Scandinavian Journal of Clinical and Laboratory Investigation
ISSN: 0036-5513 (Print) 1502-7686 (Online) Journal homepage: http://www.tandfonline.com/loi/iclb20
Noise Reduction of ECG by Averaging an Experimental Study of the Procedure and a Validated Method Lisbet Jansson, B. Jonson & O. Werner To cite this article: Lisbet Jansson, B. Jonson & O. Werner (1975) Noise Reduction of ECG by Averaging an Experimental Study of the Procedure and a Validated Method, Scandinavian Journal of Clinical and Laboratory Investigation, 35:8, 781-787, DOI: 10.3109/00365517509095810 To link to this article: http://dx.doi.org/10.3109/00365517509095810
Published online: 08 Jul 2009.
Submit your article to this journal
Article views: 2
View related articles
Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=iclb20 Download by: [McMaster University]
Date: 17 March 2016, At: 21:37
Noise Reduction of ECG by Averaging An Experimental Study of the Procedure and a Validated Method
Downloaded by [McMaster University] at 21:37 17 March 2016
LISBET JANSSON, B. JONSON & 0. WERNER Dept. of Clinical Physiology, University of Lund, Lund, Sweden
Jansson, Lisbet, Jonson, B. & Werner, 0. Noise Reduction of ECG by Averaging. An Experimental Study of the Procedure and a Validated Method, Scand. J. elin. Lab. Invest. 35, 781-787, 1975. Electrical noise often hampers the interpretation of ECG and other heartsynchronous signals. Time-coherent averaging is a commonly used method to reduce the noise. This method has been implemented on a computer with 12,000 12-bit words of core memory, suitable for on-line use in any clinical laboratory. An analysis was made of which demands apply to the exactness of synchronization of added heart cycles. The method was tested with respect to these demands and also, on a limited scale, for its efficiency in eliminating ectopic beats from the average. Key-words: Averaging; digital computers; ECG; exercise test; heart synchronous events 0. Werner, M.D., Kliniskt fysiologiska avdelningen, Lasarettet, S-221 85 Lund, Sweden
The interpretation of ECG signals is often hampered by noise, especially during exercise. The noise may partly be compensated for by visual methods. However, it is well known that the ST-T segment is difficult to judge even qualitatively in the presence of slow base-line shifts (9). Real difficulties appear when one attempts to make quantitative measurements. The most important method to reduce the noise is a correct choice of the skin electrodes and of the ECG amplifier. The noise may be further reduced by filtering, which, however, implies signal distorsion (12). To avoid this, methods that make use of the fact that the ECG is a repetitive signal must be used. One may, for example, let the median of several ECG complexes represent the signal (10). However, this does not reduce high-frequency noise of a low amplitude. A commonly used method for noise reduction is time-coherent averaging (1, 2, 4-7, 11, 12).
Averaging has the advantages of being a simple and fast procedure and of reducing all sorts of random noise. Under ideal conditions, i.e. an infinite sampling rate, a correct timing, and a uniformly repetitive pattern, it does not distort the signal. In practice several problems must be solved. Most important, a correct time reference point must be found in each ECG complex, so that the various heart cycles are correctly synchronized when added onto each other. Second, ectopic beats must be eliminated, .since otherwise the prerequisite that only uniform patterns may be added is violated. In previous studies these problems were well recognized. Surprisingly, no attempts seem to have been made to analyze how exact the synchronization must be and whether the methods used are adequate in this respect. The present paper describes an analysis of this type. A program for averaging the ECG and other heart-synchronous signals was de-
782
Lisbet Jansson, B. Jonson & 0 . Werner
signed with a special effort to arrive at an adequate synchronization of heart cycles, also under unfavorable conditions such as a high noise level. A method is described that actually measures the timing error. The program was designed so that ectopic beats were eliminated from the average without too much computing time being spent. METHODS
Downloaded by [McMaster University] at 21:37 17 March 2016
Equipment (Fig. 1 )
The ECG and other heart-synchronous signals were recorded with a direct-writing Mingograph 81 (Siemens-Elema AB) and simultaneously fed into the computer for averaging. Program
Details on the algorithms used for averaging are available from the authors. Four signals were fed into the computer. Three of these were ECG leads used for triggering. The fourth signal could be ECG or any other heart-synchronous signal. A reference complex was checked for mag-
nitude and signs of slopes. On this basis incidences preliminarily regarded as QRS complexes were recognized and the intervals between them memorized. I n a later step each such incident was further analyzed with regard to magnitudes and signs of slopes in the triggering leads and with regard to the preceding and subsequent R-R intervals. Only such incidences, which were found to belong to a ‘majority type’ with regard to these measures, were finally added to the average. Special precaution was taken to find the most distinctive point in the QRS complex to be used for synchronization between beats. To obtain realistic noise signals, an ECG was recorded from leads CH2, CH,, CH5, and CH, during strenuous exercise in a subject. The ECG was stored on digital magnetic tape with a sampling rate of 400/s. It was ascertained that the rhythm was regular. Next, the digitalized ECG signals were reexamined by a special program, which again located the time reference points belonging to each QRS. The average obtained in each of the four channels was then subtracted beat for beat from the corresponding original signal, leaving signals with only insignificant residuals of the QRS complexes (see noise signal in Fig. 4). RESULTS
Fig. 1. General set-up of the system. Signals from the patient are recorded on a Mingograph 81 (Siemens-Elema AB). Four signals are fed into the computer and are AD-converted. Two analogue output channels connect the computer to the oscilloscope and the mingograph. The latter thus serves two purposes - both as a direct analogue writing equipment and as an output organ for the averaged signals and text from the computer. When the analogue output channels are occupied with information to the oscilloscope, a special relay provides zero signals to the mingograph. The numerical keyboard gives information to the computer and also directs program flow.
No errors in the average could be visually observed when it was computed under optimal circumstances, i.e. a high sampling rate (400/s) and a low noise in the original signal. Fig. 2 shows to what extent the average would be affected by errors in the time reference point. The introduction of artificial errors in the time reference point even within 2.5 ms yielded an average with slightly blurred fine details and slightly reduced QRS amplitudes. A sampling frequency of 200/s will theoretically limit the exactness of the time reference point to about k 2 . 5 ms. An average obtained at this sampling rate (Fig. 3) is similar to that obtained with a n artificial error in the time reference point within 2.5 ms (Fig. 2B). An illustration of program performance is
Noise Reduction of ECG by Averaging
1 mV
783
I
CHS
(i A 'i
I
Downloaded by [McMaster University] at 21:37 17 March 2016
I
1
i;
1
h Fig.2. Effects of varying the position of the time reference points. Pictures A-D show the averages of the same 27 beats recorded on digital magnetic tape from a subject at rest. I n picture A no interference with the time reference point was made. An average was obtained that was visually identical to the original signal. In pictures B-D, welldefined but varying displacement of the time reference point for each complex is made, before adding the complex to the sum. For example, picture C shows the result when the time reference points in the individual ECG complexes included in the average were moved from their correct positions by -5, -2.5, 0, 4-2.5, or + 5 ms. Minus denotes a shift to the left and plus a shift to the right. Each of the time shifts mentioned was used in every fifth beat. This should simulate a situation in which the position of the time reference point has an estimated error within 5 ms. Similarly, pictures B and D show the result when the correct position of the time reference point has errors within 2.5 ms and 7.5 ms, respectively, as defined for picture C. shown in Fig. 4. The realistic noise was added to a noise-free ECG from another subject. Averaging was done on the combined signal. A striking degree of noise reduction was achieved. Even fine details of the QRS complexes are clearly demonstrated (compare the complexes in Fig. 4B). A certain base-line instability re-
Fig. 3. Effects of two different sampling rates. The average of exactly the same stretch of signal (26 ECG complexes) using sampling rates of 400/s and 2OO/s, respectively.
mained, which is not surprising in view of the very marked breath-synchronous base-line shifts in the noise signal. The noise still remaining after averaging was estimated by subtracting the average of the noise-free ECG from the average of the noisy ECG. The remnants of the QRS complexes were very small, which indicates that the distorsion of the ECG
~~erog ofa24cornplexes
-
A r c r ~ g ~ o l I I r o m p l e x ~ ~Avcra8col 27noi'o-frcc ----
-10ms -ZSms -5mr -25ms Oms shifts to the left
n
LI' smallshift tothelsft
2 -
a'-
-'-
'
Jw I
*2.5ms +5mr *ZSms *10ms shifts to the right
"residual signal "
Fig. 5. Principle for estimation of error in time reference point. A. (Schematic drawing.) Pointwise subtraction of one identical waveform from another. When the two are synchronous, the result is of course uniformly zero. In the two pictures to the right, small time shifts between the waveforms above and below are introduced. A given time shift will after subtraction result in a characteristic residual waveform. B. In this picture the average ECG has been subtracted from itself. Before subtraction artificial time shifts were introduced. A certain time shift will cause residuals, characteristic of the time shift in shape and amplitude. These residuals were used as a scale to estimate time shifts in picture C. C. Above, an original ECG is shown. Below, the average of this ECG has been subtracted from the original beat-for-beat, so that the trigger points coincide with the corresponding point in the average - i.e. no artificial time shift has been used. It can be seen that the residual 'QRS complexes' below each QRS complex are small. This in turn indicates that the trigger procedure cannot have been greatly in error in any of the ECG complexes, since otherwise large residual 'QRS complexes' would have remained. A comparison with figure B shows that the residual complexes are in fact smaller than the ones that were arrived at when timing differences of +2.5 ms and -2.5 ms were used intentionally.
due to inexact triggering caused by noise must have been insignificant. The same procedure applied to another noise-free ECG gave a similar result. The result sJlown in Fig. 4 made it probable that the exactness of the time reference point was adequate in the presence of heavy realistic noise. To prove this further, a similar subtraction technique was used. The details are explained in relation to Fig. 5. The virtually noise-free ECGs sampled at 400/s from two persons (channels CH2, CH,, CH5, and CH,) were analyzed. The individual QRS complexes were visually uniform. The estimated variation in the time reference point was found to be less than 2.5 ms in all the 50 heart cycles studied in each subject. Next the noise signal was added. This introduced an estimated error in the time reference point of 5-7.5 ms in one beat, 2.5-5 ms in nine, and less than 2.5 ms in 90 beats. The test just described was made on two members of the staff during the development of the program; the same technique was applied to ten patients under realistic clinical circumstances. Ten patients taking an exercise test for clinical reasons were randomly selected, and a total of 1,457 beats were recorded at the highest work load reached. In 93% of all beats the error was below 2.5 ms; in 6%) 2.5-5 ms; in 0.7%, 5-7.5 ms; and in 1 beat, 7.5-10 ms (Table I). The results thus agreed with those of the staff. Triggering on artifacts (which were not included in the average because of the trigger interval criterion) occurred in 10 instances. As mentioned above, a beat was included into the average if both the beat-to-beat intervals and the qualified trigger quantity were within certain automatically determined limits. Firm conclusions on the ability of the program to reject ectopic beats must await the results of clinical applications. A preliminary study of a few patients allows us to predict that most ectopic beats will be rejected by the program (Table 11). It is notable that several of the ectopic beats that were erroneously accepted into the average were in fact fusion beats from one patient (I.A.). Erroneous acceptance of ectopic
Noise Reduction of ECG by Averaging
785
Table I. Estimated timing error of the qualified trigger point during exercise (at highest work load reached by each patient)
Patient K.A. D.O. K.B. B.O.
W.P. E.N. K.H.
Downloaded by [McMaster University] at 21:37 17 March 2016
B.M. A.L. M.V.
Work load Number of (W) heart beats* 150 40 150 90 150 90
0-2.5 ms 134 197 163 147 148 156 153 83 117 64 1362
143 207 171 149 158 161 175 I09 120 64 1457
150 150 100
50
Total
* Ectopic beats excluded. ** Not included in average because of
Estimated errors 2.5-5 ms 5-7.5 ms
Table 11. Number of ventricular ectopic beats accepted into the average in nine patients ~
All ventricular Ventricular ectopic beats
Initials
ectopic beats
H.N.* J.S.* H.M.
3 3
-
1
-
87 5 173 5 1 46 324
-
L.L.
accepted into average 3
1 27** -
H.H. S.T. Total 31 * These patients also had one atrial ectopic beat each, which was eliminated because of the trigger interval criterion. ** Of these at least twelve were judged to be fusion beats. 6 - Scand. J. din. Lab. Invest.
1 2 4
7 1
1
1
10 4 22 25 3 84
Triggering on artifacts**
2 1
10
1
10
trigger interval criterion.
beats occurred only at heart rates exceeding 120/min. It was noted that both the selection criterion based on the magnitudes and signs of slopes in the QRS and the criterion based on beat-tobeat intervals were necessary to achieve these results. Thus 59 of the ventricular ectopic beats were rejected only because the first-mentioned criterion was not satisfied, whereas 27 beats were rejected only because of the beat-to-beat intervals. The latter mechanism failed especially often at a high heart rate in spite of the narrowing trigger interval limits that were used.
U.Q. K.F. I.A.
9 3 7
7.5-10 ms
DISCUSSION The averaging technique has been used during the last ten years to reduce the noise in the ECG. Various theoretical limitations of the technique have been discussed ( 2 , 3, 12). However, we have not seen any study in which, with an experimental approach, one has tried to estimate how different timing errors would affect the average and how great the timing errors are when averaging is done on a signal with realistic noise, e.g. noise arising during heavy exercise. To restrain the influence of noise occurring in a single lead we have used three ECG channels for triggering. When triggering is made on several ECG leads, the ‘spatial velocity’, i.e. the square root of the sum of the squares of the time derivatives, or some similar filtered function thereof, has most frequently been used as a trigger quantity. We have not seen any study where the actual slopes in all the ‘trigger leads’ are simultaneously taken into account. The procedure described here has this feature. Its major advantage is that the QRS upslope cannot be mistaken for the downslope. There are also some restrictions imposed on the shape of those artifacts or ectopic beats that can cause a false trigger point. For similar reasons the program adjusts the trigger level so that a QRS complex
Downloaded by [McMaster University] at 21:37 17 March 2016
786
Lisbet Jansson, B. Jonson & 0 . Werner
belonging to the majority type of complexes barely manages to provide a trigger point, whereas minor artifacts and ectopic beats with slow QRS gradients fail to do so. The present technique has so far been found adequate (Table I). It cannot be compared with other techniques owing to lack of data that allow evaluation of the exactness of the triggering procedures used by others. It is evident that a major variability in the QRS - for example, due to ventricular ectopic beats - could invalidate the average. We believe that the results obtained are encouraging, since two simple procedures made is possible to exclude a majority of ventricular ectopic beats. Further improvement would probably require some analysis of the morphology of the QRS - for example, its width and amplitude (12) an analysis that would call for significant additional computing time. The low-frequency noise is often very distinct during exercise, and to allow proper judgement of ST-T patterns the average must therefore include a number of heart cycles in the range of 60-120. Even with a large number of heart cycles, significant residual baseline instability in the average has to be expected now and then, since the respiratory base-line shifts are often periodic (3). Problems certainly arise when the heart rate and breathing rate are coupled. Since the aim of reducing noise by averaging is the quantitative measurement of ECG during exercise, averaging should definitely not be used as an alternative to enhancing the quality of the original signal. One of the most important ambitions in the development of the program has been to make it suitable for on-line use in the clinical routine and to make it easy to operate. To this end program flow is entirely directed from a simple numerical keyboard. Another important quality is that the same recording equipment is used for output of the original analogue signal, the computed average, and the relevant alphanumeric information. The use of time-coherent averaging is not restricted to the ECG: other heart-synchronous signals may also be averaged. Thus we have made preliminary tests on the use of phono-
cardiography and heart-synchronous pressure variations in the esophagus in combination with averaging, with good technical results. We hope that these techniques will allow easy study of the pathophysiology of the heart during exercise. ACKNOWLEDGEMENTS The present study was supported by the Swedish Medical Research Council, grant 04X-2872, and by the Swedish National Association against Chest and Heart Diseases.
REFERENCES 1. Arvedson, 0. Methods f o r Data Acquisition and Evaluation of Electrocardiograms and Vectorcardiograms with the Digital Computer. M.D. dissertation, Umei, 1968. 2. Blomqvist, G. The Frank lead exercise ECG, a quantitative study based on averaging technique and digital computer analysis. Acta med. scand. 178, Suppl. 440, 1965. 3. Evanich, M., Newberry, A. & Partridge, L. Some limitations on the removal of periodic noise by averaging. J. appl. Physiol. 33, 536, 1972.
4.Hornsten, T. & Bruce, R. Computed STforces of Frank and bipolar exercise electrocardiograms. Amer. Heart J. 78, 346, 1969. 5. McHenry, P. L., Stowe, D. E. & Lancaster, M. C. Computer quantitation of the ST-segment response during maximal treadmill exercise. Circulation 38, 691, 1968. 6. Rautaharju, P. M. Average transient computing techniques applied to electrocardiographic investigation. Chapt. 3 in Quantitative Spatial Vector Analysis of ECG Data by Electronic Computer Means. Ph. D. thesis, University of Minnesota, Minneapolis, June 1963. 7.Rautaharju, P. & Blackburn, H. The exercise
electrocardiogram. Experience in analysis of ‘noisy’ cardiograms with a small computer. Amer. Heart J. 69, 515, 1965. 8. Redwood, D. & Epstein, S. Uses and limita-
tions of stress testing in the evaluation of ischemic heart disease. Circulation 46, 1115, 1972. 9. Rose, G. A. & Blackburn, H. Cardiovascular survey methods. Wld HIth Org. Monog. Ser. 56, 1968. 10. Sheffield, L. T., Holt, J. H., Lester, F. M., Conroy, D. V. & Reeves, T. J. On-line analysis of exercise electrocardiograms. Circulation 40, 935, 1969. 11. Simoons, M. L., Bloom, H. B. K. & Tan, T. H.
Noise Reduction of ECG by Averaging Comparison of the diagnostic value of amp& tude and slope measurements from the STsegment during exercise. Abstract, VI European Congress of Cardiology, Madrid, 1972.
Downloaded by [McMaster University] at 21:37 17 March 2016
Received 15 November 1974 Accepted 7 July 1975
787
12. Wolf, H. K., MacInnes, P. J., Stock, S., Hilippi, R. K. & Rautaharju, P. M. Computer analysis of rest and exercise electrocardiograms. Computers biomed. Res. 5, 329, 1972.
ISSN: 0036-5513 (Print) 1502-7686 (Online) Journal homepage: http://www.tandfonline.com/loi/iclb20
Noise Reduction of ECG by Averaging an Experimental Study of the Procedure and a Validated Method Lisbet Jansson, B. Jonson & O. Werner To cite this article: Lisbet Jansson, B. Jonson & O. Werner (1975) Noise Reduction of ECG by Averaging an Experimental Study of the Procedure and a Validated Method, Scandinavian Journal of Clinical and Laboratory Investigation, 35:8, 781-787, DOI: 10.3109/00365517509095810 To link to this article: http://dx.doi.org/10.3109/00365517509095810
Published online: 08 Jul 2009.
Submit your article to this journal
Article views: 2
View related articles
Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=iclb20 Download by: [McMaster University]
Date: 17 March 2016, At: 21:37
Noise Reduction of ECG by Averaging An Experimental Study of the Procedure and a Validated Method
Downloaded by [McMaster University] at 21:37 17 March 2016
LISBET JANSSON, B. JONSON & 0. WERNER Dept. of Clinical Physiology, University of Lund, Lund, Sweden
Jansson, Lisbet, Jonson, B. & Werner, 0. Noise Reduction of ECG by Averaging. An Experimental Study of the Procedure and a Validated Method, Scand. J. elin. Lab. Invest. 35, 781-787, 1975. Electrical noise often hampers the interpretation of ECG and other heartsynchronous signals. Time-coherent averaging is a commonly used method to reduce the noise. This method has been implemented on a computer with 12,000 12-bit words of core memory, suitable for on-line use in any clinical laboratory. An analysis was made of which demands apply to the exactness of synchronization of added heart cycles. The method was tested with respect to these demands and also, on a limited scale, for its efficiency in eliminating ectopic beats from the average. Key-words: Averaging; digital computers; ECG; exercise test; heart synchronous events 0. Werner, M.D., Kliniskt fysiologiska avdelningen, Lasarettet, S-221 85 Lund, Sweden
The interpretation of ECG signals is often hampered by noise, especially during exercise. The noise may partly be compensated for by visual methods. However, it is well known that the ST-T segment is difficult to judge even qualitatively in the presence of slow base-line shifts (9). Real difficulties appear when one attempts to make quantitative measurements. The most important method to reduce the noise is a correct choice of the skin electrodes and of the ECG amplifier. The noise may be further reduced by filtering, which, however, implies signal distorsion (12). To avoid this, methods that make use of the fact that the ECG is a repetitive signal must be used. One may, for example, let the median of several ECG complexes represent the signal (10). However, this does not reduce high-frequency noise of a low amplitude. A commonly used method for noise reduction is time-coherent averaging (1, 2, 4-7, 11, 12).
Averaging has the advantages of being a simple and fast procedure and of reducing all sorts of random noise. Under ideal conditions, i.e. an infinite sampling rate, a correct timing, and a uniformly repetitive pattern, it does not distort the signal. In practice several problems must be solved. Most important, a correct time reference point must be found in each ECG complex, so that the various heart cycles are correctly synchronized when added onto each other. Second, ectopic beats must be eliminated, .since otherwise the prerequisite that only uniform patterns may be added is violated. In previous studies these problems were well recognized. Surprisingly, no attempts seem to have been made to analyze how exact the synchronization must be and whether the methods used are adequate in this respect. The present paper describes an analysis of this type. A program for averaging the ECG and other heart-synchronous signals was de-
782
Lisbet Jansson, B. Jonson & 0 . Werner
signed with a special effort to arrive at an adequate synchronization of heart cycles, also under unfavorable conditions such as a high noise level. A method is described that actually measures the timing error. The program was designed so that ectopic beats were eliminated from the average without too much computing time being spent. METHODS
Downloaded by [McMaster University] at 21:37 17 March 2016
Equipment (Fig. 1 )
The ECG and other heart-synchronous signals were recorded with a direct-writing Mingograph 81 (Siemens-Elema AB) and simultaneously fed into the computer for averaging. Program
Details on the algorithms used for averaging are available from the authors. Four signals were fed into the computer. Three of these were ECG leads used for triggering. The fourth signal could be ECG or any other heart-synchronous signal. A reference complex was checked for mag-
nitude and signs of slopes. On this basis incidences preliminarily regarded as QRS complexes were recognized and the intervals between them memorized. I n a later step each such incident was further analyzed with regard to magnitudes and signs of slopes in the triggering leads and with regard to the preceding and subsequent R-R intervals. Only such incidences, which were found to belong to a ‘majority type’ with regard to these measures, were finally added to the average. Special precaution was taken to find the most distinctive point in the QRS complex to be used for synchronization between beats. To obtain realistic noise signals, an ECG was recorded from leads CH2, CH,, CH5, and CH, during strenuous exercise in a subject. The ECG was stored on digital magnetic tape with a sampling rate of 400/s. It was ascertained that the rhythm was regular. Next, the digitalized ECG signals were reexamined by a special program, which again located the time reference points belonging to each QRS. The average obtained in each of the four channels was then subtracted beat for beat from the corresponding original signal, leaving signals with only insignificant residuals of the QRS complexes (see noise signal in Fig. 4). RESULTS
Fig. 1. General set-up of the system. Signals from the patient are recorded on a Mingograph 81 (Siemens-Elema AB). Four signals are fed into the computer and are AD-converted. Two analogue output channels connect the computer to the oscilloscope and the mingograph. The latter thus serves two purposes - both as a direct analogue writing equipment and as an output organ for the averaged signals and text from the computer. When the analogue output channels are occupied with information to the oscilloscope, a special relay provides zero signals to the mingograph. The numerical keyboard gives information to the computer and also directs program flow.
No errors in the average could be visually observed when it was computed under optimal circumstances, i.e. a high sampling rate (400/s) and a low noise in the original signal. Fig. 2 shows to what extent the average would be affected by errors in the time reference point. The introduction of artificial errors in the time reference point even within 2.5 ms yielded an average with slightly blurred fine details and slightly reduced QRS amplitudes. A sampling frequency of 200/s will theoretically limit the exactness of the time reference point to about k 2 . 5 ms. An average obtained at this sampling rate (Fig. 3) is similar to that obtained with a n artificial error in the time reference point within 2.5 ms (Fig. 2B). An illustration of program performance is
Noise Reduction of ECG by Averaging
1 mV
783
I
CHS
(i A 'i
I
Downloaded by [McMaster University] at 21:37 17 March 2016
I
1
i;
1
h Fig.2. Effects of varying the position of the time reference points. Pictures A-D show the averages of the same 27 beats recorded on digital magnetic tape from a subject at rest. I n picture A no interference with the time reference point was made. An average was obtained that was visually identical to the original signal. In pictures B-D, welldefined but varying displacement of the time reference point for each complex is made, before adding the complex to the sum. For example, picture C shows the result when the time reference points in the individual ECG complexes included in the average were moved from their correct positions by -5, -2.5, 0, 4-2.5, or + 5 ms. Minus denotes a shift to the left and plus a shift to the right. Each of the time shifts mentioned was used in every fifth beat. This should simulate a situation in which the position of the time reference point has an estimated error within 5 ms. Similarly, pictures B and D show the result when the correct position of the time reference point has errors within 2.5 ms and 7.5 ms, respectively, as defined for picture C. shown in Fig. 4. The realistic noise was added to a noise-free ECG from another subject. Averaging was done on the combined signal. A striking degree of noise reduction was achieved. Even fine details of the QRS complexes are clearly demonstrated (compare the complexes in Fig. 4B). A certain base-line instability re-
Fig. 3. Effects of two different sampling rates. The average of exactly the same stretch of signal (26 ECG complexes) using sampling rates of 400/s and 2OO/s, respectively.
mained, which is not surprising in view of the very marked breath-synchronous base-line shifts in the noise signal. The noise still remaining after averaging was estimated by subtracting the average of the noise-free ECG from the average of the noisy ECG. The remnants of the QRS complexes were very small, which indicates that the distorsion of the ECG
~~erog ofa24cornplexes
-
A r c r ~ g ~ o l I I r o m p l e x ~ ~Avcra8col 27noi'o-frcc ----
-10ms -ZSms -5mr -25ms Oms shifts to the left
n
LI' smallshift tothelsft
2 -
a'-
-'-
'
Jw I
*2.5ms +5mr *ZSms *10ms shifts to the right
"residual signal "
Fig. 5. Principle for estimation of error in time reference point. A. (Schematic drawing.) Pointwise subtraction of one identical waveform from another. When the two are synchronous, the result is of course uniformly zero. In the two pictures to the right, small time shifts between the waveforms above and below are introduced. A given time shift will after subtraction result in a characteristic residual waveform. B. In this picture the average ECG has been subtracted from itself. Before subtraction artificial time shifts were introduced. A certain time shift will cause residuals, characteristic of the time shift in shape and amplitude. These residuals were used as a scale to estimate time shifts in picture C. C. Above, an original ECG is shown. Below, the average of this ECG has been subtracted from the original beat-for-beat, so that the trigger points coincide with the corresponding point in the average - i.e. no artificial time shift has been used. It can be seen that the residual 'QRS complexes' below each QRS complex are small. This in turn indicates that the trigger procedure cannot have been greatly in error in any of the ECG complexes, since otherwise large residual 'QRS complexes' would have remained. A comparison with figure B shows that the residual complexes are in fact smaller than the ones that were arrived at when timing differences of +2.5 ms and -2.5 ms were used intentionally.
due to inexact triggering caused by noise must have been insignificant. The same procedure applied to another noise-free ECG gave a similar result. The result sJlown in Fig. 4 made it probable that the exactness of the time reference point was adequate in the presence of heavy realistic noise. To prove this further, a similar subtraction technique was used. The details are explained in relation to Fig. 5. The virtually noise-free ECGs sampled at 400/s from two persons (channels CH2, CH,, CH5, and CH,) were analyzed. The individual QRS complexes were visually uniform. The estimated variation in the time reference point was found to be less than 2.5 ms in all the 50 heart cycles studied in each subject. Next the noise signal was added. This introduced an estimated error in the time reference point of 5-7.5 ms in one beat, 2.5-5 ms in nine, and less than 2.5 ms in 90 beats. The test just described was made on two members of the staff during the development of the program; the same technique was applied to ten patients under realistic clinical circumstances. Ten patients taking an exercise test for clinical reasons were randomly selected, and a total of 1,457 beats were recorded at the highest work load reached. In 93% of all beats the error was below 2.5 ms; in 6%) 2.5-5 ms; in 0.7%, 5-7.5 ms; and in 1 beat, 7.5-10 ms (Table I). The results thus agreed with those of the staff. Triggering on artifacts (which were not included in the average because of the trigger interval criterion) occurred in 10 instances. As mentioned above, a beat was included into the average if both the beat-to-beat intervals and the qualified trigger quantity were within certain automatically determined limits. Firm conclusions on the ability of the program to reject ectopic beats must await the results of clinical applications. A preliminary study of a few patients allows us to predict that most ectopic beats will be rejected by the program (Table 11). It is notable that several of the ectopic beats that were erroneously accepted into the average were in fact fusion beats from one patient (I.A.). Erroneous acceptance of ectopic
Noise Reduction of ECG by Averaging
785
Table I. Estimated timing error of the qualified trigger point during exercise (at highest work load reached by each patient)
Patient K.A. D.O. K.B. B.O.
W.P. E.N. K.H.
Downloaded by [McMaster University] at 21:37 17 March 2016
B.M. A.L. M.V.
Work load Number of (W) heart beats* 150 40 150 90 150 90
0-2.5 ms 134 197 163 147 148 156 153 83 117 64 1362
143 207 171 149 158 161 175 I09 120 64 1457
150 150 100
50
Total
* Ectopic beats excluded. ** Not included in average because of
Estimated errors 2.5-5 ms 5-7.5 ms
Table 11. Number of ventricular ectopic beats accepted into the average in nine patients ~
All ventricular Ventricular ectopic beats
Initials
ectopic beats
H.N.* J.S.* H.M.
3 3
-
1
-
87 5 173 5 1 46 324
-
L.L.
accepted into average 3
1 27** -
H.H. S.T. Total 31 * These patients also had one atrial ectopic beat each, which was eliminated because of the trigger interval criterion. ** Of these at least twelve were judged to be fusion beats. 6 - Scand. J. din. Lab. Invest.
1 2 4
7 1
1
1
10 4 22 25 3 84
Triggering on artifacts**
2 1
10
1
10
trigger interval criterion.
beats occurred only at heart rates exceeding 120/min. It was noted that both the selection criterion based on the magnitudes and signs of slopes in the QRS and the criterion based on beat-tobeat intervals were necessary to achieve these results. Thus 59 of the ventricular ectopic beats were rejected only because the first-mentioned criterion was not satisfied, whereas 27 beats were rejected only because of the beat-to-beat intervals. The latter mechanism failed especially often at a high heart rate in spite of the narrowing trigger interval limits that were used.
U.Q. K.F. I.A.
9 3 7
7.5-10 ms
DISCUSSION The averaging technique has been used during the last ten years to reduce the noise in the ECG. Various theoretical limitations of the technique have been discussed ( 2 , 3, 12). However, we have not seen any study in which, with an experimental approach, one has tried to estimate how different timing errors would affect the average and how great the timing errors are when averaging is done on a signal with realistic noise, e.g. noise arising during heavy exercise. To restrain the influence of noise occurring in a single lead we have used three ECG channels for triggering. When triggering is made on several ECG leads, the ‘spatial velocity’, i.e. the square root of the sum of the squares of the time derivatives, or some similar filtered function thereof, has most frequently been used as a trigger quantity. We have not seen any study where the actual slopes in all the ‘trigger leads’ are simultaneously taken into account. The procedure described here has this feature. Its major advantage is that the QRS upslope cannot be mistaken for the downslope. There are also some restrictions imposed on the shape of those artifacts or ectopic beats that can cause a false trigger point. For similar reasons the program adjusts the trigger level so that a QRS complex
Downloaded by [McMaster University] at 21:37 17 March 2016
786
Lisbet Jansson, B. Jonson & 0 . Werner
belonging to the majority type of complexes barely manages to provide a trigger point, whereas minor artifacts and ectopic beats with slow QRS gradients fail to do so. The present technique has so far been found adequate (Table I). It cannot be compared with other techniques owing to lack of data that allow evaluation of the exactness of the triggering procedures used by others. It is evident that a major variability in the QRS - for example, due to ventricular ectopic beats - could invalidate the average. We believe that the results obtained are encouraging, since two simple procedures made is possible to exclude a majority of ventricular ectopic beats. Further improvement would probably require some analysis of the morphology of the QRS - for example, its width and amplitude (12) an analysis that would call for significant additional computing time. The low-frequency noise is often very distinct during exercise, and to allow proper judgement of ST-T patterns the average must therefore include a number of heart cycles in the range of 60-120. Even with a large number of heart cycles, significant residual baseline instability in the average has to be expected now and then, since the respiratory base-line shifts are often periodic (3). Problems certainly arise when the heart rate and breathing rate are coupled. Since the aim of reducing noise by averaging is the quantitative measurement of ECG during exercise, averaging should definitely not be used as an alternative to enhancing the quality of the original signal. One of the most important ambitions in the development of the program has been to make it suitable for on-line use in the clinical routine and to make it easy to operate. To this end program flow is entirely directed from a simple numerical keyboard. Another important quality is that the same recording equipment is used for output of the original analogue signal, the computed average, and the relevant alphanumeric information. The use of time-coherent averaging is not restricted to the ECG: other heart-synchronous signals may also be averaged. Thus we have made preliminary tests on the use of phono-
cardiography and heart-synchronous pressure variations in the esophagus in combination with averaging, with good technical results. We hope that these techniques will allow easy study of the pathophysiology of the heart during exercise. ACKNOWLEDGEMENTS The present study was supported by the Swedish Medical Research Council, grant 04X-2872, and by the Swedish National Association against Chest and Heart Diseases.
REFERENCES 1. Arvedson, 0. Methods f o r Data Acquisition and Evaluation of Electrocardiograms and Vectorcardiograms with the Digital Computer. M.D. dissertation, Umei, 1968. 2. Blomqvist, G. The Frank lead exercise ECG, a quantitative study based on averaging technique and digital computer analysis. Acta med. scand. 178, Suppl. 440, 1965. 3. Evanich, M., Newberry, A. & Partridge, L. Some limitations on the removal of periodic noise by averaging. J. appl. Physiol. 33, 536, 1972.
4.Hornsten, T. & Bruce, R. Computed STforces of Frank and bipolar exercise electrocardiograms. Amer. Heart J. 78, 346, 1969. 5. McHenry, P. L., Stowe, D. E. & Lancaster, M. C. Computer quantitation of the ST-segment response during maximal treadmill exercise. Circulation 38, 691, 1968. 6. Rautaharju, P. M. Average transient computing techniques applied to electrocardiographic investigation. Chapt. 3 in Quantitative Spatial Vector Analysis of ECG Data by Electronic Computer Means. Ph. D. thesis, University of Minnesota, Minneapolis, June 1963. 7.Rautaharju, P. & Blackburn, H. The exercise
electrocardiogram. Experience in analysis of ‘noisy’ cardiograms with a small computer. Amer. Heart J. 69, 515, 1965. 8. Redwood, D. & Epstein, S. Uses and limita-
tions of stress testing in the evaluation of ischemic heart disease. Circulation 46, 1115, 1972. 9. Rose, G. A. & Blackburn, H. Cardiovascular survey methods. Wld HIth Org. Monog. Ser. 56, 1968. 10. Sheffield, L. T., Holt, J. H., Lester, F. M., Conroy, D. V. & Reeves, T. J. On-line analysis of exercise electrocardiograms. Circulation 40, 935, 1969. 11. Simoons, M. L., Bloom, H. B. K. & Tan, T. H.
Noise Reduction of ECG by Averaging Comparison of the diagnostic value of amp& tude and slope measurements from the STsegment during exercise. Abstract, VI European Congress of Cardiology, Madrid, 1972.
Downloaded by [McMaster University] at 21:37 17 March 2016
Received 15 November 1974 Accepted 7 July 1975
787
12. Wolf, H. K., MacInnes, P. J., Stock, S., Hilippi, R. K. & Rautaharju, P. M. Computer analysis of rest and exercise electrocardiograms. Computers biomed. Res. 5, 329, 1972.
