Automatic Recognition of Electrocardiographic
Waves by Digital Computer
By FRIEDEMANN W. STALLMANN, SC.D., AND
HUBERT V. PIPBERGER, M.D.
A
Downloaded from http://circres.ahajournals.org/ by guest on June 18, 2017
PPLICATIONS of electronic computers
.in analysis of medical and biological data
are of increasing interest. Large-scale computers offer the advantages of high speed of
operation, relatively great accuracy, and extreme versatility. Great amounts of data can
be processed in a relatively short time. Thus,
extensive statistical studies, mass screening
of the population, and data handling in large
medical centers should be facilitated.
The electrocardiogram was chosen for a
pilot study in automatic processing of medical
data because of its widespread use as a diagnostic aid. Since even the largest available
digital computers are limited in memory capacity, a study in data reduction had to precede the data processing proper. It could be
shown that corrected orthogonal three-lead
electrocardiograms contain the same clinical
information which is being used in conventional 12-lead analysis.1 Thus, data reduction
by a factor of 4:1 could be accomplished without sacrificing clinical information. An automatic procedure for electroeardiographic data
conversion from the original analog into digital form has been described previously.2 Commercially available digital computers can be
used, once the electrocardiogram is converted
into proper digital format.
Since the electrocardiogram consists of several distinct wave forms with different electrophysiological significance, the prime objective of any automatic electrocardiographic
analysis has to be the development of a computer program which allows automatic recognition of each electrocardiographic wave.
Their beginning and end have to be determined accurately in order to obtain measurement^ of duration of the complexes and of
the time intervals in between them. Although
automatic screening of electrocardiograms for
certain abnormalities is feasible when the time
interval over a complete cardiac cycle is obtained, 3 a detailed analysis of the electrocardiogram is not possible without proper
identification of each wave. An automatic
procedure which allows accurate determination of beginning and end of P waves, QRS
complexes, and end of T waves is described
in the present report. The computer program
was successfully tested in a series of 395 electrocardiograms.
Methods
Out of a total of 395 eleetroeardiographie records used for the present study, 173 were read as
normal and the remainder as abnormal. The technically poorest records were selected from a taperecorded electrocardiogram library of approximately 2,500 cases. In most of these tracings, a
large amount of interference, such as 60-eycle A.C.
noise and/or muscle tremor, was present. It was
felt that an automatic wave recognition program
could be tested more rigidly with technically poor
tracings. Frequently, small signals, such as P and
Q waves, were hardly visible. Although the amount
of noise in most records was much greater than
that of average tracings in clinical electroeardiography, it is common experience that under unfavorable conditions such tracings may be encountered. Since the described procedure is
proposed for routine use, such unfavorable conditions have to be taken into account.
From the Veterans Administration Hospital, Mt.
Alto, and the Department of Medicine, Georgetown
University, School of Medicine, "Washington, D. C.
Supported in part by a research grant of the
American Heart Association and a Public Health
Service research grant (H-4576) of the National
Heart Institute. Work performed under contract
with the Heart Disease Control Program, U. S.
Public Health Service.
Presented at the Fifth Annual Meeting of the
Biophysical Society, February 16 to 18, 1961, St.
Louis, Missouri.
Received for publication May 11, 1961.
1138
Circulation Research, Volume IX, November 1961
ELECTROCARDIOGRAPHS "WAVES
1139
V/1
Downloaded from http://circres.ahajournals.org/ by guest on June 18, 2017
Figure 1
Frequency response function of the filtering procedure used for elimination of noise. The ordinate
(F) indicates the ratio between filtered and unfiltered electrocardiographie signals for the frequencies shown on the abscissa. For details of the
procedure, see Appendix.
Sehmitt's SVEC III lead system4 was used in
130 records. The remainder were taken with
Frank's lead system.5 For the purpose of wave
recognition, the choice of a specific lead system
is of no consequence. The three orthogonal leads
were recorded simultaneously on magnetic analog
tape, using FM channels as previously described.2
A special purpose analog-to-digital converter0
was used for digitizing the records. The sampling
rate was 1,000 per second for each lead. A threelead record of one cardiac cycle, comprising P,
QRS, and T, contains 1,800 digits on the average.
For further numerical analysis, an IBM 704 digital computer was used.
First, a filtering procedure was required hecause of the large amount of noise. The initial
process consisted, therefore, in determination of
the frequency spectrum of electrocardiographie
signals and noise. It was found that the signal-tonoise ratio was high below and low above 60 c.p..s.
An essential part of the noise consisted of 60
c.p.s. interference and its higher harmonics. Frequencies above the 60-c.p.s. level had to be eliminated, therefore. The frequency response function
for the filtering of the records is shown in figure
1. Details of the computational process are given
in the appendix to this report. Figure 2 shows
a typical example of the performance of the filter.
The noise superimposed on the electrocardiographie records was almost completely eliminated.
However, the electrocardiographie signals, especially the time of beginning and end of each wave,
•were not significantly affected. Computation of
time intervals in filtered and unfiltered records,
therefore, were identical. After determining the
Circulation Research, Volume IX, November 19G1
Figure 2
The three electrocardiographie records on top represent three orthogonal leads (y, x, z from top to
bottom). Underneath the tracings shown in heavy
lines are enlarged plots of the P wave (middle)
and QRS complex (beloiv) of the digitized lead x.
The same wave forms are shown after the filtering
procedure as thin lines. Note that the extraneous
noise of the original records is almost completely
eliminated, whereas the electrocardiographie signals are not noticeably affected.
time limits of all waves, the points were transferred to the original unfiltered electrocardiogram.
Thus, the higher frequency components which
might have been suppressed in the filtering process
were maintained for further analvsis.
1140
STALLMANN, PIPBERGER
Table 1
Comparison Beticeen Computed and Visual Measurements of Electroccirdiographic Wave
Durations and Time Intervals Obtained from 106 Records'
Computed minus visual
determinations of
wave durations and time
intervals
P wave
P-R interval
23
7
9
16
44
1
0
0
2
5
16
58
14
5
More than -f" 32 msec.
Between -f- 32 and + 24 msec.
Between -f- 24 and + 16 msec.
Between -+- 16 and -f- 8 msec.
Between + 8 and — 8 msec.
Between — 8 and — 16 msec.
More than — 16 msec.
QRS duration Q-T interval
6
9
24
42
19
0
0
19
9
11
25
30
4
3
*For method of measurement, see text.
Downloaded from http://circres.ahajournals.org/ by guest on June 18, 2017
Beginning and end of eleetrocardiographic
waves can be defined through voltage changes of
the orthogonal leads Ax, Ay, Az per unit time At.
Ax = x(t + At) — x(t), etc.
Since the three orthogonal leads are considered as
scalar components of a vector loop in three-dimensional space, the three voltage changes can
also be combined to a vector whose magnitude is
called spatitil velocity V.
V =
A/AX 2 + Ay-
+
Az2
At
This entity proved very useful for wave recognition. Spatial velocities exceeding 3 /xY per msec,
were found only in P waves, QRS complexes, and
T waves but not during T-P and P-R intervals or
S-T segments.
For a given cardiac cycle, the computer then
proceeded as follows:
1. The point where the spatial velocity Y
reached a maximum was determined. In cases
where the same maximum, value was reached more
than once, the first maximum was used.
2. Starting from the point determined first, V
was determined in forward and backward direction
until values below 3 /xY per msec, were found.
These points were defined as beginning and end of
QRS, respectively.
3. From the beginning of QRS, V was determined again in retrograde direction until a value
exceeding 3 /xY per msec, was found. This point
indicated the end of the P wave. If such a value
for V was not found, it was assumed that no P
wave was present.
4. In records with P waves, the computer
started determinations of V from the beginning
of the tracing until the same critical value for Y
was exceeded. This point was defined as the beginning of the P wave. In cases where this point
differed from the end of P by a few milliseconds
only, it was assumed that the beginning of P had
been missed.
5. Starting at the end of the record, V was
determined in retrograde direction until the specified value for V was exceeded. This point indicated the end of the T wave. Since the beginning
of T waves is poorly defined and not used in clinical electrocardiography, a search for this point
was not included in the computer program.
Computation time for the entire wave recognition program was, on the average, 15 seconds for
each record.
Results
At first, the computer results were compared with the analog records iu order to determine whether all waves present iu the
tracings had beeu properly identified. The
findings were as follows:
1. In 22 cases, no P wave had been recognized by the computer. The analog records of
these cases showed premature ventricular
systoles without preceding P waves in 10 instances. In nine further cases, auricular fibrillation was present. The regular P waves had
been replaced by irregular F waves of small
amplitude. Three additional cases showed
nodal rhythm with P waves hidden in the
QRS complex. In all cases with normal sinus
rhythm, the P waves had been properly identified.
2. In nine cases, the beginning of P waves
had not been recognized by the computer. Six
of these were premature ventricular systoles,
and two records showed auricular fibrillation.
F waves or other unidentified potential
changes preceding QRS had apparently been
strong enough to exceed the set limits of spatial velocity for a short period of time. A
Circulation Research, Volume IX, Novrmber 1901
ELECTROCARDIOGRAPHS WAVES
Downloaded from http://circres.ahajournals.org/ by guest on June 18, 2017
further ease showed a "wandering pacemaker." The beginning of P was missed by
the computer because of a very gradual onset
of this wave. The potential changes were only
strong enough during the downstroke to exceed the limits of spatial velocity.
3. In five more cases, the analog records
showed auricular fibrillation, but the computer had identified P waves of extremely long
duration.
4. In four out of a total of 16 complexes
representing premature ventricular systoles,
the end of QES was not properly identified.
The QRS complexes were followed immediately by T waves without S-T segments. In these
cases, either the QRS-T was taken by the computer as one complex, or the peak of T was
mistaken as end of QRS.
5. The end of QRS was falsely determined
in one case with right ventricular conduction
defect where the QRS contained a plateau
without potential changes. This plateau was
falsely recognized as end of QRS.
Thus, not more than 24 points were falsely
determined by the computer. This figure represents 1.3 per cent of a total of 1,911 points
expected to be determined.
In a further test, a random sample of 106
records out of the total series was selected for
a more exact comparison between computed
and visual measurements. The original three
orthogonal leads were reproduced from magnetic analog tape, displayed on an oscilloscope
screen, photographed on 35-mm. film, and enlarged five times for viewing. This display
allowed consistent visual time measurements
with a maximal accuracy down to 5 msec.
Differences between visual measurements and
computer results are listed in table 1. The
following three major types of discrepancies
were found:
1. QRS durations obtained from the computer were systematically longer. This discrepancy was due to the gradual ending of the
complex which can hardly be pinpointed visually.
2. In about 20 per cent of the cases, the computed Q-T interval was eonsiderabty longer
than that obtained visually. In all these cases,
Circulation Research, Volume IX, November 1861
1141
Table 2
Ten Consecutive Cardiac Cycles of Three Randomly Selected Records Measured by Three Independent Readers*
Tracing
number
4-58
7-58
7-60
4-58
7-58
7-60
4-58
7-58
7-60
4-58
7-58
7-60
ECG time
interval
measured
P
P
P
PR
PR
PR
QRS
QRS
QRS
Q-T
Q-T
Q-T
Maximum
discrepancies
between 10
consecutive
cardiac cycles
[1 reader]
(msec.)
25
11
14
29
10
17
10
14
11
28
22
18
Maximum
discrepancies
between
measurements
of 3 independent
readers [10 cardiac
cycles]
(msec.)
25
21
24
32
10
21
14
14
14
35
29
32
*If consistency in measurements is assumed for
each reader, the second column from the right indicates variability in duration from one heart beat to
the next. Results in the last column indicate discrepancies between the three readers for the same
records.
small potential changes followed the T waves
which had not been recognized visually.
Whether or not these potential changes belonged to the T waves or represented small U
waves could not be determined.
3. In about 20 per cent of the records, the
computed P duration was also considerably
longer than that determined visually. This
difference was due to small potential changes
during the P-R segment.
In a third test, the discrepancies between
visual and computed measurements were compared with the beat-to-beat variability of wave
durations and with the consistency of visual
time measurements for different observers.
The results of this comparison are given in
table 2. Ten consecutive cardiac cycles in each
of three randomly selected records were measured by three independent readers. Differences in time measurements of a sequence of
cardiac cycles were largest for Q-T intervals,
P-R intervals, the P durations, in decreasing
order. Even for QRS complexes, with their
more abrupt onset, differences of up to 14
1142
msec, were found in 10 consecutive heart
beats. Discrepancies between different observers were also largest for Q-T intervals and
smallest for QRS complexes.
Downloaded from http://circres.ahajournals.org/ by guest on June 18, 2017
Discussion
With the described method for automatic
electrocardiographic wave recognition by digital computer, the durations of P waves, P-R
intervals, P-R segments, QRS complexes, and
Q-T intervals are obtained. These data are the
basic parameters of any ECG analysis. In
addition, the separation of the different waves
provides the basis for a more detailed analysis
of each of them.
Since in the described procedure the computer determined the beginning and end of
electrocardiographic waves by the use of only
one measurement, the spatial velocity, a failure rate of 1.3 per cent appears very small.
Except for one case with right ventricular
conduction defect, failures of the method were
encountered only in cardiac arrhythmias.
Automatic analysis of irregularities in cardiac
rhythm will require a new set of computer
instructions which is being prepared at present. The described procedure can serve as the
basis for such a program. Even in its present
form, with a failure rate of 1.3 per cent, the
described procedure can be recommended for
large-scale statistical and epidemiological
studies where such a failure rate cannot be
considered excessive.
The beginning and end of electrocardiographic waAres were not noticeably changed
through the relatively strong filtering procedure. Since it is known, however, that electrocardiograms may contain higher frequency
components7 which are affected by the filter,
it is necessarjr to transfer the computed endpoints of each wave to the original unfiltered
record. Thus, higher frequency information
content is maintained for further analysis,
and the filtering procedure is used as an intermediary step only.
The systematic discrepancies between computed and visual determinations of the end
of P, QRS, and T appeared to be due to true
potential variations. These variations were too
STALLMANN, PIPBBEGER
small to be recognized visually when enlargements of analog records, as described above,
were used. These enlargements, however, were
considerably greater in size than conventional
tracings used in clinical electrocardiography.
It remains an open question whether, in these
cases, the computed or visual measurements
represent " t r u e " wave durations. It is fairly
simple, however, to modify the described computer program in order to simulate visual
measurements. This will require only an increase in the critical value for spatial velocity
for wave recognition.
When the limits of visual accuracy and
beat-to-beat variations in wave duration were
taken into account, the computed results indicated a satisfactory agreement with visual
measurements. The discrepancies between visual readings of different observers emphasize
the importance of consistency. In this respect,
the performance of the computer is superior
to that of the electrocardiographer because a
constant mathematical term is used for delineation of electrocardiographie waves.
Summary
A digital computer program for automatic
recognition of electrocardiographic waves has
been described. First, a filtering procedure
was applied in order to eliminate extraneous
noise. Consequently, the spatial velocity derived from three orthogonal electrocardiographic leads was determined for an entire
cardiac cycle. It was found that a critical
value of 3 /xV per msec, was never exceeded
during T-P intervals, P-R segments, and S-T
segments. This limit for the spatial velocity
could be used to indicate the beginning and
end of electrocardiographic waves. The method
was tested in a series of 395 records. The computer failed in 1.3 per cent of all expected
measurements. With one exception, failures
were encountered only in eases with cardiac
arrhythmias but not with regular sinus
rhythm. Comparison between computed and
visual time measurements showed close agreement, especially when limits of visual accuracy and beat-to-beat variations in wave duration were taken into account. The described
Circulation Research, Volume IX, November 1961
ELECTROCARDIOGRAPHIC WAVES
procedure can serve as the basis for a complete electroeardiographic analysis by digital
computer.
Appendix
A filter is mathematically a transformation of
time series. Let x(t) be a function of time given
for the points t = 0, ± 1 , ± 2 , . . . . , which represents a signal disturbed by noise. By choosing a
suitable computational process, the function x(t)
can then be transformed into another function of
the same type x*(t), which is assumed to be a
better representation of the original signal. If
we restrict ourselves to linear and finite filters,
since the computer can perform only finite operations, the filter procedure can be written explicitly
as
1143
This gives
=
32- ( 1 + ~ COS I 6 r ) ' T
_ _ 1
= 16 .
~
"64'
The frequency response function is then
F(a.) = — sin 16
cotg —
2 sin
COS u) -
COS 7T
as shown in figure 1.
Downloaded from http://circres.ahajournals.org/ by guest on June 18, 2017
+N
References
1. PIPBERGER, H . V., BlALEK, S. M., PERLOPT, J . K.,
AND SCHNAPER, H. W.: Correlation of clinical
T = -N
This is a moving average type of operation with a
weight function g(r).
If x(t) is a pure harmonic function
i <o t
x(t) = C e
then x*(t) is also pure harmonic
x«(t) = 0( u ) - C e ' "
with
0(M) z= r((,)) e
information in the standard 12-lead ECG and
in a corrected orthogonal 3-lead ECG. Am.
Heart J. 61: 34, 1961.
2. PlPBERGER, H . V., FREIS, E . D., TABACK, L.,
AND MASON, H. L.: Preparation of eleetro-
cardiographic data for analysis by digital
electronic computer. Circulation 21: 413, 1960.
3. PIPBERGER, H. V., ARMS, R. J., AND STALLMANN,
=
_^> g (T) e
T = -N
Here F(<o) indicates the frequency response function and 8 (<i)) the phase shift introduced by the
filter. The phase shift is zero if, and only if, the
weight function g ( r ) is symmetric
g(-r) = g(r) •
Only such filters will be considered here.
The weight function g(r) can be chosen so that
F (to) has given values onN + 1 points &> = k— ,
k = 0, 1, 2 , . . . . N. In our case, the main part of
the noise was close to — (sampling rate 1,000/s :
16
16 ~ 60 c.p.s.) and its higher harmonics. Therefoi'e, we have chosen N = 16 and
- .. _ 1 for k = 0, 1
16 ;
0 for k = 2, 3,....16 .
Circulation Research, Volume IX, November 1961
F. W.: Automatic screening of normal and
abnormal electrocardiograms by means of digital electronic computer. Proc. Soc. Exper.
Biol. & Med. 106: 130, 1961.
4. SCHMITT, O. H., AND SIMONSON, E.:
Present
status of veetorcardiography. A.M.A. Arch.
Int. Med. 96: 574, 1955.
5. FRANK, E.: Accurate, clinically practical system for spatial veetorcardiography. Circulation 13: 737, 1956.
6. TABACK, L., MARDEN, E., MASON, H. L., AND
PIPBERGER, H. V.: Digital recording of electrocardiographie data for anatysis by digital
computer. IRE Transactions on Medical Electronics ME-6: 167, 1959.
7. LANGNER, P . H., JR., AND GESELOWITZ, D. B.:
Characteristics of the frequency spectrum in
the normal electrocardiogram and in subjects
following myocardial infarction. Circulation
Research 8: 577, 1960.
Automatic Recognition of Electrocardiographic Waves by Digital Computer
FRIEDEMANN W. STALLMANN and HUBERT V. PIPBERGER
Downloaded from http://circres.ahajournals.org/ by guest on June 18, 2017
Circ Res. 1961;9:1138-1143
doi: 10.1161/01.RES.9.6.1138
Circulation Research is published by the American Heart Association, 7272 Greenville Avenue, Dallas, TX 75231
Copyright © 1961 American Heart Association, Inc. All rights reserved.
Print ISSN: 0009-7330. Online ISSN: 1524-4571
The online version of this article, along with updated information and services, is located on the
World Wide Web at:
http://circres.ahajournals.org/content/9/6/1138
Permissions: Requests for permissions to reproduce figures, tables, or portions of articles originally published in
Circulation Research can be obtained via RightsLink, a service of the Copyright Clearance Center, not the
Editorial Office. Once the online version of the published article for which permission is being requested is
located, click Request Permissions in the middle column of the Web page under Services. Further information
about this process is available in the Permissions and Rights Question and Answer document.
Reprints: Information about reprints can be found online at:
http://www.lww.com/reprints
Subscriptions: Information about subscribing to Circulation Research is online at:
http://circres.ahajournals.org//subscriptions/