- Email: [email protected]
Elimination of EKG artifacts from EEG records: a new method of non-cephalic referential EEG recording
89
Electroencephalograplzv and clinical Neurophysiologv, 1987, 6 6 : 8 9 - 9 2 Elsevier Scientific Publishers Ireland, Ltd.
EE(I 03130
Shorl communication
Elimination of E K G artifacts from E E G records: a new method of non-cephalic referential E E G recording Masatoshi Nakamura and Hiroshi Shibasaki * Department of Electrical Engineering, Saga Uni~ersttv, ffo*!]omachi, Saga 840, and * D&ision of Neuroh)gv, Department of Internal Medicine, Saga Medical School, Nabe~hima, Saga ~4t) Ol (Japan) (Accepted for publication: 14 April, 1986)
Summa~'
A new method for eliminating EKG artifacts from EEGs was reported. Based on the simultaneously recorded EEGs and EKG, the procedure consisted of 4 steps: synchronized partition of the raw EEG record into segments with respect to the QRS complex of the EKG, averaging of the segments time-locked to EKG, repetition of the average artifact synchronized to EKG and subtraction of the estimate of the artifacts from the raw data. This method enabled the elimination of EKG artifacts from EEGs recorded by using the chin or the hand electrode as a reference. Features of the proposed method include: ~1) all signals are processed in digital instead of analogue form, (2) the average E K G wave form is computed throughout the subtraction part of the recording instead of being computed only before the subtraction starts, (3) in case of an on-line algorithm, the average EKG wave form is always updated to the point immediately preceding the real time of EKG subtraction, and (4) by using the recursive form of exponential weighted averaging equation, the average EKG wave form can adapt appropriately to a change of the artifact wave form. By the high speed implementation of the procedure, the processed data of multi-channel EEGs can be obtained on-line with a delay of only 200 msec.
Key words: EKG artifact elimination: non-cephalic referential EEG recording
Non-cephalic referential recording of the electroencephalogram lEE(I) has become increasingly important, as the technique of mapping EEG activities over the scalp began to be widely used. Conventional use of the ear reference, however, is frequently' misleading because of activation of the reference electrode by the background alpha rhythm, and above all by a focal slow wave or spike localized to the temporal region close to the earlobe (Kiloh et al. 1972). The non-cephalic reference derivation is ideal in this respect, but its practical use has been extremely limited due to its inevitable contamination by a large amplitude electrocardiogram (EKG). To eliminate the EKG artifacts from EEG records, there have been a few attempts such as the use of the balanced non-cephalic reference electrode (Stephenson and Gibbs 1951) and automatic elimination of EKG artifacts from the EE(i recorded by the balanced non-cephalic method (Bickford et al. 1971: Barlow and Dubinsky 19801 Ishiyama et ah 1980, 1982a, b). Bickford et al. (1971), by employing one EEG channel recorded in reference to the balanced non-cephalic electrode and a separate EKG channel in the computer evaluation of integrated brain activity in cases of suspected electro-cerebral inactivity, obtained an average E K G wave form for the EKGcontaminated EEG, the R wave in the E K G monitor channel serving as the trigger for the computer, and repeatedly sub-
tracted the predetermined averaged EK(I wave form from the subsequent incoming EKG-contaminated EEG channel. Barlow and Dubinsky (1980) implemented this basic technique in such a way as to permit EKG subtraction of as many as 16 channels of a clinical EEG in real time. Ishiyama et al. (1980, 1982a) used a hybrid computer to make the EKG elimination device more practical. All of these previous methods, however, were based on the EEGs recorded by using the balanced non-cephalic reference because of insufficient EKG elimination. In this communication, we report a new method for suc cessfully eliminating EKG artifacts from EEGs referentially recorded without a balanced non-cephalic reference.
Methods (1) Data acquisition The subjects of this study were 3 healthy young adults. Recording of EEG took place in a quiet, dimly lit room with the subject placed in a sitting position and with the eyes closed. An exploring electrode was fixed to the scalp at point O1 of the international 10-20 system, and 3 reference electrodes were placed at the left earlobe (A1), the chin and the right hand, respectively. The EEG was recorded using a time constant of
0013-4649/87/$03.50 ~:' 1987 Elsevier Scientific Publishers Ireland, Ltd.
90
M. NAKAMURA, H. SHIBASAKI
0.1 sec and a high frequency cut-off of 30 Hz. The EKG was simultaneously recorded from a pair of electrodes placed about 10 cm apart on the left chest, using a time constant of 0.1 sec and a high frequency cut-off of 30 Hz. The EEG and EKG records were digitized on-line at a high sampling rate of 1 kHz by the 12-bit analog-to-digital converters, and divided into 10 blocks of 5 sec each due to restrictcd memo D' size of a SORD M343 computer. The high sampling rate in the digitization was necessa W to make the synchronization mentioned below as accurately as possible.
to be the sum of the real EEG x(t) and the EKG artifacts z(t) as
y(t) = x(t) + z(t).
(1)
The peaks of EKG R waves p(t) were detected from the EKG record and used as the trigger signals. The starting point t i for the synchronized partition of the raw EEG was determined as the point 200 msec before the trigger pulse, so that the point preceded the E K G P wave. The raw EEG was then partitioned into segments, synchronized with the trigger pulse of the EKG as
KII) Artifact elimination procedure
y,*(r + t j ) = y ( r + t i ) { u ( r )
The artifact elimination method, implemented by the SORD computer, involved 4 separate steps: synchronized partition, averaging, synchronized repetition and subtraction (Fig. 1). Synchronized partition. The raw EEG data were assumed
u ( r 4 t,
lj , ) }
(2)
where u(t) denoted the step function. The term {u(r) u(r t t i tj + t)} then represented the boxar type window. Accuracy of the synchronization was 1 msec (sampling rate of 1 kHz). The total number of segments ",,','as L (l:ig. 1A/. Aueraging. Those segments obtained by the synchronized partition were then averaged, time-locked to each starting point as follows: L
~*(~) : ±
tl (A) ~y~(:+tj
t2
lr~3 tL ~,~ Synchronized Partition )=y(z+tj)(u(T)-u(~+tj-tj+l)} i
(B)
~v~'Averaging 2 . ( z ) = - 1-
L:
L: j=l
y~(:+tj )
1
t
-T v
-'T--
I
I
[
'
tI t2 t3 tL D) @Subtraction Processed EEG Data: ~(t)=y(t)-~(t)
Fig. 1. Procedure of the EKG artifact elimination: synchronized partition, averaging, synchronized repetition and subtraction. The symbol x(t) denotes the real EEG, z(t) the artifacts, and u(t) the step function.
(B)
As the length of segment differed from segment to segment, the number of average L~ depended on the elapsed tinle r from each starting point. The synchronized components z(t) of the raw EE(L namely the EKG artifacts, became evident in proportion to the number of segments averaged, whereas the asynchronized components x(t) decreased gradually (Fig. 1B). ~vnchroni;ed repetition. The wave form of the average artifact thus obtained was repeated, synchronized with the trigger pulse of the EKG. This operation was expressed by the following equation: L
z-(t)- ~ 2 * ( t - t i ) { u ( t - t i )
C) U'Synchronized Repetition l Estimate of Artifact: I L Z(t)= .~ Z*(t-tj){u(t-tj)-u(t-tj+l)} I
Y'. v,*(~ +t,),
u(t t,~,)}.
(4)
1
This iterative wave form served as an estimate of the EKG artifacts actually contaminated in the raw EEG (Fig. 1C). Subtraction. By subtracting the estimate of the artifacts z(t) from the raw EEG y(t), the processed EEG data was obtained as follows:
~,(t) = y ( t )
z(t).
(5)
These data were the final objective obtained bv the present artifact elimination method (Fig. 1D),
Results Figs. 2, 3 and 4 show the raw EEG data and the EKG recorded for obtaining triggers (A), the average artifact (B) and the processed EEG data (C) for O1-A1, Ol-chin and Ol-hand, respectively. The wave form in A and C each is a block of 5 sec data out of an overall 50 sec record, and the average artifact in B is a grand average of 50 sec data.
Ol ,41 record In the raw EEG data (Fig. 2A), no evidence of the artifact
EK(.; ARTIFACT E L I M I N A T I O N Pulse (EKG)
(A) Raw EEG Data and Trigger 2
91 (A) Raw EEG Data and Trigger Pulse (EKG)
2
~e e
t~g :> o.
(g) Average Artifact
(B) Average Artifact
(c) Processed EEG Data
(C) Processed EEG Data
e ,×
-°'I"
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
.
.
.
.
.
.
.
.
.
.
.
2
.
.
3
.
.
.
.
.
.
.
.
.
4
.
.
.
.
.
.
]
.
5
0
TIME [secl
[
2
3
a
5
TIME lsec)
Fig. 2. Wave forms derived from O I - A I
for a healthy young
adult: (A) raw EEG and EKG data, (B) average artifact, (C) processed EEG data. The average artifact (B) has an upward deflection at the end of the record due to a small number of segments with sufficient duration to be entered into the average at that point in time.
was recognized by visual inspection, but the average artifact record demonstrated a small amplitude EKG artifact (Fig. 2B). O1 chin record
EKG artifacts were prominent in the raw EEG data (Fig. 3A). Low amplitude slow waves were occasionally intermixed corresponding to the T wave of EKG (arrows). In the processed EEG data (Fig. 3C), the EKG artifacts were not recognized any more. Those low amplitude slow waves seen in the raw data (arrows in Fig. 3A) were also eliminated.
Fig. 4. Wave forms derived from e l - h a n d for the same subject as in Figs. 2 and 3: (A) raw' EEG and E K G data, (B) average artifact. (C) processed EEG data. The processed EEG wave form (CI appears a little noisy, probably because the low gain at which the e l - h a n d derivation was recorded resulted in a loss of resolution through the computer.
O l - h a n d record
In the raw data (Fig. 4A), the EEG activity was almost totally masked by extremely large EKGs. Subtracting the average artifact (Fig. 4B) from the raw data and magnifying the processed data by 20 times in scale, the EEG activit,~ became recognizable (Fig. 4C). The wave form in Fig. 4C was fairly similar to those of e l - A 1 and e l - c h i n derivations, although low amplitude EKG artifacts, mainly R waves, still remained from time to time.
Discussion
(A) Raw EEG Data and Trlgger Pulse (EKG)
L~e
(B) Average Artifact
(C) Processed EEG Data e ,x
0
1
3
e
4
5
TIME [~eCl
Fig. 3. Wave forms derived from e l - c h i n for same subject as in Fig. 2: (A) raw EEG and EKG data, (B) average artifact, (C) processed EEG data. The average artifact (B) has an upward deflection at the end of the record due to a small number of segments with sufficient duration to be entered into the average at that point in time.
The newly proposed method in this paper successfull,, eliminated E K G artifacts from the EEGs recorded by using the chin electrode as a reference (Fig. 3). The artifact elimination from the EEGs recorded against the hand reference seems to be also successful, when the incomparably large size of EK(i artifacts in this tracing is taken into account (Fig. 4). The basic principle of this method is similar to that by Biekford et al. (1971), Barlow and Dubinsky (1980) and Ishiyama et al. (1980, 1982a). In contrast with their methods, however, the present method does not require a balanced non-cephalic electrode as a reference due to sufficient EKG elimination. In the present method, the average EKG wave form is computed throughout the subtraction part of the recording, instead of being cam puted only before the subtraction part. Also in case of an on-line algorithm, which will be discussed later, the average EK(I wave form is always updated to the point immediateh preceding the real time of EKG subtraction. In the present method, moreover, all signals are processed in digital instead of analogue form, which will enable a more efficient data processing. The usefulness of this technique will lead to an accurate evaluation of the scalp topography of both physiological and
92
M. NAKAMURA, H. SHIBASAKI
abnormal EEG activities, and especially to a significant contribution to various mapping techniques. False localization of focal EEG abnormalities originating from the temporal region as a result of the car reference activation will be avoided by applying the present EKG elimination technique to the mapping techniques currently available. It is noteworthy that, even in the car reference derivation (Fig. 2). the background EEG activity is slightly improved by the EKG artifact elimination procedure. This finding is more pronounced in thc chin reference derivation (Fig. 3). This improvcmcnt in the quality of EEG records appears to be duc to the elimination not only of the QRS complex but also of slower components, cspccially the T wave, from the raw EEG data. This fact also suggests the possibility that EEG records without EKG elimination, even though the R waves of thc E K G arc not visible in the record, might result in an erroneous mtcrprctation of the normal record, for example, as a poorly, organized background activity or an intermixture of slow waves. This discussion obviously holds true for the pulse artifacts as wclh which are not infrequently encountered even in the bipolarly recorded EEGs. The basic assumption for thc EKG elimination methods described above is that the artifact wave form is constant from beat to beat during a whole recording period. Obviously, however, this assumption does no.t hold true for real recording. To improve the present method in this regard, the recursive form of exponential weighted least squares equation (Nakamura 1982. 1983) can be adopted instead of the equation (3l as followes: 2*,(r)=X~*,
~(r)~(1
X)y,*('r+t,).
(6)
The first term in the right hand side of the equation (6) implies thc a priori average artifact determined in the previous stage m 1. and the second term implies the innovation derived from the current measurement. The index X is appropriately determined in the range 0 < X ~<1. By using the recursive equation, the current average artifact f_*,(r) can adapt appropriately to a gradual change of the artifact wave form and can also adapt in several seconds after its sudden change. In the present analysis, the data were processed off-line for about 10 min for one channel. If the artifact elimination procedure with the recursive equation is programmed in mac+ hine language, the processed EEG can now be obtained on-line with a delay of only 200 msec. Multi-channel on-line processing of EEG can also be implemented by logic circuits, in which the artifact elimination procedure is designed for each channel. In addition, and probably more importantly, this technique could be applied to the recording of short latency evoked potentials that requires a non-cephalic reference derivation.
R6sume l'limination d'art~jacts E K G sur un enregistrement E E G : une nou+,e]lc re&bode d'enregixtrement avec r~f~rence non c@halique
Unc nouvelle methode pour eliminer les artefacts EKG est decrite. Basec sur l'enregistrement simultane de I'EEG ct de
I'EKG, cette methode possede 4 dtapes: segmentation synchronisdc du trace EEG par rapport au complexe QRS de I'EKG, moyennage des segments synchronises sur I'EKG, repetition de l'artefact moyenne synchronis,5 sur FEKG et soustraction dc cet artefact estim6 de la sdrie d'enregistrements. Cctte mdthode permet l'elimination des artefacts EK(i dans lcs enregistremcnts EEG, en utilisant commc reference lc mcnton ou la main. Entre autres caracteristiques de cette methode on pout citer: (1) le traitement des signaux qui doivent Cztre sous formc digitale et non analogique, (2) la forint de l'oude EKG qui cst calculde pendant toute la periodc de sa soustraction de l'enregistrement au lieu de n'~tre calculee qu'avant. (3) dans le cas d'un algorithme en temps reel, la forme movcnnec de l'ondc EKG est toujours mise '~ jour immediatcmcnt avant le mmnent reel de sa soustraction, et (4) la possibilit6 d'adaptation au changement de forme de l'artefact, en utilisant la forme recursire de l'equation exponentielle de moyennage pon&Sre. Par la rapidite d'action de cette mdthode, des donndes EEG sur canaux multiples peuvent etre traitees, en temps rdel. avcc un delai de 200 msec seulement.
References
Barlow, J.S. and Dubinsky, J. EKG-artifact minimization in referential EEG recordings by computer subtraction. Elcctroenceph, olin. Neurophysiol., 1980+ 48:470 472. Bickford, R.G., Sims, J.K., Bilinger, T.W. and Aung, M.H. Problems in EEG estimation of brain death and uses of computer techniques for their solution. Trauma, 1971, 12: 61-95. Ishiyama, Y., Ebe, M., Homma, I. and Abe, Z. Non-cephalic reference electrode method in EEG recording. Elimination of ECG artifacts mixed in EEG. ME & BE, 1'~80, 18: 334 340. Ishiyama, Y., Ebe, M., Homma, 1. and Abe, Z. Elimination of EKG artifacts from EEGs recorded with balanced noncephalic reference electrode method Electroenceph. ,,din. Neurophysiol., 1982a, 53:662 665. Ishiyama, Y.. Ebe, M., Homma, I., Abe, Z. and Hino, M. Automatic elimination of ECG artifacts mixing in EE(; recorded by BNE method. ME & BE, 1982b, 20: $438. Kiloh, L.G., McComas, A.J. and Osselton. J.W. Clinical Electrocncephalography+ 3rd edn. Butterworth, London, 1972: 42. Nakamura, M. Relationship between steady state Kahnan filter gain and noise variance. Int. J. Syst. Sci., 1982, Ill: 1153-1163. Nakarnura, M. Exponential weighted least squares method for nonlinear stochastic systems. J. Soc. Instrum. Control Engng. 1983. 19:607 613. Stephenson, W.A. and Gibbs, F.A. A balanced non-cephalic reference electrode. Electroenceph. olin. Neurophvsiol., 1951, 3: 237-240.
Electroencephalograplzv and clinical Neurophysiologv, 1987, 6 6 : 8 9 - 9 2 Elsevier Scientific Publishers Ireland, Ltd.
EE(I 03130
Shorl communication
Elimination of E K G artifacts from E E G records: a new method of non-cephalic referential E E G recording Masatoshi Nakamura and Hiroshi Shibasaki * Department of Electrical Engineering, Saga Uni~ersttv, ffo*!]omachi, Saga 840, and * D&ision of Neuroh)gv, Department of Internal Medicine, Saga Medical School, Nabe~hima, Saga ~4t) Ol (Japan) (Accepted for publication: 14 April, 1986)
Summa~'
A new method for eliminating EKG artifacts from EEGs was reported. Based on the simultaneously recorded EEGs and EKG, the procedure consisted of 4 steps: synchronized partition of the raw EEG record into segments with respect to the QRS complex of the EKG, averaging of the segments time-locked to EKG, repetition of the average artifact synchronized to EKG and subtraction of the estimate of the artifacts from the raw data. This method enabled the elimination of EKG artifacts from EEGs recorded by using the chin or the hand electrode as a reference. Features of the proposed method include: ~1) all signals are processed in digital instead of analogue form, (2) the average E K G wave form is computed throughout the subtraction part of the recording instead of being computed only before the subtraction starts, (3) in case of an on-line algorithm, the average EKG wave form is always updated to the point immediately preceding the real time of EKG subtraction, and (4) by using the recursive form of exponential weighted averaging equation, the average EKG wave form can adapt appropriately to a change of the artifact wave form. By the high speed implementation of the procedure, the processed data of multi-channel EEGs can be obtained on-line with a delay of only 200 msec.
Key words: EKG artifact elimination: non-cephalic referential EEG recording
Non-cephalic referential recording of the electroencephalogram lEE(I) has become increasingly important, as the technique of mapping EEG activities over the scalp began to be widely used. Conventional use of the ear reference, however, is frequently' misleading because of activation of the reference electrode by the background alpha rhythm, and above all by a focal slow wave or spike localized to the temporal region close to the earlobe (Kiloh et al. 1972). The non-cephalic reference derivation is ideal in this respect, but its practical use has been extremely limited due to its inevitable contamination by a large amplitude electrocardiogram (EKG). To eliminate the EKG artifacts from EEG records, there have been a few attempts such as the use of the balanced non-cephalic reference electrode (Stephenson and Gibbs 1951) and automatic elimination of EKG artifacts from the EE(i recorded by the balanced non-cephalic method (Bickford et al. 1971: Barlow and Dubinsky 19801 Ishiyama et ah 1980, 1982a, b). Bickford et al. (1971), by employing one EEG channel recorded in reference to the balanced non-cephalic electrode and a separate EKG channel in the computer evaluation of integrated brain activity in cases of suspected electro-cerebral inactivity, obtained an average E K G wave form for the EKGcontaminated EEG, the R wave in the E K G monitor channel serving as the trigger for the computer, and repeatedly sub-
tracted the predetermined averaged EK(I wave form from the subsequent incoming EKG-contaminated EEG channel. Barlow and Dubinsky (1980) implemented this basic technique in such a way as to permit EKG subtraction of as many as 16 channels of a clinical EEG in real time. Ishiyama et al. (1980, 1982a) used a hybrid computer to make the EKG elimination device more practical. All of these previous methods, however, were based on the EEGs recorded by using the balanced non-cephalic reference because of insufficient EKG elimination. In this communication, we report a new method for suc cessfully eliminating EKG artifacts from EEGs referentially recorded without a balanced non-cephalic reference.
Methods (1) Data acquisition The subjects of this study were 3 healthy young adults. Recording of EEG took place in a quiet, dimly lit room with the subject placed in a sitting position and with the eyes closed. An exploring electrode was fixed to the scalp at point O1 of the international 10-20 system, and 3 reference electrodes were placed at the left earlobe (A1), the chin and the right hand, respectively. The EEG was recorded using a time constant of
0013-4649/87/$03.50 ~:' 1987 Elsevier Scientific Publishers Ireland, Ltd.
90
M. NAKAMURA, H. SHIBASAKI
0.1 sec and a high frequency cut-off of 30 Hz. The EKG was simultaneously recorded from a pair of electrodes placed about 10 cm apart on the left chest, using a time constant of 0.1 sec and a high frequency cut-off of 30 Hz. The EEG and EKG records were digitized on-line at a high sampling rate of 1 kHz by the 12-bit analog-to-digital converters, and divided into 10 blocks of 5 sec each due to restrictcd memo D' size of a SORD M343 computer. The high sampling rate in the digitization was necessa W to make the synchronization mentioned below as accurately as possible.
to be the sum of the real EEG x(t) and the EKG artifacts z(t) as
y(t) = x(t) + z(t).
(1)
The peaks of EKG R waves p(t) were detected from the EKG record and used as the trigger signals. The starting point t i for the synchronized partition of the raw EEG was determined as the point 200 msec before the trigger pulse, so that the point preceded the E K G P wave. The raw EEG was then partitioned into segments, synchronized with the trigger pulse of the EKG as
KII) Artifact elimination procedure
y,*(r + t j ) = y ( r + t i ) { u ( r )
The artifact elimination method, implemented by the SORD computer, involved 4 separate steps: synchronized partition, averaging, synchronized repetition and subtraction (Fig. 1). Synchronized partition. The raw EEG data were assumed
u ( r 4 t,
lj , ) }
(2)
where u(t) denoted the step function. The term {u(r) u(r t t i tj + t)} then represented the boxar type window. Accuracy of the synchronization was 1 msec (sampling rate of 1 kHz). The total number of segments ",,','as L (l:ig. 1A/. Aueraging. Those segments obtained by the synchronized partition were then averaged, time-locked to each starting point as follows: L
~*(~) : ±
tl (A) ~y~(:+tj
t2
lr~3 tL ~,~ Synchronized Partition )=y(z+tj)(u(T)-u(~+tj-tj+l)} i
(B)
~v~'Averaging 2 . ( z ) = - 1-
L:
L: j=l
y~(:+tj )
1
t
-T v
-'T--
I
I
[
'
tI t2 t3 tL D) @Subtraction Processed EEG Data: ~(t)=y(t)-~(t)
Fig. 1. Procedure of the EKG artifact elimination: synchronized partition, averaging, synchronized repetition and subtraction. The symbol x(t) denotes the real EEG, z(t) the artifacts, and u(t) the step function.
(B)
As the length of segment differed from segment to segment, the number of average L~ depended on the elapsed tinle r from each starting point. The synchronized components z(t) of the raw EE(L namely the EKG artifacts, became evident in proportion to the number of segments averaged, whereas the asynchronized components x(t) decreased gradually (Fig. 1B). ~vnchroni;ed repetition. The wave form of the average artifact thus obtained was repeated, synchronized with the trigger pulse of the EKG. This operation was expressed by the following equation: L
z-(t)- ~ 2 * ( t - t i ) { u ( t - t i )
C) U'Synchronized Repetition l Estimate of Artifact: I L Z(t)= .~ Z*(t-tj){u(t-tj)-u(t-tj+l)} I
Y'. v,*(~ +t,),
u(t t,~,)}.
(4)
1
This iterative wave form served as an estimate of the EKG artifacts actually contaminated in the raw EEG (Fig. 1C). Subtraction. By subtracting the estimate of the artifacts z(t) from the raw EEG y(t), the processed EEG data was obtained as follows:
~,(t) = y ( t )
z(t).
(5)
These data were the final objective obtained bv the present artifact elimination method (Fig. 1D),
Results Figs. 2, 3 and 4 show the raw EEG data and the EKG recorded for obtaining triggers (A), the average artifact (B) and the processed EEG data (C) for O1-A1, Ol-chin and Ol-hand, respectively. The wave form in A and C each is a block of 5 sec data out of an overall 50 sec record, and the average artifact in B is a grand average of 50 sec data.
Ol ,41 record In the raw EEG data (Fig. 2A), no evidence of the artifact
EK(.; ARTIFACT E L I M I N A T I O N Pulse (EKG)
(A) Raw EEG Data and Trigger 2
91 (A) Raw EEG Data and Trigger Pulse (EKG)
2
~e e
t~g :> o.
(g) Average Artifact
(B) Average Artifact
(c) Processed EEG Data
(C) Processed EEG Data
e ,×
-°'I"
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
.
.
.
.
.
.
.
.
.
.
.
2
.
.
3
.
.
.
.
.
.
.
.
.
4
.
.
.
.
.
.
]
.
5
0
TIME [secl
[
2
3
a
5
TIME lsec)
Fig. 2. Wave forms derived from O I - A I
for a healthy young
adult: (A) raw EEG and EKG data, (B) average artifact, (C) processed EEG data. The average artifact (B) has an upward deflection at the end of the record due to a small number of segments with sufficient duration to be entered into the average at that point in time.
was recognized by visual inspection, but the average artifact record demonstrated a small amplitude EKG artifact (Fig. 2B). O1 chin record
EKG artifacts were prominent in the raw EEG data (Fig. 3A). Low amplitude slow waves were occasionally intermixed corresponding to the T wave of EKG (arrows). In the processed EEG data (Fig. 3C), the EKG artifacts were not recognized any more. Those low amplitude slow waves seen in the raw data (arrows in Fig. 3A) were also eliminated.
Fig. 4. Wave forms derived from e l - h a n d for the same subject as in Figs. 2 and 3: (A) raw' EEG and E K G data, (B) average artifact. (C) processed EEG data. The processed EEG wave form (CI appears a little noisy, probably because the low gain at which the e l - h a n d derivation was recorded resulted in a loss of resolution through the computer.
O l - h a n d record
In the raw data (Fig. 4A), the EEG activity was almost totally masked by extremely large EKGs. Subtracting the average artifact (Fig. 4B) from the raw data and magnifying the processed data by 20 times in scale, the EEG activit,~ became recognizable (Fig. 4C). The wave form in Fig. 4C was fairly similar to those of e l - A 1 and e l - c h i n derivations, although low amplitude EKG artifacts, mainly R waves, still remained from time to time.
Discussion
(A) Raw EEG Data and Trlgger Pulse (EKG)
L~e
(B) Average Artifact
(C) Processed EEG Data e ,x
0
1
3
e
4
5
TIME [~eCl
Fig. 3. Wave forms derived from e l - c h i n for same subject as in Fig. 2: (A) raw EEG and EKG data, (B) average artifact, (C) processed EEG data. The average artifact (B) has an upward deflection at the end of the record due to a small number of segments with sufficient duration to be entered into the average at that point in time.
The newly proposed method in this paper successfull,, eliminated E K G artifacts from the EEGs recorded by using the chin electrode as a reference (Fig. 3). The artifact elimination from the EEGs recorded against the hand reference seems to be also successful, when the incomparably large size of EK(i artifacts in this tracing is taken into account (Fig. 4). The basic principle of this method is similar to that by Biekford et al. (1971), Barlow and Dubinsky (1980) and Ishiyama et al. (1980, 1982a). In contrast with their methods, however, the present method does not require a balanced non-cephalic electrode as a reference due to sufficient EKG elimination. In the present method, the average EKG wave form is computed throughout the subtraction part of the recording, instead of being cam puted only before the subtraction part. Also in case of an on-line algorithm, which will be discussed later, the average EK(I wave form is always updated to the point immediateh preceding the real time of EKG subtraction. In the present method, moreover, all signals are processed in digital instead of analogue form, which will enable a more efficient data processing. The usefulness of this technique will lead to an accurate evaluation of the scalp topography of both physiological and
92
M. NAKAMURA, H. SHIBASAKI
abnormal EEG activities, and especially to a significant contribution to various mapping techniques. False localization of focal EEG abnormalities originating from the temporal region as a result of the car reference activation will be avoided by applying the present EKG elimination technique to the mapping techniques currently available. It is noteworthy that, even in the car reference derivation (Fig. 2). the background EEG activity is slightly improved by the EKG artifact elimination procedure. This finding is more pronounced in thc chin reference derivation (Fig. 3). This improvcmcnt in the quality of EEG records appears to be duc to the elimination not only of the QRS complex but also of slower components, cspccially the T wave, from the raw EEG data. This fact also suggests the possibility that EEG records without EKG elimination, even though the R waves of thc E K G arc not visible in the record, might result in an erroneous mtcrprctation of the normal record, for example, as a poorly, organized background activity or an intermixture of slow waves. This discussion obviously holds true for the pulse artifacts as wclh which are not infrequently encountered even in the bipolarly recorded EEGs. The basic assumption for thc EKG elimination methods described above is that the artifact wave form is constant from beat to beat during a whole recording period. Obviously, however, this assumption does no.t hold true for real recording. To improve the present method in this regard, the recursive form of exponential weighted least squares equation (Nakamura 1982. 1983) can be adopted instead of the equation (3l as followes: 2*,(r)=X~*,
~(r)~(1
X)y,*('r+t,).
(6)
The first term in the right hand side of the equation (6) implies thc a priori average artifact determined in the previous stage m 1. and the second term implies the innovation derived from the current measurement. The index X is appropriately determined in the range 0 < X ~<1. By using the recursive equation, the current average artifact f_*,(r) can adapt appropriately to a gradual change of the artifact wave form and can also adapt in several seconds after its sudden change. In the present analysis, the data were processed off-line for about 10 min for one channel. If the artifact elimination procedure with the recursive equation is programmed in mac+ hine language, the processed EEG can now be obtained on-line with a delay of only 200 msec. Multi-channel on-line processing of EEG can also be implemented by logic circuits, in which the artifact elimination procedure is designed for each channel. In addition, and probably more importantly, this technique could be applied to the recording of short latency evoked potentials that requires a non-cephalic reference derivation.
R6sume l'limination d'art~jacts E K G sur un enregistrement E E G : une nou+,e]lc re&bode d'enregixtrement avec r~f~rence non c@halique
Unc nouvelle methode pour eliminer les artefacts EKG est decrite. Basec sur l'enregistrement simultane de I'EEG ct de
I'EKG, cette methode possede 4 dtapes: segmentation synchronisdc du trace EEG par rapport au complexe QRS de I'EKG, moyennage des segments synchronises sur I'EKG, repetition de l'artefact moyenne synchronis,5 sur FEKG et soustraction dc cet artefact estim6 de la sdrie d'enregistrements. Cctte mdthode permet l'elimination des artefacts EK(i dans lcs enregistremcnts EEG, en utilisant commc reference lc mcnton ou la main. Entre autres caracteristiques de cette methode on pout citer: (1) le traitement des signaux qui doivent Cztre sous formc digitale et non analogique, (2) la forint de l'oude EKG qui cst calculde pendant toute la periodc de sa soustraction de l'enregistrement au lieu de n'~tre calculee qu'avant. (3) dans le cas d'un algorithme en temps reel, la forme movcnnec de l'ondc EKG est toujours mise '~ jour immediatcmcnt avant le mmnent reel de sa soustraction, et (4) la possibilit6 d'adaptation au changement de forme de l'artefact, en utilisant la forme recursire de l'equation exponentielle de moyennage pon&Sre. Par la rapidite d'action de cette mdthode, des donndes EEG sur canaux multiples peuvent etre traitees, en temps rdel. avcc un delai de 200 msec seulement.
References
Barlow, J.S. and Dubinsky, J. EKG-artifact minimization in referential EEG recordings by computer subtraction. Elcctroenceph, olin. Neurophysiol., 1980+ 48:470 472. Bickford, R.G., Sims, J.K., Bilinger, T.W. and Aung, M.H. Problems in EEG estimation of brain death and uses of computer techniques for their solution. Trauma, 1971, 12: 61-95. Ishiyama, Y., Ebe, M., Homma, I. and Abe, Z. Non-cephalic reference electrode method in EEG recording. Elimination of ECG artifacts mixed in EEG. ME & BE, 1'~80, 18: 334 340. Ishiyama, Y., Ebe, M., Homma, 1. and Abe, Z. Elimination of EKG artifacts from EEGs recorded with balanced noncephalic reference electrode method Electroenceph. ,,din. Neurophysiol., 1982a, 53:662 665. Ishiyama, Y.. Ebe, M., Homma, I., Abe, Z. and Hino, M. Automatic elimination of ECG artifacts mixing in EE(; recorded by BNE method. ME & BE, 1982b, 20: $438. Kiloh, L.G., McComas, A.J. and Osselton. J.W. Clinical Electrocncephalography+ 3rd edn. Butterworth, London, 1972: 42. Nakamura, M. Relationship between steady state Kahnan filter gain and noise variance. Int. J. Syst. Sci., 1982, Ill: 1153-1163. Nakarnura, M. Exponential weighted least squares method for nonlinear stochastic systems. J. Soc. Instrum. Control Engng. 1983. 19:607 613. Stephenson, W.A. and Gibbs, F.A. A balanced non-cephalic reference electrode. Electroenceph. olin. Neurophvsiol., 1951, 3: 237-240.