: 513-523.
1
© 1992 Elsevier Scientific Publishers Ireland, Ltd. 0013-4649/92/$05.00
EEG 91142
L o n g - t e r m E E G - v i d e o - a u d i o m o n i t o r i n g : c o m p u t e r d e t e c t i o n o f focal E E G seizure patterns * Flavia Pauri a, Francesco Pierelli a, Gian-Emilio Chatrian a,b and William W. Erdly c '~ Section of EEG and Clinical Neurophysiology, Division of Neurology, Department of Medicine, h Department of Neurological Surgery, University of Washington, School of Medicine, and c Department of Psychology, University of Washington, School of Arts and Sciences, Seattle, WA (U.S.A.) (Accepted for publication: 16 September 1991)
Summary Twelve individuals with medically refractory partial seizures had undergone EEG-video-audio (EVA) monitoring over 1-15 (mean 10.5) days. We selectively reexamined available 15-channel EEGs (video-cassettes) totaling 461 h and containing 253 EEG focal seizures. Computer analysis (CA) of these bipolar records was performed using a mimetic method of seizure detection at 6 successive computer settings. We determined the computer parameters at which this method correctly detected a reasonably large percentage of seizures (81.42%) while generating an acceptable rate of false positive results (5.38/h). These parameters were adopted as the default setting for identifying focal EEG seizure patterns in all subsequent long-term bipolar scalp and sphenoidal recordings. Factors hindering or facilitating automatic seizure identification are discussed. It is concluded that on-line computer detection of focal EEG seizure patterns by this method offers a satisfactory alternative to and represents a distinct improvement over the extremely time consuming and fatiguing off-line fast visual review (FVR). Combining CA with seizure signaling (SS) by the patients and other observers increased the correct detections to 85.38~. CA is best used in conjunction with SS. Key words: EEG-video-audio monitoring; Focal EEG seizures; Partial seizures; Computer seizure detection; Visual seizure detection; Seizure signaling
Long-term EEG-video-audio (EVA) monitoring is an established method for the evaluation and treatment of selected patients suffering from intractable epileptic seizures (Gotman et al. 1985; Gumnit 1987). The success of this technique depends in large measure on the method followed to detect the patients' seizures. This may consist of one or more of the following: continuous or intermittent, direct or indirect, patient observation and E E G recording by trained personnel; seizure signaling (SS) by the patients and other observers; fast visual review (FVR) of all EEGs obtained during monitoring; and computer analysis (CA) of these EEGs. In our initial practice of EVA monitoring, we strived to recognize our patients' seizures as completely as possible by FVR of all their E E G records aided by SS by the patients and other observers (Pierelli et al. 1989). The present work reports on the results of
* Presented in part at the 5th European Congress of Clinical Neurophysiology, Paris, France, September 1990.
Correspondence to: Dr. G.E. Chatrian, Section of EEG and Clinical Neurophysiology, NN 282A, University of Washington Medical Center, RC-97, Seattle, WA 98195 (U.S.A.). Tel.: (206) 548-6465; Fax: (206) 548-4102.
CA of the EEGs of the same patients included in our previous investigation.
Patients and methods
We studied 12 patients who suffered from medically intractable partial seizures (Pierelli et al. 1989). All subjects had undergone 24 h EVA monitoring over 1-15 days (mean 10.5, S.D. 3.86, median 11) at the University of Washington Medical Center in Seattle between July 1986 and April 1988. These individuals ranged in age from 10 to 59 years (mean 29.1, S.D. 15.1, median 29.0). Eight were females and 4 males. Fifteen channels of E E G recorded bipolarly with scalp and, in 5 patients, sphenoidal electrodes and 1 channel of ECG were amplified and multiplexed (Telefactor TM 100-16 CTE), transmitted by cable telemetry to the E E G laboratory and stored on video cassette together with the video and the audio signals (Pierelli et al. 1989). At the time they were monitored, 216 focal seizure patterns had been detected on these individuals by combined FVR and SS. In the present investigation, we selectively reexamined those EEGs (video cassettes) that contained one or more focal seizures recognized earlier by FVR and SS (total: 461 h). These
2
F. P A U R I ET'AL.
records were: decoded (Telefactor TM 101-16D), lowpass filtered ( - 3 dB at 50 Hz) by a 6-pole switch capacitor filter (Biomedical Monitoring Systems MF32), digitized at the rate of 200 s a m p l e s / s e c / c h a n n e l by a 32-bit word length processor personal computer (Compaq Deskpro 386/20) and stored on a 130 Mb fixed disk drive with 135 Mb tape backup. The digitized EEGs were analyzed using a mimetic seizure detection method (Gotman 1982, 1985, 1986, 1990; Gotman et al. 1985) based on modifications of the the algorithm described by Gotman and Gloor (1976). The computer program (Stellate System Monitor 3.0) broke individual E E G waves into half-waves and determined the following 3 measures for each 2 sec E E G epoch: (1) the average amplitude of the half-waves relative to that of a sample of background EEG activity consisting of a 16 sec epoch ending 20 sec prior to the beginning of the current epoch. This parameter established whether or not the E E G epoch under scrutiny had "paroxysmal" features; (2) the average duration of the half-waves as a measure of frequency; and (3) the coefficient of variation of half-wave duration (S.D./mean) expressing degree of rhythmicity. A seizure was detected in any given channel when the computer recognized in 2 successive 2 sec epochs a series of waves fulfilling certain established criteria, i.e., having a sufficiently large relative amplitude, a frequency comprised within a range and a coefficient of variation below a maximum. Specific values of these parameters used in our study are given in Table I. A seizure was also recognized when the E E G demonstrated a sudden increase in frequency, even if the amplitude was unchanged from, or smaller than, that of the background activities. Seizure detections so determined in individual channels were rated from 1 to 4 according to their likelihood of being genuine. Detection ratings were then summed for all channels and a minimum sum of 4 was required to indicate a detection. The sensitivity of detection was adjusted by using a detection threshold. A threshold of 0 corresponded to the minimum sum of 4 with higher thresholds requiring correspondingly higher sums. Computations TABLE I Computer settings for detection of E E G focal seizures. Computer setting a
Amplitude threshold
Detection threshold
Frequency range (Hz)
Maximum coefficient of variation (%)
A B C D E F
2.7 2.5 2.7 2.5 2.3 2.1
1.0 1.0 0.0 0.0 0.0 0.0
3.0-20.0 3.0-20.0 3.0-20.0 3.0-20.0 3.0-20.0 3.0-20.0
40.0 40.0 40.0 40.0 40.0 40.0
a Monitor 3.0 program, Stellate Systems Inc., Westmount, Quebec, Canada.
were facilitated by the use of a math co-processor (Compaq 80387). In our study, the digitized EEGs were successively analyzed at each of 4 computer settings corresponding to increasing seizure detection sensitivity (Table I A - D ) . Each setting combined a seizure "detection threshold" of 1.0 or 0.0 with an "amplitude threshold" of 2.7 or 2.5. Values of 3.0-20.0 for "seizure frequency range" and 40.0 for "maximum coefficient of variation" were utilized at all settings. Some records (video-cassettes) containing one or more seizure patterns undetected at setting D (total: 154 h) were again analyzed with amplitude thresholds of 2.3 and 2.1 respectively, without altering the remaining detection parameters. After every E E G was analyzed at each computer setting, a printout was generated (IBM Proprinter II) which listed the seizure detections and their time of occurrence. This made it possible to retrieve and display on the computer monitor (Compaq 295 videographic monitor with 297 video-graphic controller board) the E E G samples recognized by the computer as seizures and to establish by visual analysis whether the computer determination had been right or wrong. An especially useful feature was that the channel numbers in which a detection had occurred were highlighted on the monitor. Whenever multiple successive seizure detections were made by the computer in the course of the same paroxysm, they were scored as a single detection. "Correct" and "missed" detections were expressed as percentages of total, and false detections were characterized by their rate of occurrence per hour. For each pattern recognized as a seizure we determined: the time of occurrence; the estimated duration; the approximate time elapsed from onset to full development (maximal amplitude and spread); the characteristic electrographic features; the patient's state of consciousness prior to seizure onset; the clinical manifestations (Commission on Classification and Terminology of the International l e a g u e Against Epilepsy 1985); and whether or not they had been signaled by the patient or other observers. As in our previous study (Pierelli et al. 1989), we arbitrarily regarded as E E G seizures the ictal patterns lasting 10 sec or longer. Eight partial seizures unaccompanied by E E G changes were excluded from consideration. "False positive" seizure detections were classified as presumably caused by: artifacts (muscular, ocular, mechanical, or mixed); normal E E G patterns (including beta, alpha and mu rhythms, vertex sharp waves and other sleep transients, spindles and theta and delta activity during sleep or in children); and, abnormal patterns (including theta and delta activities in waking adults and interictal spikes, sharp waves and their variants). Attempts were made to determine the causes of failure of CA to detect E E G seizure patterns unam-
C O M P U T E R DETECTION OF FOCAL S E I Z U R E S
3
biguously identified by combined FVR and SS. These included analyzing all E E G parameters of these paroxysms, searching for potentially confounding artifacts, and repeating several times the CA of these episodes.
Results
Computer analysis, seizure signaling, and visual analysis Our data base consisted of 253 E E G focal seizures. Their estimated mean duration was 44.87 sec (S.D. 3.32, median 35 sec) and the approximate mean time elapsed from their onset to their full development was 8.42 sec (S.D. 8.15, median 7 sec). These paroxysms occurred during both wakefulness (N = 132 or 52%) and sleep (N = 121 or 47.8%). In 175 instances (69.17%), they were associated with clinical manifestations of partial seizures, either complex (N = 158) or simple (N = 17). Symptoms and signs of these attacks were included in an earlier description (Pierelli et al. 1989). The remaining E E G seizure patterns were unaccompanied by detectable clinical changes. Table II summarizes the results of CA of the EEGs of our patients. Changing successively the parameters of CA from setting A through D yielded a progressive increase in correct seizure detections from 121 (47.82%) to 206 (81.42%). The increase in percentage of correct seizure detections achieved by ehanging computer parameters from A to D was paralleled by an increment in false positive detections from 2.70 to 5.38/h. Some of the tapes containing seizures missed by CA at setting D were re-analyzed using amplitude threshold of 2.3 and 2.1 (Table IE and F) with otherwise unchanged detection parameters. Correct seizure detections increased by 5.5% at E and 7.83% at F. However, this improvement was paralleled by further increase of false detections to 7.42/h at E and 11.60/h at F. Thus, no further attempt was made to quantify the effects of CA at these extreme settings. CA detected a much larger number of seizures than had been identified by SS at the time EVA monitoring
was performed (69 or 27.27%). This difference was statistically significant at all computer settings ( P < 0.001), with the X 2 ranging from 22.79 at A to 258.36 at D. However, SS recognized some seizures that CA had missed. Thus, combining CA at setting D with SS augmented the number of correct detections from 206 to 216 (85.38%). FVR by skilled technologists of all EEGs recorded on our patients during each 24 h EVA monitoring period (Pierelli et al. 1989) demanded from 2 to over 6 h of unrelenting visual effort, depending on the presence or absence and the number and features of seizures. A realistic approach to this labor intensive and fatiguing examination required that, when large numbers of highly stereotyped focal seizures occurred in a given 24 h period, some be disregarded to alleviate the burden of this analysis. These paroxysms were not logged in our monitoring records. In addition, some focal seizures were missed by FVR. These primarily consisted of ictal patterns, which barely exceeded 10 sec in duration, were characterized by relatively subtle E E G changes, or were partly obscured by artifacts. FVR performed under these constraints with the aid of SS resulted in the logging in our monitoring records of 186 of 253 E E G focal seizures (73.52%). This result was in contrast to the recognition of 216 seizure patterns by joint CA at setting D and SS. This difference was statistically significant (X 2 = 10.89, 1 dr, P = 0.001). However, 37 focal ictal patterns identified by FVR were missed by CA at setting D. Hence, combining FVR, SS and CA (D) resulted in the detection of a total of 253 seizures (100%) which represented our data base.
EEG seizures correctly detected or missed by computer analysis Our study demonstrated no apparent relationship between the electrographic features of the E E G seizure pattern and the computer's ability to detect it (Table liD. Similarly, no statistically significant difference existed between the number of seizures identified by CA
TABLE II Detection of 253 E E G focal seizures by computer analysis, seizure signaling, and both. Method of detection Computer analysis
Seizure signaling Combined computer analysis (D) and seizure signaling a Specified in Table I.
Computer setting a A B C D -
Detections Correct N(%)
Missed N (%)
False positive (N/h)
121 (47.82) 146 (57.70) 185 (73.12) 206 (81.42) 69 (27.27)
132 (52.18) 107 (42.30) 68 (26.88) 47 (18.58) 184 (72.73)
2.70 3.30 5.01 5.38 0.00
216 (85.38)
37 (14.62)
5.38
4
F. PAURI ET AL.
TABLE II1 EEG focal seizure patterns correctly detected or missed by computer analysis ~. Seizure pattern
Correctly detected Missed N
Repetitive spikes, sharp waves and their variants Paroxysmal alpha, theta, and delta Electrodecremental Combinations of the above Paroxysmal alpha, theta and delta, and artifacts Total number of seizures
(patients)
N
(patients)
17
(7)
1
( 1)
133 2 54
(8) (2) (6)
17 0 11
(4) (0) (2)
0
(0)
18
(1)
2()6
(12)
47
(5)
~' Setting D, Table 1.
d u r i n g wakefulness (n = 1 l l ) a n d sleep ( n = 9 5 ) , respectively (X 2 = 1.30, 1 dr, n.s.). In contrast, the n u m ber of c o m p u t e r - d e t e c t e d seizures was significantly larger when the E E G recordings utilized scalp and
s p h e n o i d a l electrodes (n = 124) c o m p a r e d to scalp electrodes alone (n = 82; X 2 = 7.09, 1 dr, P < 0.01; Fig. 1). T h e estimated d u r a t i o n of the E E G focal seizures detected by C A at settings A - D was significantly longer than that of seizures missed (Table IV). Moreover, when the c o m p u t e r settings were c h a n g e d from A through D, correct detections occurred for seizure patterns characterized by increasingly short time elapsed from onset to full d e v e l o p m e n t (Table V). T h e existence of a statistically significant difference as regards this last p a r a m e t e r b e t w e e n seizures detected and missed at settings A a n d C may reflect the use of a higher a m p l i t u d e threshold (2.7) than at B a n d D (2.5). Forty-seven of 253 E E G focal seizures were unrecognized by C A at setting D. T h e p r e s u m e d causes for these repeatedly verified false negative results i n c l u d e d o v e r w h e l m i n g E M G artifact in 18 seizures (38.30%) or the p r e s e n c e of E E G seizure activity in both the epoch u n d e r scrutiny a n d that utilized for comparatively assessing b a c k g r o u n d activity, in 15 (31.91%). In some
F8-Fp2 Fpl-F7 F8-Sp2 Sp2-Spl Spl-F7 F8-F4 F4-Fz Fz-F3 F3-F7 J
T4-C4 C4-Cz Cz-C3 C3-T3 T 6 - P4
.
.
.
.
.
.
P3-T5 01-02 TT 606741U
no clinical signs
chewing
iso~vT
1 sec
,
Fig. 1. Thirty-two-year-oldwoman with complex partial seizures characterized by automatisms, and generalized tonic-clonic attacks. This EEG seizure pattern occurred during wakefulness at 22:54 h. It began with a 5 Hz rhythmic discharge best developed at the left sphenoidal electrode and was computer detected within 8.5 sec of onset. The paroxysm subsequently involved the temporal and occipital areas of the scalp and lasted a total of 2 min and 16 sec. Neither the patient nor other observers signaled this episode. In this and subsequent figures, black arrow indicates time of first computer detection.
5
COMPUTER DETECTION OF FOCAL SEIZURES
Fp2-Pz F8-Pz Sp2-Pz T2-Pz T4-Pz T6-Pz Fz-Pz Cz-Pz Oz-Pz
E
,
l
_ ;.._L _ J i ~ ' : ' . ~ ~.; ; ~ , "
.aa
I
--
.,=,,.k
~__--~.r~Y. !--'.'--; . . . . . . . . . .
Ja.JtJ.
m
ill--
l,...=.Jil~uJIl=d.
"t"-,----lr.m~--
--r~
~ I
-,---, . rr,r
t
~
.......
~
=
L,~
. ; q . ;L. ~
tl
-,r~,'~
O=
.
a . ~ * l . i
; ~ ' ~ : " IR ~ ' r - -
-.
~tl
=
~
t
. . . .
d
I . a l l
~"
.....
...IJl
"
"
Fpl -Pz F7-Pz Spl-Pz T1 -Pz T3-Pz T5-Pz
w.,,.J~,.-,~ ..l,=.,~.~, . . . . . . . . . . . . . . . ,~a,.,. ~.=.~.~,=,=,u,JL. ,.aJ.,WL~,l,~lt,,,,,.a,.i_..~.j ,Li~,l,.a..a,a.L,../,,i.£6~..=~
LA 454563U
20o .vT i 1 sec
tonic posturing
Fig. 2. Twenty-four-year-old woman with partial seizures characterized by tonic posturing of one or both upper extremities, followed by automatisms, and infrequent generalized tonic-clonic attacks. During this episode, which occurred during sleep at 13:59 h, prominent E M G activity became apparent in the E E G and was followed within 18 sec by a 3 - 4 Hz rhythmic discharge. This was computer detected at Fz-Pz and Cz-Pz were E M G contamination was less evident. Total duration of the paroxysm was 40 sec. This seizure was not signaled by the patient or other observers.
puzzling
instances
ble explanation recognize
(14 seizures,
was found
certain
or 29.79%),
no plausi-
for the computer's
failure to
unambiguous
seizure patterns
ActiL'ities falsely detected as seizures by computer analysis
identi-
In
f i e d b y v i s u a l a n a l y s i s ( F i g . 3).
normal
some or
instances,
artifacts
abnormal
EEG
and,
less
patterns
commonly,
demonstrated
T A B L E IV Estimated duration of E E G focal seizures correctly detected or missed by computer analysis. Computer
Seizures detected
Seizures missed
setting a
Time (sec)
Time (sec)
Mean
(S.D.)
Median
Mean
(S.D.)
Median
A B C D
55.52 51.50 48.68 46.85
(38.50) (36.55) (34.29) (33.12)
41 40 35 35
35.10 35.81 34.50 36.15
(18.10) (18.98) (17.60) (19.69)
30 30 30 30
a Specified in Table I. t test, 2-tailed, df = 251.
Mean difference (sec)
Statistical comparison b
20.42 15.69 14.18 10.70
t t t t
= = = =
5.47, 4.06, 3.25, 2.13,
P P P P
< < = =
0.0001 0.0001 0.001 0.034
6
F. PAURI ET AL.
TABLE V Estimated time from onset to full development of EEG focal seizures correctly detected or missed by computer analysis. Computer setting a
Seizures detected N Time (sec)
A B C D
121 146 185 206
Seizures missed N Time (sec)
Mean
(S.D.)
Median
10.34 9.21 9.17 8.79
(8.83) (8.57) (8.77) (8.47)
8 7 7 7
132 107 68 47
Mean
(S.D.)
Median
6.66 7.34 6.38 6.79
(7.05) (7.44) (5.69) (6.34)
5 6 5 5
Mean difference (sec)
Statistical comparison
3.68 1.87 2.79 2.00
t t t t
= = = =
3.58, P 1.78, P 2.43, P 1.53, P
< 0.0001 < 0.07 c < 0.016 < 0.128 c
Specified in Table I. b t test, 2-tailed, df = 251. c Not significant.
p a r o x y s m a l f e a t u r e s , c h a n g e s in f r e q u e n c y , r h y t h m i c i t y , or combinations of these features which mimicked those c h a r a c t e r i z i n g E E G s e i z u r e p a t t e r n s . N o t surprisingly, s o m e o f t h e s e a c t i v i t i e s w e r e e r r o n e o u s l y d e t e c t e d as s e i z u r e s by t h e c o m p u t e r . I n t h e E E G s e x a m i n e d , t h e t o t a l n u m b e r o f false p o s i t i v e d e t e c t i o n s p r o g r e s s i v e l y i n c r e a s e d f r o m 1203 at s e t t i n g A to 2483 at D ( T a b l e
VI). A r t i f a c t s c a u s i n g 2030 false s e i z u r e d e t e c t i o n s at D c o n s i s t e d o f p o t e n t i a l s o f p r i m a r i l y m e c h a n i c a l (553 o r 2 7 . 2 4 % ) , m u s c u l a r (173 o r 8 . 5 2 % ) o r o c u l a r (112 o r 5 . 5 2 % ) origins. C o m b i n a t i o n s o f t h e s e e x t r a c e r e b r a l p o t e n t i a l s (Fig. 4) o r a r t i f a c t s f r o m o t h e r s o u r c e s acc o u n t e d f o r 1192 a d d i t i o n a l false p o s i t i v e r e s u l t s ( 5 8 . 7 2 % ) . N o statistically s i g n i f i c a n t r e l a t i o n was
F8-Fp2 Fpl-F7 F8-Sp2 Sp2-Spl Spl-F7 F8-F4 F4-Fz Fz-F3 F3-F7 T4-C4 C4-Cz Cz-C3 C3-T3 T6-P4 P3-T5 01 -O2
100~vT MT 404954U
paresthesias
1 sec
Fig. 3. Sixty-year-old man with simple partial seizures characterized by paresthesias, feeling of unreality and inability to move. This seizure occurred during wakefulness at 19:09 h. It was characterized by a rhythmic discharge which began at 5 Hz, involved the right sphenoidal lead and the adjacent inferior frontal and mid-temporal areas of the scalp, and subsequently decreased in frequency while increasing in amplitude. Total duration of the paroxysm was 16 sec. This seizure was signaled by the patient after its termination, but was not detected by the computer.
,
COMPUTER DETECTION OF FOCAL SEIZURES TABLE VI Presumed causes of false EEG seizure detections by computer analysis. Computer setting a A B C D
Artifacts
Normal EEG patterns
Abnormal EEG patterns
Total
N
(%)
N
(%)
N
(%)
N
(%)
964 1172 1872 2030
(80.13) (79.57) (83.83) (81.76)
110 144 206 260
(9.14) (9.78) (9.23) (10.47)
129 157 155 193
(10.72) (10.66) (6.94) (7.77)
1203 1473 2233 2483
(100) (100) (100) (100)
" Specified in Table 1.
d e m o n s t r a t e d b e t w e e n false d e t e c t i o n s c a u s e d by artifacts a n d time of day or night. N o r m a l E E G activities e r r o n e o u s l y d e t e c t e d as s e i z u r e p a t t e r n s in 260 i n s t a n c e s at setting D i n c l u d e d : p r o l o n g e d b u r s t s o f p r o m i n e n t p o t e n t i a l s o f b e t a (15 or 5.76%), a l p h a (46 o r 17.70%), t h e t a (11 o r 4.23%) or d e l t a (98 or 37.70%) frequency; long-lasting s l e e p spindles of large a m p l i t u d e (49 o r 18.84%); a n d r e p e t i t i v e V waves (8 or 3.07%) o r K c o m p l e x e s (33 or 12.70%). A m o n g the causes o f false d e t e c t i o n s in w a k i n g p a -
tients, o f special i n t e r e s t was the t r a n s i t i o n following eye c l o s u r e from a mixed f r e q u e n c y E E G p a t t e r n o f small v o l t a g e to a well s y n c h r o n i z e d a l p h a r h y t h m o f l a r g e r a m p l i t u d e (Fig. 5). A b n o r m a l E E G p a t t e r n s r e s p o n s i b l e for 193 s p u r i o u s s e i z u r e d e t e c t i o n s (7.7% at setting D) c o n s i s t e d of b u r s t s o f d e l t a waves (122 or 63.21%) o r r e p e t i t i v e spikes, s h a r p waves a n d t h e i r v a r i a n t s which lasted < 10 sec a n d w e r e t h e r e f o r e a r b i t r a r i l y c a t e g o r i z e d as interictal by the criteria of this study (71 o r 36.78%).
Fp2-Pz F8-Pz Sp2-Pz T2-Pz T4-Pz T6-Pz Fz-Pz Cz-Pz Oz-Pz Fpl -Pz F7-Pz Spl-Pz T1 -Pz T3-Pz T5-Pz
TT606741U
15o scratching head
vT 1 sec
Fig. 4. Same patient as in Fig. 1. Repeated scratching of the scalp caused rhythmic mechanical and muscular artifacts which mimicked an EEG seizure pattern and were detected by the computer as a seizure.
,
F. PAURI ET AL.
Fp2-F8 F8-T4 T4-T6 F4-C4 C4-P4 P4-O2 Fz-Cz Cz-Pz Pz-Oz F3-C3 C3-P3 P3-O1 Fpl-F7 F7-T3 T3-T5 ECG DM 608077U
i
J
I
/X eye closure
50.vT , 1 sec
Fig. 5. Thirty-five-year-oldwoman with complexpartial seizures. Eye closure was followedby the appearance of a prominent alpha rhythm which was computer detected as a seizure. Discussion
During the first 3 years of operation of our long-term monitoring unit, the identification of focal E E G seizure patterns in the entire 24 h E E G recording depended on combined FVR by skilled personnel and SS by the patients and other observers. This approach yielded a number of seizure detections that we felt at the time to be adequate and perhaps optimal for diagnostic purposes. However, the "fast" visual review proved very time consuming and fatiguing (Pierelli et al. 1989). Thus, we searched for alternative methods of seizure recognition and chose to test the computer detection method described by Gotman (1982, 1985, 1986, 1990) and Gotman et al. (1985). One of the attractive features of this technique was that it closely mimicked the approach followed by clinical electroencephalographers in visually identifying ictal E E G events. To our knowledge, no quantitative study comparable to ours exists in the available literature. The results of off-line CA of the tapes containing 253 E E G focal seizures varied according to the computer setting utilized. Changing computer parameters
from setting A through D (Table I) resulted in an increase of correct seizure detections from 47.82 to 81.42%. This improved detection efficiency was achieved at the price of an increase of false positive detections from 2.70 to 5.38/h. At settings E and F, further increments in correct seizure detections were associated with major increases in false positive results. The false positive detections generated by CA at setting D (5.38/h) proved far less burdensome than we had expected. Only 5-20 min were required for experienced technologists visually to review and categorize as correct or false detections all E E G samples identified and stored as seizures by the computer over each 24 h period. The relative brevity of this process was explained by the finding that patterns falsely recognized as seizures by CA mostly consisted of gross artifacts or normal or abnormal E E G activities which were readily identified by skilled personnel. By contrast, the assessment of substantially larger numbers of false positive detections at computer settings E and F was sufficiently time consuming to partly offset the benefits of the attending increase in correct detections. We concluded that CA at setting D detected a reasonably
COMPUTER DETECTION OF FOCAL SEIZURES
large proportion of focal EEG seizures while generating an acceptable rate of false positive results. Combining SS and CA raised the percentage of correct seizure detections to 85.38%. Our monitoring records indicated that the percentage of EEG focal seizures logged as detected by FVR (with the aid of SS) was 73.52%. This disappointingly small figure should not be construed as indicative of the potential efficiency of FVR in identifying EEG focal seizures. The paroxysms not logged in our records (26.48%) included, not only truly missed seizures but also ictal patterns deliberately rejected as redundant in recordings containing large numbers of highly stereotyped seizures. Thus, it would be misleading to compare the results of the earlier FVR with those of the present CA without taking into consideration the limitations of FVR performed during EVA monitoring. A precise assessment of the efficiency of this last method in detecting focal EEG seizures in our patients would have required repeating the visual analysis of all EEGs under scrutiny, an extremely time consuming effort which was beyond the scope of this study. However, we believe that our findings are revealing of the level of efficiency that FVR may realistically be expected to attain in most EVA monitoring units operating, like ours, under certain time constraints. Some of the findings of our study were not surprising. These included the observations that the detection of focal EEG seizure patterns by CA was facilitated by the use of sphenoidal electrodes, longer seizure durations, and shorter times from seizure onset to full development. By contrast, the lack of apparent relationships to seizure recognition of both the patient's state of consciousness and the time of occurrence of the paroxysms were somewhat unexpected. We had assumed that seizures occurring during sleep, hence mostly at night, would be more readily recognized by CA due to the manifestation of fewer artifacts. Upon completion of our investigation, computer setting D (Table I), was adopted as the default setting for detecting focal EEG seizure patterns in all subsequent long-term bipolar scalp and sphenoidal recordings. This setting proved satisfactory over a 2 year period, thus making it unnecessary to attempt to optimize detection parameters by varying them in individual patients according to the specific features of the seizures occurring during the early phases of monitoring. Our study indicates that CA of bipolarly recorded scalp and sphenoidal EEGs by the method of Gotman (1982, 1985, 1986, 1990) and Gotman et al. (1985) using appropriate detection parameters achieves a reason-
9
ably large proportion of correct detections of seizure patterns, while generating a tolerable rate of false positive detections. When this analysis is performed on-line, results are immediately available upon termination of and at any time during each 24 h monitoring period. Hence, we regard this method as a satisfactory alternative to and a distinct improvement over the extremely time consuming and fatiguing off-line FVR of all E E G records. Best results are obtained by combining this type of CA with SS by the patients and other observers which is insufficiently reliable when used alone (Pierelli et al. 1989). We express appreciation to Dr. Jean Gotman, Montreal Neurological Institute and Department of Neurology and Neurosurgery, McGill University, Montreal, Quebec, Canada for his suggestions and criticisms. This research was supported in part by grants of the Dottorato di Ricerca in Neurofisiopatologia e Neuroriabilitazione IV Ciclo, Universit~ degli Studi "La Sapienza," Rome, Italy to Dr. Flavia Pauri, and No. 106701/00/8902386 from the Italian Consiglio Nazionale delle Ricerche to Dr. Francesco Pierelli.
References Commission on Classification and Terminology of the International League Against Epilepsy. Proposal for classification of epilepsy and epileptic syndromes. Epilepsia, 1985, 26: 268-278. Gotman, J. Automatic recognition of epileptic seizures in the EEG. Electroenceph. clin. Neurophysiol., 1982, 54: 530-540. Gotman, J. Seizure recognition and analysis. In: J. Gotman, J.R. Ives and P. Gloor (Eds.), Long-Term Monitoring in Epilepsy. Electroenceph, clin. Neurophysiol., Suppl. 37. Elsevier, Amsterdam, 1985: 133-145. Gotman, J. Computer analysis during intensive monitoring of epileptic patients. In: R.J. Gumnit (Ed.), Intensive Neurodiagnostic Monitoring. Advances in Neurology, Vol. 46. Raven Press, New York, 1986: 249-269. Gotman, J. Automatic seizure detection: improvements and evaluation. Electroenceph. clin. Neurophysiol., 1990, 76: 317-324. Gotman, J. and Gloor, P. Automatic recognition and quantification of interictal epileptic activity in the human scalp EEG. Electroenceph. clin. Neurophysiol., 1976, 41: 513-529. Gotman, J., Ives, J.R. and Gloor, P. (Eds.). Long-Term Monitoring in Epilepsy. Electroenceph. clin. Neurophysiol., Suppl. 37. Elsevier, Amsterdam, 1985. Gumnit, R.J. (Ed.). Intensive Neurodiagnostic Monitoring. Advances in Neurology, Vol. 46. Raven Press, New York, 1987. Pauri, F., Pierelli, F. and Chatrian, G.-E. Automatic detection of partial epileptic seizures: comparison with fast visual review. Neurophysiol. Clin., 1990, 20 (Suppl.): 95S. Pierelli, F., Chatrian, G.-E., Erdly, W.W. et al. Long-term EEGvideo-audio monitoring: detection of partial epileptic seizures and psychogenic episodes by 24 h EEG record review. Epilepsia, 1989, 30: 513-523.
