- Email: [email protected]
Automatic EEG interpretation: a new computer-assisted system for the automatic integrative interpretation of awake background EEG
Electroencephalography and clinical Neurophysiology, 82 (1992) 423-431
423
© 1992 Elsevier Scientific Publishers Ireland, Ltd. 0013-4649/92/$05.00
EEG91056
Automatic EEG interpretation: a new computer-assisted system for the automatic integrative interpretation of awake background EEG Masatoshi Nakamura, Hiroshi Shibasaki a, Kaoru Imajoh, Shigeto Nishida, Ryuji Neshige b and Akio Ikeda a Department of Electrical Engineering, Saga University, Honjomachi, Saga 840 (Japan), a Department of Brain Pathophysiology, Kyoto University Faculty of Medicine, Shogoin, Sakyo-ku, Kyoto 606 (Japan), and b Department of Internal Medicine, Saga Medical School, Nabeshima, Saga 849 (Japan) (Accepted for publication: 9 December 1991)
Summary A new computer-assisted system for automatic interpretation of the awake electroencephalogram (EEG) was developed. First, all the items necessary for EEG interpretation were determined in accordance with the procedure that a qualified electroencephalographer (EEGer) goes through for the visual inspection of the background EEG activity, and then each item was defined quantitatively. For the automatic interpretation, specific EEG parameters were determined for each item so that they could fit the graded judgement of the item by the qualified EEGer as closely as possible. These specific EEG parameters were actually calculated from periodograms obtained from the time series of EEG records of 14 patients with various neurological diseases. The automatic EEG interpretation system thus established was applied to the EEG data of these 14 subjects and to 3 additional EEGs, and the results were compared with those obtained through the visual interpretation by the EEGer. This automatic EEG interpretation was found to be in good agreement with the visual interpretation by the EEGer in most EEG records. In contrast with the previous automatic analyses of EEG which were focussed on certain aspects of EEG such as the dominant rhythm, the present system is unique in its capability of providing an integrative interpretation of the spontaneous awake EEG by taking into account all its features except for paroxysmal abnormalities. Key words: Automatic EEG interpretation; Awake EEG; Quantitative scores of EEG
The electronencephalogram (EEG) is a summation of electrical activities generated by nerve cells in the cerebral cortex and recorded from the scalp. As the EEG reflects at least a part of the functional status of the subject's brain, the EEG interpreted by electroencephalographers (EEGers) is used as a diagnostic aid in diseases affecting the brain. When an EEGer interprets an EEG record, he analyses its time and spatial features by visual inspection. Automatic analyses of EEG have been attempted by many investigators. Among various kinds of automatic EEG interpretation previously reported, spike detection in either the scalp-recorded or depth-recorded EEG in epileptic patients has drawn particular attention (Gotman and Gloor 1976; Frost 1985; Gotman 1985). Automatic analyses of sleep stages by using EEG data have also been implemented, and an accuracy of about 80% in the recognition was achieved (Inoue et al. 1982). Auto-
Correspondence to: Dr. Masatoshi Nakamura, Department of Electrical Engineering, Saga University, Honjomachi, Saga 840 (Japan).
matic diagnosis of awake EEGs has been investigated by Kowada et al. (1971) and Yamamoto et al. (1975). However, most previous studies were restricted to a certain aspect of EEG characteristics such as the dominant rhythm or slow waves. An automatic integrative interpretation of a full EEG record has not yet been accomplished, mainly because a large number of complex items are involved in the interpretation of the EEG. In previous works the authors (Nakamura et al. 1985, 1989b) succeeded in an automatic scoring of 'organization' of the dominant or basic rhythm of EEG activity which is one of the most essential issues for the interpretation of the spontaneous awake EEG obtained while the subject is at rest with the eyes closed. In the present study, we further advanced our previous 'organization' scoring method to accomplish an automatic integrative interpretation of the awake EEG record, taking into account all items necessary for EEG interpretation except for paroxysmal abnormalities such as spikes and spike-and-waves. This system can be used clinically as an assistant tool for neurologists and even for EEGers themselves, and may be the first step for its clinical application to hospital use.
424
M. NAKAMURA ET AL.
Methods
automatic interpretation so that the results can be compared with those derived from visual inspection by the EEGer.
(1) Data acquisition EEGs of 14 subjects recorded on paper were arbitrarily selected from consecutive EEG records of Saga Medical School for development of the present system for automatic EEG interpretation. These EEGs were recorded from 14 patients, aged 20-64 years, with various neurological diseases. The diagnosis of these diseases is not listed here because it does not matter for the purpose of the present study, although all patients needed EEG examination for diagnostic purposes. All EEGs were recorded in a quiet, dimly lit room where the subject was placed in a supine position in a bed with the eyes closed. Exploring cup electrodes were fixed to the scalp at points Fp a, F3, C3, P3, O1, FP2, F4, C4, P4, 0 2 , F7, T3, Ts, F8, T4 and T 6 o f t h e international 10-20 system, and all electrodes were referenced to the ipsilateral ear electrode (A~ or A2). The 16-channel EEGs were recorded by an electroencephalograph with a time constant of 0.3 sec and a high frequency cut-off at 120 Hz ( - 3 dB) at a paper speed of 3 cm/sec and a sensitivity of 0.5 cm/50 ~V. Five artifact-free strips of EEG 10 sec long were selected from the entire record of the above montage for each subject and were subjected to visual inspection by a qualified EEGer (one of the authors, H.S.) as well as to the source of the automatic interpretation. Later EEG records of 3 other subjects were subjected to
(2) Visual interpretation of EEGs by a qualified EEGer The procedure for visual inspection of EEGs was sequentially categorized into 16 items and was arranged in order of interpretation (Table I). The 50 sec long EEG record of each subject was inspected quantitatively according to the following criteria (Table I), without any prior knowledge of the clinical picture of each subject. First of all, the dominant rhythm was defined as a dominant EEG activity that consisted of waves with an approximately constant period and with the maximum amplitude at the occipital or parieto-occipital regions of the head. As the first step, the dominant rhythm was checked as to whether it was recognizable or absent. If the dominant rhythm was identified, it was then analysed with respect to its organization, frequency and amplitude. Differences of these 3 quantities of the dominant rhythm between the corresponding sites of the two hemispheres were defined as 'asymmetry.' Extension of the dominant rhythm to anterior regions of the head was also taken into account. For beta rhythm, the amplitude and asymmetry in its amount within the band (13 Hz <) at each electrode were taken into account. As regards the waves of theta (4-8 Hz) and delta ( < 4 Hz) frequency and the activity
TABLE I Items and criteria (threshold values) for grading each item used for the present visual interpretation of EEG. Item
Normal
Abnormal Mild
Moderate
Marked
Score 0
Score 1
Score 2
Score 3
Yes 1 0.3 < 8< 0.5 < 100< 50 < till F, AT
Yes
No
0 < 0.3 9< < 0.5 < 100 < 50 till C, MT
Dominant rhythm Existence Organization Asymmetry (%) Frequency (Hz) Asymmetry (Hz) Amplitude (/zV) Asymmetry (%) Extension (/zV)
< 0.6 <9 < 1.0 < 130 < 60
2 0.6 < 6< 1.0 < 130< 60_< till Fp (low)
< 100 < 60
< 100 60 <
< 80
80<
5<
<50
50<
< 1.0 < 8 < 2.0 < 80
3 1.0 < <6 2.0 < 80__< till Fp (high)
Beta rhythm Amplitude (/zV) Asymmetry (%)
< 50 < 50
50 < 50 <
Theta rhythm Duration (%) Electrodes
0
<5 Active electrodes
Delta rhythm Duration (%) Electrodes
0
-
<
50
50<
Active electrodes
Non-dominant alpha rhythm Duration (%) Electrodes
< 10
10< < 30 Active electrodes
30<
< 75
75 <
AUTOMATIC EEG INTERPRETATION
425
of alpha frequency (8-13 Hz) other than the dominant rhythm, which was defined as 'non-dominant alpha rhythm,' the proportional duration of the respective rhythm at each electrode was estimated and expressed as a percentage, and the electrodes active for each component were listed. Every item of EEG was graded into 4 scores: normal (0), mildly abnormal (1), moderately abnormal (2) and markedly abnormal (3). The criteria (threshold value) for each score were quantitatively specified for each item, as shown in Table I.
(3) Computation of periodogram parameters The EEG records that were interpreted by the EEGer as described above were divided into 10 segments of 5 sec duration each and were subjected to the automatic interpretation. First, numeral time series of each EEG segment were obtained by transforming the actual EEG record into dotted data by using an image scanner (NEC PC-In503G) at a sampling interval (ordinate period) of 9.4 msec and a sampling amplitude of 2.8/xV. The data were input to a personal computer NEC PC-9801 RX, and again transformed into digital data by a computer program made by the authors (Nakamura et al. 1989a). The digital EEG time series at the sampling interval of 18.4 msec was adopted for the following analyses, because the Nyquist frequency
27 Hz was enough for analysis of awake EEGs except for spikes. The digital data of the EEG time series were then transformed into the Fourier components by the FFT method (Cooley and Tukey 1965), and the periodogram of each segment was obtained for each of the 16 channels as the squares of the Fourier components for each resolving frequency of 0.2 Hz (see Figs. 1 and 2). Features of the periodogram for each 5 sec EEG segment for each individual channel were expressed by the periodogram parameters for each of the 5 frequency bands: the dominant rhythm (d), and beta (/3), theta (0), delta (3) and the non-dominant alpha (a) bands. The periodogram consisted of 11 parameters: the amount of the periodogram within the respective band (So, St3, So, S~ and S,~), the peak frequency fro, ft3, f0, f~ and f~), and the standard deviation of the alpha rhythm (~,,). Thus, the time and spatial features of the 16-channel EEG record of 50 sec duration for each subject were condensed into 1760 periodogram parameters (11 parameters × 16 channels × 10 segments).
(4) Construction of EEG parameters By using the periodogram parameters thus obtained, EEG parameters were constructed so that they could
TABLE II EEG parameters for each item employed for the present automatic interpretation. Item
EEG parameters
Dominant rhythm Existence Organization Asymmetry (%) Frequency (Hz) Asymmetry (Hz) Amplitude (tzV)
Sd//S T >__0.1, S d / S a > 1.0, 10 g~/2 > 10.0 Yo = 0.49 + 0.58t7~ -0.13 S~ + 4.82 X 10 -5 (Sa) 2 - 0.41S~/S T +3.12 S~/S T lYd (X2)--Yd (Xl)l fd Ifd (X2)- fd (Xl)] 10 ~1/2
Asymmetry (%) Extension (/~V)
{Sd (X2)I/2 --Sd (Xl)l/2}//max{Sd (X2)1/2, Sd (X1)l/2}X 100 10 S~,/2
Beta rhythm Amplitude (pN) Asymmetry (%)
6
S~/, 2 >_ 10.0
{8/3, (X2)-5/3, (Xl)l/max{St~, (X2) , So, (Xl)lX 100
Theta rhythm Duration (%) Electrodes
((S 0 / S T) x 10016 Sl0/2 ~ 15l Active electrodes
Delta rhythm Duration (%) Electrodes
I(S~/S T) × 10016 S 1/2 > 25} Active electrodes
Non-dominant alpha rhythm Duration (%) Electrodes
Residual of dominant rhythm {(S,~,/S T) × 10016 S~/,2 > 15} Active electrodes
S d, stands for the amount of the periodogram within the dominant at the bipolar lead (Fp-F), St3, for the amount of the periodogram which excludes the higher harmonics, S~, for the amount of the non-dominant alpha rhythm.
426
M. N A K A M U R A E T AL.
closely conform to the procedures of visual EEG interpretation by the EEGer (Table II). First, the dominant rhythm, as defined for visual interpretation, was expressed by the following 3 equations: S d / S T >_0.1
(1.1)
Sd/Sd> 1.0
(1.2)
I0 S~OO _> I0.0 (/zV).
(1.3)
mum amount of the periodograms within the dominant band at all electrodes other than O and P (Fp, F, C, T) was designated as Sa. Therefore eqn. 1.2 fulfills the spatial dominancy of the rhythm over the posterior regions of the head. The threshold amplitude for the existence of the dominant rhythm was expressed by eqn. 1.3. The threshold value 10 (/xV) and the following threshold values for other frequency bands were given by the EEGer based on his standard criteria for visual EEG interpretation. If all of the above 3 criteria were fulfilled, the dominant rhythm was judged to exist, and then the features of the dominant rhythm were evaluated by using the EEG parameters for the organization (y_z~), the peak frequency (fd) and the amplitude ( I O V S d ) (Table II). The 'organization' is the most important measure of the dominant rhythm (Nakamura et al. 1985) and is defined by the International Federation of Societies for Electroencephalography and Clinical Neurophysiology as 'the degree to which physiologic EEG rhythms
The peak frequencies in the interval 7.2-13 Hz were detected at O~ and O 2, respectively. The amount of the periodograms within ___1 Hz around the respective peak frequency was measured at the electrodes O and P on both hemispheres separately. The amount at either the electrode O or P, depending on whichever was larger, for each hemisphere was designated as Sd for that hemisphere, and the rhythm in question was determined to be dominant in time when the Sd was 10% or more of the total amount (ST) of the periodogram at that electrode (eqn. 1.1). Then the maxiPeriodograms
EEG time series FPI-A
EEGer F3-A
Value
Item
Score
Lt i Rt
C3-A P3-A O,-A
L ..... F,-A2
~
C,-A2
~
L....At~tA~. ,..
P,-A~ 0 2 - Az ~
'
Ts-A,
_ . .._ . _
L,
.
,
olo
o
010
o
MiId ab.
i
1.5:1.9
2
Asymmetry [%]
0.0
0
-0.3
Frequency [Hz]
10.3i 10.3
0
Asymetry [Hz]
0.0
0
-0.2
0
Amplitude [~V]
30.0130.0
0
34.0!28.7
0
As~met ry [%]
0.0
0
15.5
0
Extension [#xV]
till P
0
15.6iJ l&3
0
~plitude [/xV]
<15.0
0
11.3
== Asymmetry [%]
0.0
0
9.3
0 0
2
17.4
[%]
Duration
Duration T , - A2 ~ T6 - A
[%J
Electrodes
~
L 2
~
_
..
L._.:...
[%]
Electrodes
.
Score
Organization
Electrodes A , "~'~'~
Value Lt i Rt
Existence
Duration
~
F--A, T3-
Aut~atic
5.0 Fp
F
C
POT 15. 0 epis. 1 2 Fp F C P 0 T
occasional ly more predominant on R 20.0 j 1 Fp F C bilateral
i
9.5i 9.7
Fp
0
2
F
C
POT 13.4 l 2 Fp2 F3 E, Ec3P'C, P~ P, O, F7 F, T3 T, 16.7 I 1 Fp F C POT
.
loofu v]]
z'[-5 50[u V2/Hz] Time [see]
Frequency [Hz]
Fig. 1. The time series of a 5 sec segment of 16-channel E E G and corresponding periodograms in subject 1. Estimates of the E E G parameters and grading scores were obtained by the automatic E E G interpretation (right) and by the E E G e r ' s visual inspection (left) based on a 50 sec long E E G record in this and the following figures.
AUTOMATIC EEG INTERPRETATION
427
conform to certain ideal characteristics displayed by a proportion of subjects in the same age group, free from personal and family history of neurologic and psychiatric diseases, and other illnesses that might be associated with dysfunction of the brain' (Chatrian et al. 1974). The following equation was adopted as the function for scoring the organization;
absolute value of the difference between the two sides. The quantity 'frequency' was expressed by the peak frequency fro) of the periodogram parameter. The 'asymmetry' for the frequency was defined as an absolute value of the difference. As the quantity 'amplitude' used in the EEGer's interpretation was in fact the median voltage of the EEG waves measured peakto-peak, it was expressed a s 10~/Sd . The 'asymmetry' for the amplitude was defined as a relative ratio in percentage of the measurement on two sides. In order to eliminate the effect of the active ear-lobe reference, 'extension' of the dominant rhythm to anterior regions of the head was determined from the bipolar derivation which was computed from the original referential derivation. The 'extension' was represented by the amount of the periodogram (S~) within the same dominant band at the bipolar lead (Fp-F) on each hemisphere (Table II). As the wave form of the dominant alpha rhythm is usually sharp in the negative phase (upward deflection in the present derivation) and blunt in the positive
Yd = 0 . 4 9 + 0 . 5 8 (7~,--0.13 S,, + 4 . 8 2 × 10 -5 ( ~ . ) z - 0 . 4 1 S~/ST + 3 . 1 2 S a / S T
(2)
This equation for scoring the organization of the dominant rhythm is a regression of 5 factors (t~, S~, (~,)2, S , / S T and S J S T) measured at either electrode O or P, depending on to which electrode So belongs; it was previously derived by the authors based on visual inspections of EEG records by 5 neurlogists (Nakamura et al. 1989b). It should be noted that parameters used in evaluating the organization were those of the a band (8-13 Hz), but not those of the dominant band (within + 1 Hz around the peak frequency). The quantity 'asymmetry' for the organization was defined as an
EEG time series
Periodograms
FpI~A
F
-A
C
-A
P
-A
EEGer 1~
k A .=- ~ - - ~
.
Item
Value
Existence
Lti
Rt
O:
x
Automatic Valee
Score
Score
Lt iRt
0
O:
Mode ab. R
0
1.6 :;---
i
Lack R
3
0.0
3
Frequency [Hz] 10.71 - - -
0
10.71---
Fp2-A
Asymmetry [Hz]
Lack R
3
0.0
3
F
-A
Amplitude[xtV] 40.0 i Asymmetry [%] Lack R
0
39.6 I - - - ;
0
3
0.0
3
C
-A
Extens ion [~uV]
0
IS.Ti---
0 0
O
P
Organization Asymmetry [%]
-A
-A
~-
F~-A
<20. 0
0
13.2
Asymmetry [%]
60. 0
2
15. 1
0
2
12.0
2
[%3 15.0
Electrodes Duration
T3-A
[%]
Electrodes
T~-A Duration
[%]
(L>R)
Right hemisphere
90.0
Electrodes T,-A
C~F°'C, Ep2 F3 F, P3 P, O, 02 F7 F, T, Ts Ts
{ 3
44. 7
T6 02 P, 20. 0
I
1
Right hemisphere
13.0 t I FD, F~2 F3 F, C3 C, P3 P, O, F7 Fs T3 Ts T6
50[/z V2/Hz]
0
s"0
; Time [see]
3
C, P3 P, O~ 02 F8 %
T6-A
r
t
F,~ F~2 F3 F4
Fa-A
lootu v]]
0
~Jplitude[xtV] Duration
O~-A
till C(L)
0
x
1'0 15 2'0 15
Frequency [Hz]
Fig. 2. T h e time series of a 5 sec segment of 16-channel E E G and corresponding periodograms in subject 2.
428
M. NAKAMURAET AL.
phase (downward deflection), higher harmonics of the rhythm can appear in the Fourier analysis. Consequently, the second harmonic of the dominant rhythm may occur in the beta band (over 13 Hz). Therefore, the amount of the periodogram in the beta band (S t) was obtained by summing up the periodogram within that band while excluding the amount of + 1 Hz around twice the peak frequency of the dominant rhythm (2 fd). The threshold value for the existence of the beta rhythm was given as 6S~-~ >_ 10 (/zV) (Table II). The features of the beta rhythm were evaluated using the EEG parameters, namely the amplitude in ~V and the asymmetry of the modified amounts in percentage. The threshold values for the existence of theta rhythm and delta rhythm were set as 6S~0 > 15 (~V) and 6 S ~ > 25 (izV), respectively. Alpha rhythm excluding the dominant rhythm (non-dominant alpha rhythm) was expressed by any rhythm in the alpha band (8-13 Hz) other than the dominant rhythm. The threshold value for the existence of the non-dominant alpha rhythm was given by 6~/~-~> 15 (/zV) (Table II). The features of these 3 rhythms were evaluated in
terms of their durations in percentage and electrodes active for respective rhythms. As the quantity 'duration' in the EEGer's interpretation was found to be the relative appearance rate of the respective wave components, it was expressed by {E(S/ST)10} × 100 (%), where S represents the amount of the periodogram of that particular wave component.
(5) Automatic EEG interpretation EEG parameters were computed for each segment data for each of the 14 subjects by using the periodogram parameters obtained in section 3 and the formulae as described above (Table II), and the final estimates of the EEG parameters were obtained by averaging the EEG parameters of all 10 segments. Then the scores for each item were calculated based on the criteria (threshold values) which were employed for the visual inspection of each item (Table I) and were compared with the results obtained by the EEGer's visual inspection for each subject. Finally, EEG data of 3 subjects, which were not included in the data used for development of the present system, were
Periodograms
EEG time series FPL-- A~,N, ~
Itern C P
3
-
3
A
-
O
,
A
,
~
~
-
A
,
L~..,,.a4._:
~
~ - ~ F , - A 2'
.
,
C, - A
P
2
,
O
~
,
,
'
2
v
.
~
~
2 A
' z
~
L,~/D~ L ~. . . . . . . L,.,.
~
A -
.
-
,~.~
. ~
~ ~
~
0 i
Asymmetry [%] 0.0 Frequency [Hz] 10.s i x0.s Asymmetry [Hz] 0.0
0 o 0
O
Amplitude[~V] 95.0 i 95.0 Asymetry [%] 0.0 Extension[~V] till C ABelitude[/JV] <35.0 Asymmetry [%] 0.0 Duration [%] 20. 0
0
O
0.0 10,3110.2 0.1 58.6;:57.6 -1.8 25.7!45.1 18.8
0
-3.0
2
14.8
O
O
t
Diffuse ,
s
-
,
-
A
A
~
2
~
~
T
,
-
A
~
~
T
,
-
A
2
~
xoo[.vl] 0
.....
2
[~ -.
~,,.-,- .
. ..
~
Electrodes
~
Duration [%]
'~'
more on R I
C
POT
40.0 I 2 !Fp F C P 0 T 0.0
0 0 O 0 1 O 0 2
Fp F
more on R
30. 5
g
T
F
Modeab.
Duration [%]
T3-At
Lti et
©:© 1.2 ! 1.3
Electrodes
O:O
Automtic Value Score
o 2
Existence Organization
~
EEGer Value Score Lt i Rt
I
2
m
Fp F
C
POT
O
Electrodes
L
.....
s i0 15 20 25
Time [sec] Frequency [8z] Fig. 3. The time series of a 5 sec segmentof 16-channelEEG and correspondingperiodogramsin subject3.
23,4 Fp F POT
I
I c
AUTOMATIC EEG INTERPRETATION
subjected to the automatic interpretation, and the results were compared with those of the visual interpretation. The algorithm for the present automatic EEG interpretation was programmed in C language and implemented by using a personal computer NEC PC-9801 RA with a DSP board for the FFT computation. Actually, transformation of the image data of the EEG record into digital data required a processing time of about 1 h for 1 subject. If the digital data of the EEG record are directly taken into the computer, the transformation time can be saved, and then the total processing time for the automatic EEG interpretation is only about 5.5 min: the data input time 2.5 min, the main calculation time 1 min and the print-out time 2 min. Results
The grading scores obtained for each item by the present automatic interpretation system were similar to those judged by the EEGer in most of the 14 subjects. Conformity between the results of the automatic interpretation and those of the visual interpretation was summarized in Table III. Scores obtained by the two methods were exactly the same in 78.6% of all items, different by 1 point in 13.7%, by 2 points in 4.4% and by 3 points in 3.3%. Figs. 1-3 show results of interpretation of EEGs from the 3 additional subjects. The EEG parameters and grading scores obtained by the automatic EEG interpretation are listed in the right two columns of the table in each figure, and the corresponding results obtained by the EEGer's visual inspection are shown in the left two columns. A 5 sec long segment of the EEG record and the corresponding periodograms for each channel are shown in each figure. Subject 1 was a 30-year-old woman presenting with focal sensorimotor seizure starting in the left hand of recent onset. On visual inspection of the EEG (Fig. 1 and a further 45 sec length of the EEG), the frequency of the dominant rhythm was estimated to be 10.3 Hz on average and symmetrical, but its organization was judged to be mildly abnormal (score 1). There was no abnormality noted with regard to the beta rhythm. Occasional theta waves were recognized diffusely, the duration of which was estimated to be 5% of the total recording time (score 2). There were episodic large amplitude delta waves diffusely which were occasionally more predominant on the right hemisphere (not included in the segment shown in Fig. 1); their proportional duration was estimated to be 15% of the total recording time (score 2). There were occasional rhythmic alpha waves other than the dominant rhythm anteriorly and bilaterally, the duration of which was estimated to be 20% of the total recording time (score 1).
429
Automatic interpretation of the same record revealed the mean frequency of the dominant rhythm to be 9.5 Hz on the left and 9.7 Hz on the right. Its organization was found to be slightly more abnormal than estimated from the visual inspection (score 2), and slightly asymmetric, being more abnormal on the right. Otherwise the same findings as judged from the visual inspection were revealed for the dominant rhythm. The proportional duration of theta waves was found to be slightly longer (17.4%) as compared with the visual inspection, but it did not make any difference to the score. The automatic interpretation detected the same amount of delta waves, although it failed to detect an occasional asymmetry which was observed on visual inspection. Activities of alpha frequency other than the dominant rhythm were found in the same amount diffusely whereas they were found only anteriorly on visual inspection. Subject 2 was a 56-year-old woman with a glioblastoma in the right parietal lobe. On visual inspection of the EEG (Fig. 2), the dominant rhythm was judged to be present on the left, but absent on the right (score 3 for asymmetry of the organization, frequency and amplitude). The dominant rhythm on the left hemisphere was judged to be normal in frequency (10.7 Hz), amplitude (40 /xV) and organization. Beta activities were judged to be less in quantity on the right hemisphere
T A B L E III Difference between the scores obtained by automatic interpretation and visual inspection in n u m b e r of records (total n u m b e r of subjects 14). Item
Difference of scores None
1
2
3
14 10 11 13 12 14 11 12
4 2 1 1 0 1 1
0 0 0 0 0 0 1
0 0 1 0 1 0 2 0
14 9
0 1
0 2
0 2
7
6
1
0
10
0
4
0
6
8
0
0
Dominant rhythm Existence Organization Asymmetry (%) Frequency Hz) Asymmetry (Hz) Amplitude (/~V) Asymmetry (%) Extension (tzV)
Beta rhythm Amplitude (~V) Asymmetry (%)
Them rhythm Duration (%) Electrodes
Delta rhythm Duration (%) Electrodes
Non-dominant alpha rhythm Duration (%) Electrodes
430
than on the left by 60% (score 2). Theta waves were occasionally (15% of the total recording time) seen diffusely on the right hemisphere (score 2). Continuous irregular delta waves were seen localized in the right posterior quadrant of the head (P4, 02 and T6). The automatic interpretation of the same record revealed essentially the same findings as the visual inspection with regard to the dominant rhythm, except that the dominant rhythm on the left hemisphere was also found to be slightly abnormal in organization (score 1). As to the beta waves, however, the automatic interpretation did not find any significant asymmetry. The same amount of theta waves was found by automatic interpretation, but diffuse instead of with rightsided predominance which was observed on visual inspection. Slow waves of delta frequency were found with a proportional duration of about half of what was estimated by visual inspection (44.7% vs. 90%), and more diffusely compared with visual inspection. Subject 3 was a 48-year-old woman in the post-operative state for a ruptured aneurysm arising from the right anterior cerebral artery. On visual inspection of the E E G (Fig. 3), the dominant rhythm was judged to be present bilaterally and symmetrical in terms of frequency and amplitude. However, its organization was judged to be moderately abnormal bilaterally (score 2). Beta rhythm was judged to be present equally on both hemispheres. Slow waves of theta frequency were judged to be present diffusely with a proportional duration of 20% (score 2). D~ Ita waves were judged to be diffuse, more on the right hemisphere, at 40% of the total recording time, including occasional paroxysmal bursts of large amplitude rhythmic delta waves which were seen bilaterally and more anteriorly. Prominent slow waves seen at Fp~ and Fp 2 were judged to be blink artifacts, based on their wave forms and restricted location. The automatic interpretation revealed essentially the same findings as regards the dominant rhythm, except that the organization was found to be mildly abnormal (score 1). Theta and delta waves were found diffusely in similar amounts to those estimated by visual inspection, but the slight asymmetry of slow waves observed on visual inspection was not found. Alpha waves other than the dominant rhythm were detected diffusely but were not observed on visual inspection. The slow waves limited to the fronto-polar electrodes were not taken as abnormal E E G activities by the automatic interpretation.
Discussion In the present novel automatic E E G interpretation system, first the items for the E E G interpretation were
M. N A K A M U R A ET AL.
determined so that they could conform to the procedure that a qualified E E G e r usually follows for visual inspection of the actual record, and all items were defined quantitatively. For automatic interpretation, the E E G parameters were chosen so that they could represent the basis of the EEGer's visual interpretation as closely as possible. The interpretation thus obtained was demonstrated to give similar results to those derived from visual interpretation. The usefulness of this system was confirmed by applying it to new E E G records which were not taken into account for developing the system. Since the results of automatic E E G interpretation were compared with those obtained through visual inspection by the same E E G e r upon whose own procedure of visual inspection the E E G parameters were based, it might be argued that the conformity between the two methods was predictable. However, the purpose of the present study was not to make a standard criterion for E E G interpretation, but to make an automatic analysis system which resembles the EEGer's visual inspection as closely as possible. Therefore, the present system is judged to be satisfactory for that purpose. In other words, another system can be made if another E E G e r wishes to make a different system which conforms to his own procedure of visual inspection. To use different threshold values given by another EEGer, this automatic system could be adaptable to the new system merely by changing the threshold values. Furthermore, by taking an appropriate weighted sum of the obtained scores, it would be possible to provide an automatic integrative interpretation of an awake E E G record just like the final summary derived by visual inspection. Upon detailed observation of the results, however, we found slight discrepancies in some items between the automatic interpretation and the EEGer's visual inspection. The causes of the differences were found to be two-fold: an artifact of eye blinks and activation of the reference electrodes. Such disagreements will be reduced by an appropriate preprocessing of E E G data for blinks, and by using a non-cephalic reference, for example, by employing an ECG elimination method (Nakamura and Shibasaki 1987) to skirt the problem of the earlobe activation. In contrast with the previous automatic analyses of E E G (Kowada et al. 1971; Yamamoto et al. 1975; Nakamura et al. 1985, 1989b), which were focussed on certain aspects such as the dominant rhythm or slow waves, the present system is unique in its capability to make an integrative interpretation of the spontaneous awake E E G by taking into account all its features except for paroxysmal abnormalities. Therefore, this system can be used clinically as an assistant tool for neurologists and EEGers and may be the first step toward its clinical application to hospital use.
AUTOMATIC EEG INTERPRETATION The authors are grateful to Mr. I. Mitsuyasu (postgraduate student) of Saga University for his help in computer computation.
References Chatrian, G.E., Bergamini, L., Dondey, M., Klass, D.W., LennoxBuchthal, M. and Peters~n, I. A glossary of terms most commonly used by clinical electroencephalographers. Electroenceph. clin. Neurophysiol., 1974, 37: 538-548. Cooley, J.W. and Tukey, J.W. An algorithm for the machine calculation of complex Fourier series. Math. Comput., 1965, 19: 297-301. Frost, Jr., .I.D. Automatic recognition and characterization of epileptiform discharges in the human EEG. J. Clin. Neurophysiol., 1985, 2: 231-249. Gotman, J. Practical use of computer-assisted EEG interpretation in epilepsy. J. Clin. Neurophysiol., 1985, 2: 251-265. 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. Inoue, K., Kumamaru, K., Sagara, S. and Matsuoka, S. Pattern recognition approach to human sleep EEG analysis and detection of sleep stages. Mem. Fac. Eng. Kyushu Univ., 1982, 42: 177-195. Kowada, M., Sato, S., Hiraga, H., Tomono, T. and Yahata, K. Automatic pattern recognition of clinical EEG. I. A method for
431 wave form recognition. Clin. Electroenceph., 1971, 13: 647-656. (In Japanese.) Nakamura, M. and Shibasaki, H. Elimination of EKG artifacts from EEG records: a new method of non-cephalic referential EEG record. Electroenceph. clin. Neurophysiol., 1987, 66: 89-92. Nakamura, M., Nishida, S., Neshige, R. and Shibasaki, H. Quantitative analysis of 'organization' by feature extraction of the EEG power spectrum. Electroenceph. clin. Neurophysiol., 1985, 60: 84-89. Nakamura, M., Nishida, S. and Imajoh, K. Deterioration of power spectrum for quantized time series and its recovering method. Preprints of the 8th SICE Conf. Kyushu Branch, 1989a, 8: 155158. (In Japanese.) Nakamura, M., Shibasaki, H. and Nishida, S. Automatic scoring system of EEG and quantitative evaluation of its visual interpretation. Proc. '89 KACC, 1989b: 967-971. Nishida, S., Nakamura, M. and Shibasaki, H. An EEG model expressed by the sinusoidal waves with the Malkov process amplitude. Jpn ,i, Med. Electron. Biol. Eng., 1986, 24: 8-14. (In Japanese.) Yamamoto, K., Nakamura, N., Shimazono, Y., Miyasaka, M. and Fukuzawa, H. The system construction on a newly developed automatic EEG diagnosing system - with special regard to the wave-form recognition method. Psychiat. Neurol. Jap., 1975, 77: 127-159. (Abstract in English.)
423
© 1992 Elsevier Scientific Publishers Ireland, Ltd. 0013-4649/92/$05.00
EEG91056
Automatic EEG interpretation: a new computer-assisted system for the automatic integrative interpretation of awake background EEG Masatoshi Nakamura, Hiroshi Shibasaki a, Kaoru Imajoh, Shigeto Nishida, Ryuji Neshige b and Akio Ikeda a Department of Electrical Engineering, Saga University, Honjomachi, Saga 840 (Japan), a Department of Brain Pathophysiology, Kyoto University Faculty of Medicine, Shogoin, Sakyo-ku, Kyoto 606 (Japan), and b Department of Internal Medicine, Saga Medical School, Nabeshima, Saga 849 (Japan) (Accepted for publication: 9 December 1991)
Summary A new computer-assisted system for automatic interpretation of the awake electroencephalogram (EEG) was developed. First, all the items necessary for EEG interpretation were determined in accordance with the procedure that a qualified electroencephalographer (EEGer) goes through for the visual inspection of the background EEG activity, and then each item was defined quantitatively. For the automatic interpretation, specific EEG parameters were determined for each item so that they could fit the graded judgement of the item by the qualified EEGer as closely as possible. These specific EEG parameters were actually calculated from periodograms obtained from the time series of EEG records of 14 patients with various neurological diseases. The automatic EEG interpretation system thus established was applied to the EEG data of these 14 subjects and to 3 additional EEGs, and the results were compared with those obtained through the visual interpretation by the EEGer. This automatic EEG interpretation was found to be in good agreement with the visual interpretation by the EEGer in most EEG records. In contrast with the previous automatic analyses of EEG which were focussed on certain aspects of EEG such as the dominant rhythm, the present system is unique in its capability of providing an integrative interpretation of the spontaneous awake EEG by taking into account all its features except for paroxysmal abnormalities. Key words: Automatic EEG interpretation; Awake EEG; Quantitative scores of EEG
The electronencephalogram (EEG) is a summation of electrical activities generated by nerve cells in the cerebral cortex and recorded from the scalp. As the EEG reflects at least a part of the functional status of the subject's brain, the EEG interpreted by electroencephalographers (EEGers) is used as a diagnostic aid in diseases affecting the brain. When an EEGer interprets an EEG record, he analyses its time and spatial features by visual inspection. Automatic analyses of EEG have been attempted by many investigators. Among various kinds of automatic EEG interpretation previously reported, spike detection in either the scalp-recorded or depth-recorded EEG in epileptic patients has drawn particular attention (Gotman and Gloor 1976; Frost 1985; Gotman 1985). Automatic analyses of sleep stages by using EEG data have also been implemented, and an accuracy of about 80% in the recognition was achieved (Inoue et al. 1982). Auto-
Correspondence to: Dr. Masatoshi Nakamura, Department of Electrical Engineering, Saga University, Honjomachi, Saga 840 (Japan).
matic diagnosis of awake EEGs has been investigated by Kowada et al. (1971) and Yamamoto et al. (1975). However, most previous studies were restricted to a certain aspect of EEG characteristics such as the dominant rhythm or slow waves. An automatic integrative interpretation of a full EEG record has not yet been accomplished, mainly because a large number of complex items are involved in the interpretation of the EEG. In previous works the authors (Nakamura et al. 1985, 1989b) succeeded in an automatic scoring of 'organization' of the dominant or basic rhythm of EEG activity which is one of the most essential issues for the interpretation of the spontaneous awake EEG obtained while the subject is at rest with the eyes closed. In the present study, we further advanced our previous 'organization' scoring method to accomplish an automatic integrative interpretation of the awake EEG record, taking into account all items necessary for EEG interpretation except for paroxysmal abnormalities such as spikes and spike-and-waves. This system can be used clinically as an assistant tool for neurologists and even for EEGers themselves, and may be the first step for its clinical application to hospital use.
424
M. NAKAMURA ET AL.
Methods
automatic interpretation so that the results can be compared with those derived from visual inspection by the EEGer.
(1) Data acquisition EEGs of 14 subjects recorded on paper were arbitrarily selected from consecutive EEG records of Saga Medical School for development of the present system for automatic EEG interpretation. These EEGs were recorded from 14 patients, aged 20-64 years, with various neurological diseases. The diagnosis of these diseases is not listed here because it does not matter for the purpose of the present study, although all patients needed EEG examination for diagnostic purposes. All EEGs were recorded in a quiet, dimly lit room where the subject was placed in a supine position in a bed with the eyes closed. Exploring cup electrodes were fixed to the scalp at points Fp a, F3, C3, P3, O1, FP2, F4, C4, P4, 0 2 , F7, T3, Ts, F8, T4 and T 6 o f t h e international 10-20 system, and all electrodes were referenced to the ipsilateral ear electrode (A~ or A2). The 16-channel EEGs were recorded by an electroencephalograph with a time constant of 0.3 sec and a high frequency cut-off at 120 Hz ( - 3 dB) at a paper speed of 3 cm/sec and a sensitivity of 0.5 cm/50 ~V. Five artifact-free strips of EEG 10 sec long were selected from the entire record of the above montage for each subject and were subjected to visual inspection by a qualified EEGer (one of the authors, H.S.) as well as to the source of the automatic interpretation. Later EEG records of 3 other subjects were subjected to
(2) Visual interpretation of EEGs by a qualified EEGer The procedure for visual inspection of EEGs was sequentially categorized into 16 items and was arranged in order of interpretation (Table I). The 50 sec long EEG record of each subject was inspected quantitatively according to the following criteria (Table I), without any prior knowledge of the clinical picture of each subject. First of all, the dominant rhythm was defined as a dominant EEG activity that consisted of waves with an approximately constant period and with the maximum amplitude at the occipital or parieto-occipital regions of the head. As the first step, the dominant rhythm was checked as to whether it was recognizable or absent. If the dominant rhythm was identified, it was then analysed with respect to its organization, frequency and amplitude. Differences of these 3 quantities of the dominant rhythm between the corresponding sites of the two hemispheres were defined as 'asymmetry.' Extension of the dominant rhythm to anterior regions of the head was also taken into account. For beta rhythm, the amplitude and asymmetry in its amount within the band (13 Hz <) at each electrode were taken into account. As regards the waves of theta (4-8 Hz) and delta ( < 4 Hz) frequency and the activity
TABLE I Items and criteria (threshold values) for grading each item used for the present visual interpretation of EEG. Item
Normal
Abnormal Mild
Moderate
Marked
Score 0
Score 1
Score 2
Score 3
Yes 1 0.3 < 8< 0.5 < 100< 50 < till F, AT
Yes
No
0 < 0.3 9< < 0.5 < 100 < 50 till C, MT
Dominant rhythm Existence Organization Asymmetry (%) Frequency (Hz) Asymmetry (Hz) Amplitude (/zV) Asymmetry (%) Extension (/zV)
< 0.6 <9 < 1.0 < 130 < 60
2 0.6 < 6< 1.0 < 130< 60_< till Fp (low)
< 100 < 60
< 100 60 <
< 80
80<
5<
<50
50<
< 1.0 < 8 < 2.0 < 80
3 1.0 < <6 2.0 < 80__< till Fp (high)
Beta rhythm Amplitude (/zV) Asymmetry (%)
< 50 < 50
50 < 50 <
Theta rhythm Duration (%) Electrodes
0
<5 Active electrodes
Delta rhythm Duration (%) Electrodes
0
-
<
50
50<
Active electrodes
Non-dominant alpha rhythm Duration (%) Electrodes
< 10
10< < 30 Active electrodes
30<
< 75
75 <
AUTOMATIC EEG INTERPRETATION
425
of alpha frequency (8-13 Hz) other than the dominant rhythm, which was defined as 'non-dominant alpha rhythm,' the proportional duration of the respective rhythm at each electrode was estimated and expressed as a percentage, and the electrodes active for each component were listed. Every item of EEG was graded into 4 scores: normal (0), mildly abnormal (1), moderately abnormal (2) and markedly abnormal (3). The criteria (threshold value) for each score were quantitatively specified for each item, as shown in Table I.
(3) Computation of periodogram parameters The EEG records that were interpreted by the EEGer as described above were divided into 10 segments of 5 sec duration each and were subjected to the automatic interpretation. First, numeral time series of each EEG segment were obtained by transforming the actual EEG record into dotted data by using an image scanner (NEC PC-In503G) at a sampling interval (ordinate period) of 9.4 msec and a sampling amplitude of 2.8/xV. The data were input to a personal computer NEC PC-9801 RX, and again transformed into digital data by a computer program made by the authors (Nakamura et al. 1989a). The digital EEG time series at the sampling interval of 18.4 msec was adopted for the following analyses, because the Nyquist frequency
27 Hz was enough for analysis of awake EEGs except for spikes. The digital data of the EEG time series were then transformed into the Fourier components by the FFT method (Cooley and Tukey 1965), and the periodogram of each segment was obtained for each of the 16 channels as the squares of the Fourier components for each resolving frequency of 0.2 Hz (see Figs. 1 and 2). Features of the periodogram for each 5 sec EEG segment for each individual channel were expressed by the periodogram parameters for each of the 5 frequency bands: the dominant rhythm (d), and beta (/3), theta (0), delta (3) and the non-dominant alpha (a) bands. The periodogram consisted of 11 parameters: the amount of the periodogram within the respective band (So, St3, So, S~ and S,~), the peak frequency fro, ft3, f0, f~ and f~), and the standard deviation of the alpha rhythm (~,,). Thus, the time and spatial features of the 16-channel EEG record of 50 sec duration for each subject were condensed into 1760 periodogram parameters (11 parameters × 16 channels × 10 segments).
(4) Construction of EEG parameters By using the periodogram parameters thus obtained, EEG parameters were constructed so that they could
TABLE II EEG parameters for each item employed for the present automatic interpretation. Item
EEG parameters
Dominant rhythm Existence Organization Asymmetry (%) Frequency (Hz) Asymmetry (Hz) Amplitude (tzV)
Sd//S T >__0.1, S d / S a > 1.0, 10 g~/2 > 10.0 Yo = 0.49 + 0.58t7~ -0.13 S~ + 4.82 X 10 -5 (Sa) 2 - 0.41S~/S T +3.12 S~/S T lYd (X2)--Yd (Xl)l fd Ifd (X2)- fd (Xl)] 10 ~1/2
Asymmetry (%) Extension (/~V)
{Sd (X2)I/2 --Sd (Xl)l/2}//max{Sd (X2)1/2, Sd (X1)l/2}X 100 10 S~,/2
Beta rhythm Amplitude (pN) Asymmetry (%)
6
S~/, 2 >_ 10.0
{8/3, (X2)-5/3, (Xl)l/max{St~, (X2) , So, (Xl)lX 100
Theta rhythm Duration (%) Electrodes
((S 0 / S T) x 10016 Sl0/2 ~ 15l Active electrodes
Delta rhythm Duration (%) Electrodes
I(S~/S T) × 10016 S 1/2 > 25} Active electrodes
Non-dominant alpha rhythm Duration (%) Electrodes
Residual of dominant rhythm {(S,~,/S T) × 10016 S~/,2 > 15} Active electrodes
S d, stands for the amount of the periodogram within the dominant at the bipolar lead (Fp-F), St3, for the amount of the periodogram which excludes the higher harmonics, S~, for the amount of the non-dominant alpha rhythm.
426
M. N A K A M U R A E T AL.
closely conform to the procedures of visual EEG interpretation by the EEGer (Table II). First, the dominant rhythm, as defined for visual interpretation, was expressed by the following 3 equations: S d / S T >_0.1
(1.1)
Sd/Sd> 1.0
(1.2)
I0 S~OO _> I0.0 (/zV).
(1.3)
mum amount of the periodograms within the dominant band at all electrodes other than O and P (Fp, F, C, T) was designated as Sa. Therefore eqn. 1.2 fulfills the spatial dominancy of the rhythm over the posterior regions of the head. The threshold amplitude for the existence of the dominant rhythm was expressed by eqn. 1.3. The threshold value 10 (/xV) and the following threshold values for other frequency bands were given by the EEGer based on his standard criteria for visual EEG interpretation. If all of the above 3 criteria were fulfilled, the dominant rhythm was judged to exist, and then the features of the dominant rhythm were evaluated by using the EEG parameters for the organization (y_z~), the peak frequency (fd) and the amplitude ( I O V S d ) (Table II). The 'organization' is the most important measure of the dominant rhythm (Nakamura et al. 1985) and is defined by the International Federation of Societies for Electroencephalography and Clinical Neurophysiology as 'the degree to which physiologic EEG rhythms
The peak frequencies in the interval 7.2-13 Hz were detected at O~ and O 2, respectively. The amount of the periodograms within ___1 Hz around the respective peak frequency was measured at the electrodes O and P on both hemispheres separately. The amount at either the electrode O or P, depending on whichever was larger, for each hemisphere was designated as Sd for that hemisphere, and the rhythm in question was determined to be dominant in time when the Sd was 10% or more of the total amount (ST) of the periodogram at that electrode (eqn. 1.1). Then the maxiPeriodograms
EEG time series FPI-A
EEGer F3-A
Value
Item
Score
Lt i Rt
C3-A P3-A O,-A
L ..... F,-A2
~
C,-A2
~
L....At~tA~. ,..
P,-A~ 0 2 - Az ~
'
Ts-A,
_ . .._ . _
L,
.
,
olo
o
010
o
MiId ab.
i
1.5:1.9
2
Asymmetry [%]
0.0
0
-0.3
Frequency [Hz]
10.3i 10.3
0
Asymetry [Hz]
0.0
0
-0.2
0
Amplitude [~V]
30.0130.0
0
34.0!28.7
0
As~met ry [%]
0.0
0
15.5
0
Extension [#xV]
till P
0
15.6iJ l&3
0
~plitude [/xV]
<15.0
0
11.3
== Asymmetry [%]
0.0
0
9.3
0 0
2
17.4
[%]
Duration
Duration T , - A2 ~ T6 - A
[%J
Electrodes
~
L 2
~
_
..
L._.:...
[%]
Electrodes
.
Score
Organization
Electrodes A , "~'~'~
Value Lt i Rt
Existence
Duration
~
F--A, T3-
Aut~atic
5.0 Fp
F
C
POT 15. 0 epis. 1 2 Fp F C P 0 T
occasional ly more predominant on R 20.0 j 1 Fp F C bilateral
i
9.5i 9.7
Fp
0
2
F
C
POT 13.4 l 2 Fp2 F3 E, Ec3P'C, P~ P, O, F7 F, T3 T, 16.7 I 1 Fp F C POT
.
loofu v]]
z'[-5 50[u V2/Hz] Time [see]
Frequency [Hz]
Fig. 1. The time series of a 5 sec segment of 16-channel E E G and corresponding periodograms in subject 1. Estimates of the E E G parameters and grading scores were obtained by the automatic E E G interpretation (right) and by the E E G e r ' s visual inspection (left) based on a 50 sec long E E G record in this and the following figures.
AUTOMATIC EEG INTERPRETATION
427
conform to certain ideal characteristics displayed by a proportion of subjects in the same age group, free from personal and family history of neurologic and psychiatric diseases, and other illnesses that might be associated with dysfunction of the brain' (Chatrian et al. 1974). The following equation was adopted as the function for scoring the organization;
absolute value of the difference between the two sides. The quantity 'frequency' was expressed by the peak frequency fro) of the periodogram parameter. The 'asymmetry' for the frequency was defined as an absolute value of the difference. As the quantity 'amplitude' used in the EEGer's interpretation was in fact the median voltage of the EEG waves measured peakto-peak, it was expressed a s 10~/Sd . The 'asymmetry' for the amplitude was defined as a relative ratio in percentage of the measurement on two sides. In order to eliminate the effect of the active ear-lobe reference, 'extension' of the dominant rhythm to anterior regions of the head was determined from the bipolar derivation which was computed from the original referential derivation. The 'extension' was represented by the amount of the periodogram (S~) within the same dominant band at the bipolar lead (Fp-F) on each hemisphere (Table II). As the wave form of the dominant alpha rhythm is usually sharp in the negative phase (upward deflection in the present derivation) and blunt in the positive
Yd = 0 . 4 9 + 0 . 5 8 (7~,--0.13 S,, + 4 . 8 2 × 10 -5 ( ~ . ) z - 0 . 4 1 S~/ST + 3 . 1 2 S a / S T
(2)
This equation for scoring the organization of the dominant rhythm is a regression of 5 factors (t~, S~, (~,)2, S , / S T and S J S T) measured at either electrode O or P, depending on to which electrode So belongs; it was previously derived by the authors based on visual inspections of EEG records by 5 neurlogists (Nakamura et al. 1989b). It should be noted that parameters used in evaluating the organization were those of the a band (8-13 Hz), but not those of the dominant band (within + 1 Hz around the peak frequency). The quantity 'asymmetry' for the organization was defined as an
EEG time series
Periodograms
FpI~A
F
-A
C
-A
P
-A
EEGer 1~
k A .=- ~ - - ~
.
Item
Value
Existence
Lti
Rt
O:
x
Automatic Valee
Score
Score
Lt iRt
0
O:
Mode ab. R
0
1.6 :;---
i
Lack R
3
0.0
3
Frequency [Hz] 10.71 - - -
0
10.71---
Fp2-A
Asymmetry [Hz]
Lack R
3
0.0
3
F
-A
Amplitude[xtV] 40.0 i Asymmetry [%] Lack R
0
39.6 I - - - ;
0
3
0.0
3
C
-A
Extens ion [~uV]
0
IS.Ti---
0 0
O
P
Organization Asymmetry [%]
-A
-A
~-
F~-A
<20. 0
0
13.2
Asymmetry [%]
60. 0
2
15. 1
0
2
12.0
2
[%3 15.0
Electrodes Duration
T3-A
[%]
Electrodes
T~-A Duration
[%]
(L>R)
Right hemisphere
90.0
Electrodes T,-A
C~F°'C, Ep2 F3 F, P3 P, O, 02 F7 F, T, Ts Ts
{ 3
44. 7
T6 02 P, 20. 0
I
1
Right hemisphere
13.0 t I FD, F~2 F3 F, C3 C, P3 P, O, F7 Fs T3 Ts T6
50[/z V2/Hz]
0
s"0
; Time [see]
3
C, P3 P, O~ 02 F8 %
T6-A
r
t
F,~ F~2 F3 F4
Fa-A
lootu v]]
0
~Jplitude[xtV] Duration
O~-A
till C(L)
0
x
1'0 15 2'0 15
Frequency [Hz]
Fig. 2. T h e time series of a 5 sec segment of 16-channel E E G and corresponding periodograms in subject 2.
428
M. NAKAMURAET AL.
phase (downward deflection), higher harmonics of the rhythm can appear in the Fourier analysis. Consequently, the second harmonic of the dominant rhythm may occur in the beta band (over 13 Hz). Therefore, the amount of the periodogram in the beta band (S t) was obtained by summing up the periodogram within that band while excluding the amount of + 1 Hz around twice the peak frequency of the dominant rhythm (2 fd). The threshold value for the existence of the beta rhythm was given as 6S~-~ >_ 10 (/zV) (Table II). The features of the beta rhythm were evaluated using the EEG parameters, namely the amplitude in ~V and the asymmetry of the modified amounts in percentage. The threshold values for the existence of theta rhythm and delta rhythm were set as 6S~0 > 15 (~V) and 6 S ~ > 25 (izV), respectively. Alpha rhythm excluding the dominant rhythm (non-dominant alpha rhythm) was expressed by any rhythm in the alpha band (8-13 Hz) other than the dominant rhythm. The threshold value for the existence of the non-dominant alpha rhythm was given by 6~/~-~> 15 (/zV) (Table II). The features of these 3 rhythms were evaluated in
terms of their durations in percentage and electrodes active for respective rhythms. As the quantity 'duration' in the EEGer's interpretation was found to be the relative appearance rate of the respective wave components, it was expressed by {E(S/ST)10} × 100 (%), where S represents the amount of the periodogram of that particular wave component.
(5) Automatic EEG interpretation EEG parameters were computed for each segment data for each of the 14 subjects by using the periodogram parameters obtained in section 3 and the formulae as described above (Table II), and the final estimates of the EEG parameters were obtained by averaging the EEG parameters of all 10 segments. Then the scores for each item were calculated based on the criteria (threshold values) which were employed for the visual inspection of each item (Table I) and were compared with the results obtained by the EEGer's visual inspection for each subject. Finally, EEG data of 3 subjects, which were not included in the data used for development of the present system, were
Periodograms
EEG time series FPL-- A~,N, ~
Itern C P
3
-
3
A
-
O
,
A
,
~
~
-
A
,
L~..,,.a4._:
~
~ - ~ F , - A 2'
.
,
C, - A
P
2
,
O
~
,
,
'
2
v
.
~
~
2 A
' z
~
L,~/D~ L ~. . . . . . . L,.,.
~
A -
.
-
,~.~
. ~
~ ~
~
0 i
Asymmetry [%] 0.0 Frequency [Hz] 10.s i x0.s Asymmetry [Hz] 0.0
0 o 0
O
Amplitude[~V] 95.0 i 95.0 Asymetry [%] 0.0 Extension[~V] till C ABelitude[/JV] <35.0 Asymmetry [%] 0.0 Duration [%] 20. 0
0
O
0.0 10,3110.2 0.1 58.6;:57.6 -1.8 25.7!45.1 18.8
0
-3.0
2
14.8
O
O
t
Diffuse ,
s
-
,
-
A
A
~
2
~
~
T
,
-
A
~
~
T
,
-
A
2
~
xoo[.vl] 0
.....
2
[~ -.
~,,.-,- .
. ..
~
Electrodes
~
Duration [%]
'~'
more on R I
C
POT
40.0 I 2 !Fp F C P 0 T 0.0
0 0 O 0 1 O 0 2
Fp F
more on R
30. 5
g
T
F
Modeab.
Duration [%]
T3-At
Lti et
©:© 1.2 ! 1.3
Electrodes
O:O
Automtic Value Score
o 2
Existence Organization
~
EEGer Value Score Lt i Rt
I
2
m
Fp F
C
POT
O
Electrodes
L
.....
s i0 15 20 25
Time [sec] Frequency [8z] Fig. 3. The time series of a 5 sec segmentof 16-channelEEG and correspondingperiodogramsin subject3.
23,4 Fp F POT
I
I c
AUTOMATIC EEG INTERPRETATION
subjected to the automatic interpretation, and the results were compared with those of the visual interpretation. The algorithm for the present automatic EEG interpretation was programmed in C language and implemented by using a personal computer NEC PC-9801 RA with a DSP board for the FFT computation. Actually, transformation of the image data of the EEG record into digital data required a processing time of about 1 h for 1 subject. If the digital data of the EEG record are directly taken into the computer, the transformation time can be saved, and then the total processing time for the automatic EEG interpretation is only about 5.5 min: the data input time 2.5 min, the main calculation time 1 min and the print-out time 2 min. Results
The grading scores obtained for each item by the present automatic interpretation system were similar to those judged by the EEGer in most of the 14 subjects. Conformity between the results of the automatic interpretation and those of the visual interpretation was summarized in Table III. Scores obtained by the two methods were exactly the same in 78.6% of all items, different by 1 point in 13.7%, by 2 points in 4.4% and by 3 points in 3.3%. Figs. 1-3 show results of interpretation of EEGs from the 3 additional subjects. The EEG parameters and grading scores obtained by the automatic EEG interpretation are listed in the right two columns of the table in each figure, and the corresponding results obtained by the EEGer's visual inspection are shown in the left two columns. A 5 sec long segment of the EEG record and the corresponding periodograms for each channel are shown in each figure. Subject 1 was a 30-year-old woman presenting with focal sensorimotor seizure starting in the left hand of recent onset. On visual inspection of the EEG (Fig. 1 and a further 45 sec length of the EEG), the frequency of the dominant rhythm was estimated to be 10.3 Hz on average and symmetrical, but its organization was judged to be mildly abnormal (score 1). There was no abnormality noted with regard to the beta rhythm. Occasional theta waves were recognized diffusely, the duration of which was estimated to be 5% of the total recording time (score 2). There were episodic large amplitude delta waves diffusely which were occasionally more predominant on the right hemisphere (not included in the segment shown in Fig. 1); their proportional duration was estimated to be 15% of the total recording time (score 2). There were occasional rhythmic alpha waves other than the dominant rhythm anteriorly and bilaterally, the duration of which was estimated to be 20% of the total recording time (score 1).
429
Automatic interpretation of the same record revealed the mean frequency of the dominant rhythm to be 9.5 Hz on the left and 9.7 Hz on the right. Its organization was found to be slightly more abnormal than estimated from the visual inspection (score 2), and slightly asymmetric, being more abnormal on the right. Otherwise the same findings as judged from the visual inspection were revealed for the dominant rhythm. The proportional duration of theta waves was found to be slightly longer (17.4%) as compared with the visual inspection, but it did not make any difference to the score. The automatic interpretation detected the same amount of delta waves, although it failed to detect an occasional asymmetry which was observed on visual inspection. Activities of alpha frequency other than the dominant rhythm were found in the same amount diffusely whereas they were found only anteriorly on visual inspection. Subject 2 was a 56-year-old woman with a glioblastoma in the right parietal lobe. On visual inspection of the EEG (Fig. 2), the dominant rhythm was judged to be present on the left, but absent on the right (score 3 for asymmetry of the organization, frequency and amplitude). The dominant rhythm on the left hemisphere was judged to be normal in frequency (10.7 Hz), amplitude (40 /xV) and organization. Beta activities were judged to be less in quantity on the right hemisphere
T A B L E III Difference between the scores obtained by automatic interpretation and visual inspection in n u m b e r of records (total n u m b e r of subjects 14). Item
Difference of scores None
1
2
3
14 10 11 13 12 14 11 12
4 2 1 1 0 1 1
0 0 0 0 0 0 1
0 0 1 0 1 0 2 0
14 9
0 1
0 2
0 2
7
6
1
0
10
0
4
0
6
8
0
0
Dominant rhythm Existence Organization Asymmetry (%) Frequency Hz) Asymmetry (Hz) Amplitude (/~V) Asymmetry (%) Extension (tzV)
Beta rhythm Amplitude (~V) Asymmetry (%)
Them rhythm Duration (%) Electrodes
Delta rhythm Duration (%) Electrodes
Non-dominant alpha rhythm Duration (%) Electrodes
430
than on the left by 60% (score 2). Theta waves were occasionally (15% of the total recording time) seen diffusely on the right hemisphere (score 2). Continuous irregular delta waves were seen localized in the right posterior quadrant of the head (P4, 02 and T6). The automatic interpretation of the same record revealed essentially the same findings as the visual inspection with regard to the dominant rhythm, except that the dominant rhythm on the left hemisphere was also found to be slightly abnormal in organization (score 1). As to the beta waves, however, the automatic interpretation did not find any significant asymmetry. The same amount of theta waves was found by automatic interpretation, but diffuse instead of with rightsided predominance which was observed on visual inspection. Slow waves of delta frequency were found with a proportional duration of about half of what was estimated by visual inspection (44.7% vs. 90%), and more diffusely compared with visual inspection. Subject 3 was a 48-year-old woman in the post-operative state for a ruptured aneurysm arising from the right anterior cerebral artery. On visual inspection of the E E G (Fig. 3), the dominant rhythm was judged to be present bilaterally and symmetrical in terms of frequency and amplitude. However, its organization was judged to be moderately abnormal bilaterally (score 2). Beta rhythm was judged to be present equally on both hemispheres. Slow waves of theta frequency were judged to be present diffusely with a proportional duration of 20% (score 2). D~ Ita waves were judged to be diffuse, more on the right hemisphere, at 40% of the total recording time, including occasional paroxysmal bursts of large amplitude rhythmic delta waves which were seen bilaterally and more anteriorly. Prominent slow waves seen at Fp~ and Fp 2 were judged to be blink artifacts, based on their wave forms and restricted location. The automatic interpretation revealed essentially the same findings as regards the dominant rhythm, except that the organization was found to be mildly abnormal (score 1). Theta and delta waves were found diffusely in similar amounts to those estimated by visual inspection, but the slight asymmetry of slow waves observed on visual inspection was not found. Alpha waves other than the dominant rhythm were detected diffusely but were not observed on visual inspection. The slow waves limited to the fronto-polar electrodes were not taken as abnormal E E G activities by the automatic interpretation.
Discussion In the present novel automatic E E G interpretation system, first the items for the E E G interpretation were
M. N A K A M U R A ET AL.
determined so that they could conform to the procedure that a qualified E E G e r usually follows for visual inspection of the actual record, and all items were defined quantitatively. For automatic interpretation, the E E G parameters were chosen so that they could represent the basis of the EEGer's visual interpretation as closely as possible. The interpretation thus obtained was demonstrated to give similar results to those derived from visual interpretation. The usefulness of this system was confirmed by applying it to new E E G records which were not taken into account for developing the system. Since the results of automatic E E G interpretation were compared with those obtained through visual inspection by the same E E G e r upon whose own procedure of visual inspection the E E G parameters were based, it might be argued that the conformity between the two methods was predictable. However, the purpose of the present study was not to make a standard criterion for E E G interpretation, but to make an automatic analysis system which resembles the EEGer's visual inspection as closely as possible. Therefore, the present system is judged to be satisfactory for that purpose. In other words, another system can be made if another E E G e r wishes to make a different system which conforms to his own procedure of visual inspection. To use different threshold values given by another EEGer, this automatic system could be adaptable to the new system merely by changing the threshold values. Furthermore, by taking an appropriate weighted sum of the obtained scores, it would be possible to provide an automatic integrative interpretation of an awake E E G record just like the final summary derived by visual inspection. Upon detailed observation of the results, however, we found slight discrepancies in some items between the automatic interpretation and the EEGer's visual inspection. The causes of the differences were found to be two-fold: an artifact of eye blinks and activation of the reference electrodes. Such disagreements will be reduced by an appropriate preprocessing of E E G data for blinks, and by using a non-cephalic reference, for example, by employing an ECG elimination method (Nakamura and Shibasaki 1987) to skirt the problem of the earlobe activation. In contrast with the previous automatic analyses of E E G (Kowada et al. 1971; Yamamoto et al. 1975; Nakamura et al. 1985, 1989b), which were focussed on certain aspects such as the dominant rhythm or slow waves, the present system is unique in its capability to make an integrative interpretation of the spontaneous awake E E G by taking into account all its features except for paroxysmal abnormalities. Therefore, this system can be used clinically as an assistant tool for neurologists and EEGers and may be the first step toward its clinical application to hospital use.
AUTOMATIC EEG INTERPRETATION The authors are grateful to Mr. I. Mitsuyasu (postgraduate student) of Saga University for his help in computer computation.
References Chatrian, G.E., Bergamini, L., Dondey, M., Klass, D.W., LennoxBuchthal, M. and Peters~n, I. A glossary of terms most commonly used by clinical electroencephalographers. Electroenceph. clin. Neurophysiol., 1974, 37: 538-548. Cooley, J.W. and Tukey, J.W. An algorithm for the machine calculation of complex Fourier series. Math. Comput., 1965, 19: 297-301. Frost, Jr., .I.D. Automatic recognition and characterization of epileptiform discharges in the human EEG. J. Clin. Neurophysiol., 1985, 2: 231-249. Gotman, J. Practical use of computer-assisted EEG interpretation in epilepsy. J. Clin. Neurophysiol., 1985, 2: 251-265. 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. Inoue, K., Kumamaru, K., Sagara, S. and Matsuoka, S. Pattern recognition approach to human sleep EEG analysis and detection of sleep stages. Mem. Fac. Eng. Kyushu Univ., 1982, 42: 177-195. Kowada, M., Sato, S., Hiraga, H., Tomono, T. and Yahata, K. Automatic pattern recognition of clinical EEG. I. A method for
431 wave form recognition. Clin. Electroenceph., 1971, 13: 647-656. (In Japanese.) Nakamura, M. and Shibasaki, H. Elimination of EKG artifacts from EEG records: a new method of non-cephalic referential EEG record. Electroenceph. clin. Neurophysiol., 1987, 66: 89-92. Nakamura, M., Nishida, S., Neshige, R. and Shibasaki, H. Quantitative analysis of 'organization' by feature extraction of the EEG power spectrum. Electroenceph. clin. Neurophysiol., 1985, 60: 84-89. Nakamura, M., Nishida, S. and Imajoh, K. Deterioration of power spectrum for quantized time series and its recovering method. Preprints of the 8th SICE Conf. Kyushu Branch, 1989a, 8: 155158. (In Japanese.) Nakamura, M., Shibasaki, H. and Nishida, S. Automatic scoring system of EEG and quantitative evaluation of its visual interpretation. Proc. '89 KACC, 1989b: 967-971. Nishida, S., Nakamura, M. and Shibasaki, H. An EEG model expressed by the sinusoidal waves with the Malkov process amplitude. Jpn ,i, Med. Electron. Biol. Eng., 1986, 24: 8-14. (In Japanese.) Yamamoto, K., Nakamura, N., Shimazono, Y., Miyasaka, M. and Fukuzawa, H. The system construction on a newly developed automatic EEG diagnosing system - with special regard to the wave-form recognition method. Psychiat. Neurol. Jap., 1975, 77: 127-159. (Abstract in English.)