Download Recommendations for Standardization and

730
Special Report
Recommendations for Standardization and
Specifications in Automated Electrocardiography:
Bandwidth and Digital Signal Processing
A Report for Health Professionals by an Ad Hoc Writing
Group of the Committee on Electrocardiography and Cardiac
Electrophysiology of the Council on Clinical Cardiology,
American Heart Association
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
James J. Bailey, MD, Alan S. Berson, PhD, Arthur Garson Jr., MD, Leo G. Horan, MD,
Peter W. Macfarlane, PhD, David W. Mortara, PhD, Christoph Zywietz, MSc
R ecent technologic advances and published
studies point to the need to reevaluate
requirements for recording and analysis of
electrocardiographic (ECG) data. Several issues have
arisen from recent developments.
First, approximately 52 million ECGs were processed by computer in 1987 (up from 4 million in
1975), and more than 15,000 digital ECG systems are
now in use (up from 85 in 1975).1 In addition,
computer-assisted electrocardiography is spreading
rapidly from academic medical centers to private
hospitals, clinics, and physicians' offices. A major
factor in this diffusion of technology is the introduction of inexpensive (less than $5,000) stand-alone
interpretative carts using microchip technology.'
Second, a large number of ECG systems, particularly stand-alone carts, convert data into digital form
immediately after amplification and process it digitally. This type of processing, combined with preamplifiers with lower noise characteristics than those
available earlier, permits higher resolutions and
signal-to-noise ratios than conventional linear analog
systems.
Third, high-frequency components and lowamplitude signals in the ECG known primarily to
researchers are now regarded with wider interest.
Distinguishing between the needs of researchers and
requirements for routine ECG recording has become
more difficult.
Approved by the SAC/Steering Review Committee of the American Heart Association in October 1989.
Single reprints are free; ask for 00-7501.
Correspondence should be addressed to the American Heart
Association, Office of Scientific Affairs, 7320 Greenville Avenue,
Dallas, TX 75231.
Fourth, increased recognition of the value of serial
analysis of ECGs requires storage and retrieval of
records and, often, transmission of raw ECG data over
telephone lines. Data compression techniques for
efficient handling of large volumes of data will probably become widespread over the next several years.
There are a variety of methods for describing the
features of these often complex systems. Some guidelines for specifying performance would be useful.
Fifth, a consortium of European investigators
seeking common standards in quantitative electrocardiography (CSE) has found a variety of systematic
biases and inaccuracies in measurements by 16 different computer programs for ECG analysis.2-5 As
they point out, investigators cannot exchange diagnostic criteria until measurements by the various
programs are uniform. Thus, criteria cannot be standardized without standardization of metrology, morphology, and terminology among all programs.
The recommendations of the American Heart
Association (AHA) have previously focused on conventional ECG strip chart recording, although in
1975, the AHA issued guidelines for tape recording
and analog-to-digital conversion.6 7 In 1983, the
American National Standards Institute (ANSI) published a report on analog specifications.8 With the
exception of bandwidth, the Committee on Electrocardiography and Cardiac Electrophysiology finds
the analog standards recommended in these reports
still acceptable. This report is primarily concerned
with digital signal processing, which has become an
integral function in many ECG systems.
Digital processing of the ECG signal requires analogto-digital conversion and includes filtering, data compression, teletransmission, storage or retrieval onto or
AlA Scientific Council Recommendations in Electrocardiography
from digital media, and measurement of waveforms. In
every instance, the result of digital processing is
judged by the fidelity with which it represents the
original ECG signal. This fidelity underlies the accuracy and uniformity of metrology (i.e., waveform
measurement) needed for final analysis and diagnostic interpretation of the ECG. This is the central
theme of the discussions that follow.
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
Signal Fidelity After Digital Processing
Fidelity may be defined as a measure of how
closely the result of digital processing represents the
"true" input signal; that is, the greater the error of a
digital processing method, the lower the fidelity.
Output data after digital transmission can be compared bit-wise with input data. After digital storage
of raw data, the retrieved data can be similarly tested
against input data. After data compression, the data
can be reconstructed and compared with the original
data, but after analog-to-digital conversion, the test
of fidelity is how well the analog signal is represented
by digital data. There are several ways to perform
such a test, but their details and relative merits are
beyond the scope of this discussion.
Testing the effect of digital filtering is a complicated process. One approach is to use an extremely
"clean" ECG signal and arbitrarily add various types
of noise (muscle, baseline wander, mains hum, etc.)
at various amplitudes before filtering. The output
data after filtering can then be compared with the
original ECG data.9,10
Requirements for fidelity may vary, depending on
the capabilities of the digital processing method.
Many systems for data transmission or storage of raw
data are capable of complete duplication of input
data down to the least significant bit, which is unrealistic for filtering and inappropriate for analogto-digital conversion. In these cases, the requirement
for fidelity will depend on how the data are to be
used.
For example, if the purpose is to visually display
the ECG for routine manual reading as currently
practiced, then previous recommendations78 are relevant. In its 1975 report (section A1), the AHA
recommends a limit of 0.25 mm error (25 ,uV) for
direct-writers at 10 mm/mV sensitivity for peakto-peak amplitudes of less than 5 mm; otherwise,
error should be limited to +5%.7 At half-standard
(5 mm/mV), this means the greater of 5% or 50 ,uV.
The ANSI standard (section 3.2.7.1) suggests the
greater of 10% or 50 JuV.8 However, the AHA's
specification (the greater of +5% or ±25 ,V) is
readily achievable by current digital technology.
If, on the other hand, the purpose is to extract
measurements from digital data and classify waveforms diagnostically, fidelity must be improved. (The
high fidelity required for surface recording of HisPurkinje activity and late potentials is not considered
here.) The AHA report on stand-alone amplifiers
(section B) provides one possible guideline: an error
731
limit of ±2% or 10 ,uV, whichever is greater, for
peak-to-peak deflections.7
Several error parameters can be used to quantify
the differences between input and output data: absolute error at any point or over any interval, root mean
square (RMS) error, maximum peak relative error
(MPE), timing errors, and waveform detection
errors. The fidelity criteria met by an automated
ECG system should be publicly stated. The following
examples of fidelity criteria are supported by either
theoretical investigations or experimental studies
reported in the literature:
Fidelity Criteria
For Routine Visual Reading
Fl: Deviation of recorded output from an exact
linear representation of the input signal may
not exceed 25 gV or 5%, whichever is greater.
(See AHA report, section A1.7)
For Morphological Diagnosis by Digital Program
F2: RMS error over the PQRST complex may not
exceed 10 gV.H
F3: Error in peak-to-peak deflections may not
exceed 10 gV or 2%, whichever is greater.
(See AHA report, section B.7)
F4: Mean squared error divided by mean squared
amplitude of a deflection may not exceed
1%.12
F5: QRS deflections with .20 ,uV amplitude and
>6 msec duration should be detected (threshold defined by CSE4'13).
F6: MPE may not exceed 10% over interval of any
QRS deflection . 20 gV and . 12 msec. 14,15
For Digital Transmission or Storage/Retrieval of Data
F7: Corresponding samples of input data and
reconstructed data should differ by less than
10 gV (half of CSE threshold; see F5 above).
Data Compression, Storage, and Retrieval
Operational and economic considerations related
to storage, retrieval, and teletransmission of ECG
data have motivated a search for efficient data compression techniques. The long-recognized importance of serial ECG analysis has made storage and
retrieval an integral part of computerized ECG systems. Efforts to improve ECG fidelity through higher
sampling rates and higher resolution will only
increase the need for efficient data compression
techniques.
Data compression can be achieved in many different ways, depending on the purpose of the data
subsequently retrieved. At one extreme, the data
compression scheme may be designed to permit
reconstruction of the entire cardiac cycle to the least
significant bit or with a correspondingly small error.
At the other extreme, data can be stored with less
accuracy if, for example, a display of the ECG for
rhythm determination only is required. Various intermediate forms of data compression can also be
visualized.
732
Special Report Circulation Vol 81, No 2, February 1990
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
One method for serial comparison16 stores only
certain amplitude values used by a particular comparison program. Such values may be derived statistically from several cycles or from a cardiac complex
that can be a single, selected cycle with very low
noise.
Another method uses algorithms for prediction or
interpolation17 so that the difference between an
actual value and a predicted or interpolated value
can be used in subsequent storage/retrieval techniques. Its advantage is that the difference will have
a much smaller amplitude and thus a smaller data
storage requirement than the actual value itself. It is
possible to go further and code the difference values
(or residuals) so that the most frequently occurring
residuals are represented by the shortest code. This is
one form of Huffman coding.18 Care must be taken
with this approach that errors due to truncation are
not introduced. This is discussed elsewhere.19
Still another approach is to form the average
cardiac cycle and subtract it from each PQRST cycle
in the record after suitable time alignment. The
residual signal should then have low amplitude and
can be treated by the Huffman coding methods.
Use of orthogonal functions to effect data compression is advocated by some. Recent studies have
been based on the Karhunen-Loeve (K-L) series,20-22
which allow ECG-like waveforms to be used as basis
functions and which in a linear combination represent the waveform to be compressed. The advantage
of this approach is the need normally to store up to
only 20 coefficients for each waveform to be compressed. The basis functions need be stored only
once.
Data compression may be implemented with techniques called "fan" algorithms, which are based on
first-order interpolation with two degrees of freedom.
The fan method does not require the storage of all
actual data points between the last transmitted point
and the present point during program execution.
Blanchard and Barr23 compared five data compression techniques, including a fan method, for cardiac
waveforms. The fan method produced reconstructed
waveforms with errors comparable to or better than
the other methods. Regardless of the method to
achieve compression, the principal concern is that
the user be fully informed about the technique.
Recommendations
1. A complete description of the following is necessary:
* data that are compressed/stored
* what can be retrieved, that is, entire ECGs,
selected cycles only, or measurements only
* how retrieved data may reasonably be used,
for example, rhythm analysis, visual interpretation, serial comparison, criteria/algorithm
development, etc.
2. The fidelity with which data can be reproduced
should be specified, using the criteria in the section on signal fidelity above.
3. The compression ratio should be stated, with
reference to the original sampling details, that is,
sampling rate, bit resolution, and number of bits
per sample. For example, if data from a system
with 500 Hz, 10-bit, equal-interval sampling are
compressed with a technique that uses 1,000 bits/
sec, the compression ratio may be characterized as
5:1, assuming the sampling details are clearly
stated. A preferable alternative is to express the
data reduction as average bits/sample; for
instance, in example above, the average length of
the reduced data is 2 bits/sample.
Low Frequency, Filters, and Phase
The lowest frequency components in the ECG
correspond to the longest RR interval. Thus, assuming a lower bound for heart rate of 30 beats/min, the
low-frequency cutoff of an ECG should be approximately 0.5 Hz. Such a cutoff will suppress a still lower
frequency baseline wander due to respiratory variation, etc. However, an amplifier with this cutoff
frequency distorts ECG signals considerably, particularly T waves and ST segments, since linear systems
with a nonuniform amplitude versus frequency
response exhibit phase nonlinearities in frequency
ranges where amplitude versus frequency response
changes rapidly. In 1975, the AHA7 recommended a
low-frequency cutoff of 0.05 Hz, based on previously
published studies.24 An analog, single-pole amplifying system with this cutoff frequency will exhibit less
than 6° phase shift for frequency components around
0.5 Hz, causing no important distortions in the ST
segment, T wave, or QT interval.
For example, such a system would respond with a
displacement of approximately 60 gV after a 200
gV-sec impulse input (2 mV x 100 msec; see Figure
1). A more realistic model for a QRS might be a
triangular waveform with 3 mV amplitude and 100
msec duration. This 150 ALV-sec impulse would be
followed by an approximate 45 gLV offset if a linear
system with a 0.05 Hz, single-pole filter were used
(Figure 2).
Since QRS fundamental frequencies are usually
less than 30 Hz and T wave fundamental frequencies
are usually greater than 1.0 Hz, the most sensitive
frequency range with respect to phase shifts is 1.0-30
Hz. A system with a phase lag of 60 at 1.0 Hz and zero
phase lag at frequencies above 10.0 Hz produces an
error in OT interval of less than 12 msec. This is well
within the tolerances suggested by the CSE studies.5
However, a frequency cutoff at 0.05 Hz does not
sufficiently suppress baseline wander, and ECGs with
higher cutoffs have introduced unacceptable phase
nonlinearities with consequent distortions in ST
segments.25-32
Figure 3 illustrates the relation of two step functions (positive followed by negative) to an impulse
function. The first and second panels of Figure 3
depict the relation between the two unfiltered step
functions and the corresponding impulse function.
The third and fourth panels of Figure 3 show the
AHA Scientific Council Recommendations in Electrocardiography
733
2000 ~~~
1500 _
_
|
(I)
I_J
0
1000 _
00
_
1500
21000 _
FIGURE 1. Top panel: Solid line shows input signal, a square
impulse with amplitude of 2 mV and duration of 100
(a 200 ,uV-sec impulse). Broken line represents output
after signal has been passed over a 0.05 Hz, single-pole filter.
Bottom panel: Error (equals output minus input) in gV over
time. Note maximum depression of segment after impulse is
about 60 ,uV
21500l
wave
msec
U)
_i
0
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
0G
G:
0
TIME (msec)
2500 _
U')
2000 _
11
1
-Jl
0
1500 _
0 1000 1
0
500)
1
1
so1
FIGURE 2: Top panel: Solid line shows input signal, a
triangular wave impulse with amplitude of 3 mVand duration
of 100 msec (a 150 ,utV-sec impulse). Broken line represents
output after signal has been passed over a 0.05 Hz, single-pole
filter. Bottom panel: Error (equals output minus inp'ut) in ,uV
over time. Note maximum depression of segment after impulse
is about 45 ,tV.
-500.
0s
--
-
-T
-
--
l-
-
-
1-.
--
-
1-
-
--
1
-
--
-T--
-5
-10
un
o
-15
-20
-25
0
°
-
30
2 -35
-40
-45
1,1
-5
-;v
300
500
0
200
400
100
TIME (msec)
_
600
700
800
effect on the same step and impulse functions of a
high-pass, single-pole, 0.5 Hz filter. Note a 500 ,uV
depression of the segment immediately after the
impulse. This model mimics the problem with artifactual displacement of the ST segment resulting
from the phase nonlinearities mentioned above.
In 1975, the AEA made no recommendations
about phase distortion since it was assumed that a
system with extended bandwidth (at least 0.05 Hz)
would not introduce unacceptable phase response
above 0.5 Hz. Requirements for phase response were
also omitted from the ANSI standard.8
Special Report Circulation Vol 81, No 2, February 1990
734
1
2000
1 500
(n
1000
0
500
0
0
a: -500
1
i
1
0
i
-1000
1
-1500
1
-2000
2000
1500
U)
1000
0
500
I-~
0
FIGURE 3: Top panel: Unfiltered step functions; Positive step
function followed 0.100 seconds later by negative step function. Second panel: Unfiltered impulse function composed of
step functions in top panel. Third panel: Effect of a high-pass,
single-pole, 0.5 Hz filter on step functions in top panel. Bottom
panel: Effect of 0.5 Hz filter on impulse function in second
panel. Note 500 ,uV depression of segment immediately following impulse. This model mimics the problem with artifactual displacement of ST segment resulting from phase nonlinearities (see text).
0
a:
-500
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
-1000
-1500
-2000
cc
(0
2000
1500
1000
Co
0
500
0
0
0 -500
-1000
- 1500
-2000
2000
1500
1000
-J
0
500
0
-~~~~~~~~~~~~~~~~
-500
-1000
cD
100
200
300
400
500
600
700
8(00
TIME (msec)
Digital technology now permits a variety of methods for increasing cutoff frequency while maintaining
phase linearity and minimizing ST distortion. A
corrected baseline preceding each QRS may be created with several types of filters. In the presence of
certain abnormalities, such as an interpolated premature ventricular complex, the baseline correction
method may distort the preceding ST-T wave. On the
other hand, many programs use an averaged complex
for measurement; if baseline wander is not mini-
mized before averaging, the final, averaged complex
show artifactual shifts in the ST segment, which
is particularly important in stress ECGs. Thus, caution must be exercised when implementing such
methods to guard against unexpected and unintended distortions of low-frequency signals, namely,
the P and ST-T waves.
At least two alternative methods are improvements
on the traditional 0.05 Hz, single-pole system. In one,
after capture of a significant time segment, the filter
may
AHA Scientific Council Recommendations in Electrocardiography
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
used during acquisition is used again, but in reverse
time. This method is most easily implemented with
digital technology and results in null phase shift for
all frequencies and time-symmetric step and impulse
responses.33
A real-time alternative is to construct a filter with
a step response that is initially flat.34 To the extent
that the flat response is maintained over the duration
of an input impulse such as a QRS, displacement in
the ST segment is minimized. However, two caveats
must be observed: 1) All systems with initially flat
step responses will have a gain at some frequency.
Care must be taken that this gain does not occur
within the range of ECG frequencies and in the
range of common noise frequencies, for example, in
the presence of baseline wander due to respiratory
variation, the gain should be limited to 20%. 2) If the
duration of flat step response extends beyond the
impulse duration with a subsequent rapid decay, then
an inverted image of the impulse can occur after
original input. If a triangular impulse (3 mV by 100
msec) is used as a model for a QRS, then slopes
outside the region of such an impulse should not
exceed 150 ,V/sec.
The remaining issue concerns minimum amplitude
versus frequency response criteria. It is necessary to
consider a lower bound for average heart rate and
the degree of heart rate variability to be expected.
Simonson's35,36 studies and data from the Framingham Heart Study noted by Robert J Garrison, MA,
and Daniel Levy, MD (personal communication,
June 26, 1989), indicate that over 99% of adults have
a mean resting heart rate greater than 44 beats/min.
Simonson's work implies that in 99% of resting
adults, the intraindividual RR interval variation is
less than 0.126 seconds. Hence, a heart rate of 40
forms a lower bound for over 99% of resting adults,
99% of the time. A low- frequency amplitude
response down to 0.67 Hz is necessary to reproduce
heart rate frequencies in such a range.
The following recommendations do not include a
specific reference to phase linearity because a system
meeting the recommended amplitude and impulse
response criteria is appropriate for accurate ECG
reproduction, regardless of phase characteristics.
However, a system that meets the amplitude
response criterion and also has phase linearity at
least equal to that of the linear 0.05 Hz, single-pole
system is likely to meet the impulse response criteria.
Recommendations
1. Amplitude response should be flat to within +6%
(0.5 dB) over a range of 1.0-30 Hz; the 3 dB
points should be less than or equal to 0.67 Hz and
greater than or equal to 150 Hz. (Compare with
the AHA's recommendations, section A8.7)
2. A 1 mV-sec impulse input should not produce a
displacement greater than 0.3 mV after the
impulse.
735
3. For a 1 mV-sec impulse input, the slope of the
response outside the region of the impulse should
nowhere exceed 1 mV/sec.
(Note: Criteria 2 and 3 are met by an analog 0.05 Hz,
single-pole filter.)
High Frequency and Analog-to-Digital Conversion:
Sampling Rate and Quantization
In the following discussion, the conventional definition of bandwidth as the frequency range between
low and high-frequency cutoffs (3 dB down) is used.
The frequency content of ECG signals has been a
subject of debate for many years. For routine recording, most electrocardiographers agree that visual
diagnostic accuracy can be maintained with a highfrequency specification between 50 and 100 Hz. The
AHA recommended the upper figure of 100 Hz for
direct writing electrocardiographs in 19676 and again
in 1975,7 based, in part, on technological limitations
at that time. The ECG committees of the AHA based
their recommendations on studies conducted in the
1960s and 1970s. Although medical and industrial
communities in the United States have since
accepted the need for 100 Hz bandwidth, controversy
still exists worldwide. The international standard37
for ECGs has been "on hold" for several years for a
number of reasons, including the negative vote by the
United States for a 75 Hz bandwidth favored by other
participating countries. In the meantime, the 1983
ANSI standard8 restates the same recommendation
for an upper frequency cutoff of 100 Hz.
However, many electrocardiographs on the current
market already have bandwidths of up to 125 Hz, a
standard proposed in 1976 by an instrumentation
subcommittee of the joint advisory committee of the
Canadian Cardiovascular Society and the Canadian
Heart Foundation.38 Furthermore, for stand-alone
amplifiers, in 1975, the AlHA recommended "a frequency response which is flat to within 6% from 0.14
to 1000 Hz."7
High-frequency components in the ECG have been
studied extensively. Langner and Geselowitz39 demonstrated low-amplitude notches at high frequencies
(greater than 100 Hz), and alterations in such highfrequency notches in the context of coronary artery
disease have been studied by many investigators.40-48
However, the diagnostic significance of lowamplitude, small-duration QRS notches has not been
uniformly accepted.49
Other studies have focused on errors in the principal Q, R, or S waves introduced as the bandwidth
was reduced.50-52 Clearly, a bandwidth of up to at
least 500 Hz is necessary to reproduce all measurable
high-frequency components.53 Below this, amplitude
errors occur more frequently. A judgment must be
made about the level of error that is acceptable.
Previously, 0.025 mV/0.050 mV (or 25 ,uV/50 ,uV) was
considered allowable, based on human visual ability
to resolve this amplitude (0.25 mm) when viewing a
tracing on standard 1 mm ruled paper, at full or half
standard, respectively.
736
Special Report Circulation Vol 81, No 2, February 1990
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
However, automated systems are capable of
exceeding this resolution. Wolf54 reported a program
that resolves waveforms down to 25 uV in amplitude
and 20 msec in duration if noise is less than 60 JLV
root mean square (RMS) error. The European CSE
group13,55 tested nine computer programs. Seven
consistently detected 12 msec Q waves with amplitude greater than or equal to 30 ,uV or 12 msec initial
R waves with amplitude greater than or equal to 40
,uV. Furthermore, based on their library of routine
ECGs, the CSE group recommended defining a QRS
deflection as "present" if it has an amplitude greater
than or equal to 20 ,LV and a duration greater than or
equal to 6 msec (see Fidelity Criterion F5).
Typical ECG analysis systems have 10-12 bit
analog-to-digital conversion, which translates to
9.76-2.44 SuV resolution for a full 10 mV scale
respectively. The equivalent input (preamplifier)
noise is in the range of 5-10 gV RMS. Amplitude
differences of 10-20 ,uV may occur, resulting in
significant changes in diagnostic statements in programs that use decision-tree logic. In such systems
there is good reason for an increased system bandwidth that keeps amplitude errors below 20 gV for
low-amplitude deflections.
Using 100 Hz bandwidth for recording adult ECGs
and 150 Hz for infant ECGs, Berson et a150-52 showed
that amplitude errors greater than 50 gV can occur
in over 10% of the recordings. Clearly, a bandwidth
exceeding 100 Hz for adult ECGs and 150 Hz for
infant ECGs will be required to reduce amplitude
errors well below 50 ,iV and to determine small R or
Q waves with reliability. Exactly how high the bandwidth must extend has yet to be determined.
A significant increase in bandwidth will, of course,
affect the digital sampling rate. Theoretical considerations guide determination of the minimum sampling rate necessary to allow reconstruction of the
analog signal to retain information. Nyquist's theorem asserts that the minimum sampling rate should
be twice the highest frequency component contained
in the original analog signal. However, this theorem
becomes valid only in the limit for an infinite sampling period. There are additional considerations for
the finite epochs of ECG recording.
First, the signal of interest may be mixed with
undesirable high-frequency components, in which
case it is necessary to choose a sampling rate based
on those components, although the signal of interest
may not contain such components. If a lower sampling
rate is chosen, errors will occur in the reconstructed
waveform as these high-frequency components superimpose on and mix with the low-frequency components.
An alternative is to filter out the undesirable highfrequency components before sampling; then a lower
sampling rate may be adequate.
Based on 100 Hz bandwidth, in 1975, the AHA7
recommended a minimum sampling rate of 500 Hz
for equal-interval sampling, recognizing the need for
a sampling rate two to three times the theoretical
minimum. Barr and Spach12 tested different sampling
rates, concluding that 500 Hz was the minimum
sampling rate required for adult ECGs if the reconstructed waveform were to be within 1% mean error
compared to the original waveform. They recommended a rate of 834 Hz for waveforms from infants
to eliminate visible errors, even when the mean error
level was below 1%. Using a triangular shape to
model the R and S waves, Mortara56 estimated that a
250 Hz sampling and 100 Hz frequency response
would result in 25-30% attenuations in R and S
amplitudes. Garson57 compared digital electrocardiographs (100 Hz frequency, 250 Hz sampling) with an
analog electrocardiograph (100 Hz) and found losses
in RS voltage of 15-25% in pediatric cases.
Yamamoto58 examined pediatric ECGs with 800 Hz
frequency and 4,000 Hz sampling. Reducing these
data to 100 Hz frequency and 250 Hz sampling
lowered the number of visible QRS notches (greater
than or equal to 40 ,uV) in 20 of 35 normal children
and 29 of 51 subjects with heart disease; a reduction
of more than 100 gV in peak-to-peak QRS voltage
was found in nearly 30% of the cases. Berson et a159
showed that for a 100 Hz bandwidth, a 250 Hz
sampling rate causes amplitude errors greater than
100 AV in less than 5% of 400 adult ECGs. QRS
averaging reduced incidence of this error to fewer
than 0.5% of records.
The method for interpolation of digital data to
reconstruct the original continuous signal must be
ideal if the minimum sampling rate determined by
Nyquist's theorem is used. To the extent that ideal
interpolation cannot be achieved, sampling rates
higher than the theoretical minimum must be used.
Zywietz'4,15 modeled ECG waves with triangular,
parabolic, and bell-shaped waveforms and showed
that a waveform of 16 msec can be reconstructed with
a maximum error of 10% when the sampling rate is
500 Hz.
If reconstruction between samples is to be precise, the combined interpolation method and sampling rate becomes more important. If, on the other
hand, the sampled points only are of interest,
interpolation methods are not important. Although
there may be a theoretical gain if optimum interpolation techniques are used, the findings of both
McRae60 and Barr and Spach12 indicate that sophisticated interpolation methods are not superior to
straight line interpolation.
A third consideration is the interacting effect of
sampling rate and quantization on error as shown by
Neubert et al.11 They found that a sampling rate of
250 Hz (equal interval) at a 10-bit resolution is not
sufficient to achieve a 10 gV RMS error and suggest
that maximum error may be up to 20 times as large as
RMS error. Zywietz14,15 studied maximum peakrelated error (MPE; see criterion F6), which is more
relevant to ECG diagnostic measurements than RMS
error and 10-100 times larger. This analysis suggests
that sampling intervals should be about 3/25 of
minimum wave duration to achieve an MPE less than
or equal to 10%. For the small initial QRS deflec-
AlIA Scientific Council Recommendations in Electrocardiography
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
tions in the CSE study (12 msec), this translates into
a sampling rate of about 680-700 Hz, assuming that
QRS complexes are not averaged.
Berson et al59 compared 8- and 10-bit resolution
with 12-bit resolution in 115 adult ECGs. When
resolution was reduced from 12 to 10 bits, differences
in waveform durations were 4 msec or more in 3% of
the records. Amplitude errors were greater than or
equal to 50 ,uV in 2%. When resolution was reduced
to 8 bits, the differences in durations were 4 msec or
more in 29% of the records. Based on Neubert's
theoretical study, it would be necessary to digitize at
650-700 Hz at 8 bits resolution to achieve an RMS
error of 10 ,gV or less.
To minimize the effects of noise, some programs
average measurements across many similar beats.
Others construct a PQRST complex by coherently
aligning similar complexes and using an average,
mode, or other parameter of samples from the different complexes. The extent to which noncoherent
noise is suppressed depends on how coherently the
complexes are aligned, the rejection of dissimilar
complexes (e.g., premature ventricular complexes),
and the pitfalls of the various methods. Zywietz's
studies14'5 of the effect of noise and sampling on
error suggest that the noise level in the constructed
complex can be reduced below 5 guV, allowing deflections of 20 glV to be estimated with 10% or less error.
Nevertheless, it is not feasible at this time to devise
criteria in this complex evolving issue.
In 1975, the AHA ECG Committee7 recommended that digitizing be accomplished with a precision of 10 ,uV, based on the 50 gV resolution
achievable with ordinary visual interpretation and
consideration of future automated measurements,
although at that time, the interaction of quantizing
and sampling rates and resolution of small waveforms
achievable by advanced technology could not have
been predicted in detail.
The above discussion assumes equal-interval sampling, which is used by most current automated ECG
programs. However, it is important to recognize that
nonuniform sampling rates and quantization levels
may be used to reduce the average sampling rate
below 500 Hz and the average number of bits below
10/sample. Barr and Pollard61,62 applied a method of
nonuniform sampling to cardiac electrograms; their
fan method of adaptive sampling results in markedly
lower average sampling rates compared with equal
interval sampling. Such rates and quantizing levels
may be quite acceptable. But, as stated in the AHA
recommendations, "when such methods are used, the
performance of such systems shall be adequately
documented."7
Although the majority of currently available automated ECG programs use a 250 Hz sampling rate,
many use 500 Hz, conforming to the AHA's recommendations. While increased storage capacity will
cost more, depending on efficiency of data compression, low cost and availability of microchip processors
make higher sampling rates economically feasible.63
737
Recommendations
1. Analog bandwidth should extend to 125 Hz, the
current de facto industrial standard and a Canadian standard since 1976.38 It should extend to 150
Hz for infant ECGs.
2. For visual reading of the ECG, analog-to-digital
conversion (sampling rate and bit resolution)
should be sufficient to maintain fidelity criterion
Fl (see "Signal Fidelity," page 731).
3. For computer analysis of the ECG, analogto-digital conversion parameters should be adequate to maintain low RMS error criteria (such as
F2 or F4), plus certain waveform resolution criteria (such as F3, F5, or F6). In general, a minimal
sampling rate of 500 Hz (equal-interval) with
minimal resolution of 10 gV) is implied. However,
this does not preclude other ways of meeting
fidelity criteria.
Calibration For Digital ECG Systems
Proper calibration of an ECG system as described
in section A13 of the AHA report7 is still a necessity.
Introduction of a known signal is the only secure
means of thoroughly testing calibration. Since many
digital systems cannot easily cope with fast-rise time
signals, a triangularly shaped calibration pulse, typically 1 mV at the patient electrode input point, is also
acceptable. The calibration pulse should be traceable
through the entire system, including analog-to-digital
converter, data compression, telecommunications
link, storage, and digital-to-analog converter. Since
the interval between such tests may be prolonged, the
stability of calibration is an important issue. Therefore, ECG recording system specifications should
include a maximum drift of the absolute system gain
over time and environmental changes.
Conclusion
Again, it is important to emphasize the difference
between digital processing in automated ECG analysis and visual interpretation by humans. As frequently observed, human visual criteria are typically
not appropriate for implementation in a computer
program that resolves the ECG in microvolts and
milliseconds.464 This report treats issues far beyond
those of the mid-1970s, which mainly addressed
problems of visual analyses and interpretation.
As mentioned previously, an explosion of digital
technology has overtaken the field of ECG analysis in
the 1980s; technical developments have outpaced
most efforts to evaluate the quality of the product
and its impact on medical care. Each innovation in
digital methodology should be extensively tested
before implementation in a commercial system so
that developers of such innovations can specify the
recommendations and fidelity criteria that their
devices meet.
It would be helpful to those who evaluate ECG
systems if developers would disclose the underlying
logic of their computer-based ECG interpretation
738
Special Report Circulation Vol 81, No 2, February 1990
programs and whether it includes diagnostic criteria
recognized by experienced ECG readers or a more
complex algorithm such as those based on multivariate statistics. If the logic incorporates demographic or clinical parameters such as age, habitus,
gender, symptoms, medications, clinical status, coronary risk factors, etc., the role or weighting of each
parameter in the diagnostic procedure should be
explicitly identified. Finally, it is recommended that
developers test the performance of their programs
against standardized databases such as the CSE
multilead ECG database, just as monitoring devices
are now tested against the AHA Arrhythmia and
MIT/BIH databases.65,66
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
Acknowledgments
We gratefully acknowledge the helpful comments
of Professors Roger C. Barr, PhD, Mitsuhara Okajima, MD, John A. Milliken, MD, Hubert V. Pipberger, MD, Ronald H. Startt/Selvester, MD, and Jos
L. Willems, MD, who reviewed this manuscript.
References
1. Drazen E, Mann N, Borun R, Laks M, Berson A: Survey of
computer-assisted electrocardiography in the United States. J
Electrocardiol 1988;21(suppl):S98-S104
2. Willems JL, Arnaud P, van Bemmel JH, Bourdillon PJ, Brohet
C, dalla Volta S, Andersen JD, Degani R, Denis B, Demeester
M, et al: Assessment of the performance of electrocardiographic computer programs with the use of a reference data
base. Circulation 1985;71:523-534
3. Willems JL: Standards for quantitative electrocardiography:
Facts and fiction, in Wagner GS, Scherlag BJ, Bailey JJ (eds):
Computerized Interpretation of the Electrocardiogram. New
York, Engineering Foundation, 1986, pp 135-144
4. Willems JL: The CSE project: Goals and achievements, in
Willems JL, van Bemmel JH, Zywietz C (eds): Computer ECG
Analysis: Towards Standardization: Proceedings of the IML4
Working Conference, Leuven, Belgium, 2-5 June, 1985. Amsterdam, North-Holland, Elsevier Science Publishing Co, Inc,
1986, pp 11-21
5. The CSE Working Party: Recommendations for measurement
standards in quantitative electrocardiography. Eur Heart J
1985;6:815-825
6. Recommendations for standardization of leads and of specifications for instruments in electrocardiography and vectorgraphy. Report of the Committee on Electrocardiography,
American Heart Association: Kossmann CE, Brody DA,
Burch GE, Hecht HH, Johnston FD, Kay C, Lepeschkin E,
Pipberger HV. Circulation 1967;35:583-602
7. Recommendations for standardization of leads and of specifications for instruments in electrocardiography and vectorcardiography. Report of the Committee on Electrocardiography, American Heart Association: Pipberger HV,
Arzbaecher RC, Berson AS, Briller SA, Brody DA, Flowers
NC, Geselowitz DB, Lepeschkin E, Oliver GC, Schmitt OH,
Spach M. Circulation 1975;52:11-31
8. Association for the Advancement of Medical Instrumentation:
American National Standard for Diagnostic Electrocardiographic
Devices. Arlington, Va, 1983
9. Alraum W, Zywietz C, Borovsky D, Willems JL: Methods for
noise testing of ECG analysis programs, in Computers in
Cardiology. New York, Institute of Electrical and Electronics
Engineers, 1983, pp 253-256
10. Zywietz C: Noise tolerance in ECG computer programs and
measurement stability of minimum waves: A summary of
results from CSE noise tests and conclusions, in Willems JL,
van Bemmel JH, Zywietz C (eds): Computer ECG Analysis:
Towards Standardization: Proceedings of the IMIA Working
Conference, Leuven, Belgium, 2-5 June, 1985. Amsterdam,
North-Holland, Elsevier Science Publishing Co, Inc, 1986, pp
87-96
11. Neubert D, Cramer E, Teppner U: Stability of measurement
parameters, in Willems JL, van Bemmel JH, Zywietz C (eds):
Computer ECG Analysis: Towards Standardization: Proceedings
of the IML4 Working Conference, Leuven, Belgium, 2-5 June,
1985. Amsterdam, North-Holland, Elsevier Science Publishing
Co, Inc, 1986, pp 129-133
12. Barr RC, Spach MS: Sampling rates required for digital
recording of intracellular and extracellular cardiac potentials
(review). Circulation 1977;55:40-48
13. Willems JL: Common Standards for Quantitative Electrocardiography. (Concerted action project 11.2.2 of the sectoral
research programme in the field of medical and public health
research, CSE 3rd progress report, 1983). Leuven, Belgium,
ACCO Publications, 1983, pp 201-223
14. Zywietz C, Spitzenberger U, Palm C, Wetjen A: A new
approach to determine the sampling rate for ECGs, in Computers in Cardiology. New York, Institute of Electrical and
Electronics Engineers, 1983, pp 261-264
15. Zywietz C: Sampling rate of ECGs in relation to measurement
accuracy, in Wagner GS, Scherlag BJ, Bailey JJ (eds): Computerized Interpretation of the Electrocardiogram. New York,
Engineering Foundation, 1986, pp 122-125.
16. Macfarlane PW, Cawood HT, Lawrie TD: A basis for computer interpretation of serial electrocardiograms. Comput
Biomed Res 1975;8:189-200
17. Ruttiman UE, Pipberger HV: Data compression and the
quality of the reconstructed ECG, in Wolf HK, Macfarlane
PW (eds): Optimization of Computer ECG Processing. Amsterdam, North-Holland, 1980, pp 77-83
18. Huffman DA: Method for construction of minimum redundancy codes. Proc Institute of Radio Engineers 1952;
40:1099-1101
19. Peden JMC: ECG data compression: Some practical considerations, in Paul JP et al (eds): Computing in Medicine.
London, McMillan Publications, Inc, 1982, pp 62-67
20. Womble ME, Zied AM: A statistical approach to ECG/VCG
data compression, in Wolf HK, Macfarlane PW (eds): Optimization of Computer ECG Processing. Amsterdam, NorthHolland, 1980, pp 91-99
21. Lux RL, Evans AK, Burgess MJ, Wyatt RF, Abildskov JA:
Redundancy reduction for improved display and analysis of
body surface potential maps. I. Spatial compression. Circ Res
1981;49:186-196
22. Evans AK, Lux RL, Burgess MJ, Wyatt RF, Abildskov JA:
Redundancy reduction for improved display and analysis of
body surface potential maps. II. Temporal compression. Circ
Res 1981;49:197-203
23. Blanchard SM, Barr RC: Comparison of methods for adaptive
sampling of cardiac electrograms and electrocardiograms. Med
Biol Eng Comput 1985;23: 377-386
24. Berson AS, Pipberger HV: The low-frequency response of
electrocardiographs: A frequent source of recording errors.
Am Heart J 1966;71:779-789
25. Berson AS, Geselowitz DB, Pipberger HV: Low-frequency
requirements for electrocardiographic recording of ST segments (letter). Am J Cardiol 1987;60:939-940
26. Krucoff MK: Critical challenges to computer-assisted ST
segment analysis, in Wagner GS, Scherlag BJ, Bailey JJ (eds):
Computerized Interpretation of the Electrocardiogram. New
York, Engineering Foundation, 1986, pp 85-88
27. Gold RG: Do we need a new standard for electrocardiographs? (editorial) Br Heart J 1985;54:119-120
28. Tayler DI, Vincent R: Artefactual ST segment abnormalities
due to electrocardiograph design. Br Heart J 1985;54:121-128
29. Nygards M-E, Ahren T, Ringquist I, Walker A: Phase correction for accurate ST segment reproduction in ambulatory
ECG recording, in Computers in Cardiology. New York, Institute of Electrical and Electronics Engineers, 1984, pp 33-38
30. Tayler DI, Vincent R: Signal distortion in the electrocardiogram due to inadequate phase response. IEEE Trans Biomed
Eng 1983;30:352-356
AHA Scientific Council Recommendations in Electrocardiography
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
31. Bragg-Remschel DA, Anderson CM, Winkle RA: Frequency
response characteristics of ambulatory ECG monitoring systems and their implications for ST segment analysis. Am Heart
J 1982;103:20-31
32. Longini RL, Giolma JP, Wall C III, Quick RF: Filtering
without phase shift. IEEE Trans Biomed Eng 1975;22:432-433
33. Pottala EW, Bailey JJ, Horton MR, Gradwohl JR: Suppression of baseline wander in the ECG using a bilinearly transformed, null phase filter. J Electrocardiol 1989;22:244-248
34. Mortara DW: Digital filters for ECG signals, in Computers in
Cardiology. New York, Institute of Electrical and Electronics
Engineers, 1977, pp 511-514
35. Simonson E: Differentiation Between Normal and Abnornal in
Electrocardiography. St. Louis, CV Mosby Co, 1961, p 158
36. Simonson E, Brozek J, Keys A: Variability of the electrocardiogram in normal young men. Am Heart J 1949;38:407-422
37. International Electrotechnical Commission: Draft Standard
Specification for the Performance of Single and Multi-Channel
Electrocardiographs. IEC document 62D. Geneva, Switzerland,
International Electrotechnical Commission 1978, p 6
38. Wolf HK, Graystone P, Lywood DW, et al: Suggested minimum performance characteristics of data acquisition instrumentation in computer-assisted ECG processing systems. J
Electrocardiol 1976;9:239-247
39. Langner PH Jr, Geselowitz DB: Characteristics of the frequency spectrum in the normal electrocardiogram and in
subjects following myocardial infarction. Circ Res 1960;
8:577-584
40. Langner PH Jr, Geselowitz DB, Mansure FT: High-frequency
components in the electrocardiograms of normal subjects and
of patients with coronary heart disease. Am Heart J 1961;
62:746-755
41. Boyle D, Carson P, Hamer J: High frequency electrocardiography in ischemic heart disease. Br Heart J 1966;28:539-545
42. Reynolds EW Jr, Muller BF, Anderson GJ, Muller BT:
High-frequency components in the electrocardiogram: A comparative study of normals and patients with myocardial disease. Circulation 1967;35:195-206
43. Flowers NC, Horan LG, Tolleson WJ, Thomas JR: Localization of the site of myocardial scarring in man by highfrequency components. Circulation 1969;40:927-934
44. Flowers NC, Horan LG: Diagnostic import of QRS notching
in high-frequency electrocardiograms of living subjects with
heart disease. Circulation 1971;44:605-611
45. Langner PH Jr, Geselowitz DB, Briller SA: Wide band
recording of the electrocardiogram and coronary heart disease. Am Heart J 1973;86:308-317
46. Anderson GJ, Blieden MF: The high frequency electrocardiogram in coronary artery disease. Am Heart J 1975;89:349-358
47. Sapoznikov D, Tzivoni D, Weinman J, Penchas S, Gotsman
MS: High fidelity ECG in the diagnosis of occult coronary
artery disease: A study of patients with normal conventional
ECG. J Electrocardiol 1977;10:137-148
48. Goldberger AL, Bhargava V, Froelicher V, Covell J: Effect of
myocardial infarction on high-frequency QRS potentials. Circulation 1981;64:34-42
739
49. Pipberger HV, Zafar SA, Pipberger HA: Diagnostic information derived from high frequency components in the electrocardiogram, in Antaloczy Z (ed): Modem Electrocardiology.
Amsterdam, The Netherlands, 1978, pp 165-180
50. Berson AS, Pipberger HV: Electrocardiographic distortions
caused by inadequate high-frequency response of directwriting electrocardiographs. Am Heart J 1967;74:208-218
51. Berson AS, Ferguson TA, Batchlor CD, Dunn RA, Pipberger
HV: Filtering and sampling for electrocardiographic data
processing. Comput Biomed Res 1977;10:605-610
52. Berson AS, Lau FY, Wojick JM, Pipberger HV: Distortions in
infant electrocardiograms caused by inadequate highfrequency response. Am Heart J 1977;93:730-734
53. Golden DP Jr, Wolthius RA, Hoffler GW: A spectral analysis
of the normal resting electrocardiogram. IEEE Trans Biomed
Eng 1973;13:366-372
54. Wolf HK, Maclnnis PJ, Stock S, Helppi RK, Rautaharju PM:
Computer analysis of rest and exercise electrocardiograms.
Comput Biomed Res 1972;5:329-346
55. Macfarlane PW, Willems JF: The CSE project: Progress as
viewed by the cooperating centres, in Selvester RH, Geselowitz DB (eds): Computerized Interpretation of the Electrocardiogram (VIII). New York, Engineering Foundation, 1983, pp
293-306
56. Mortara DW: Differences between filters with similar cut-off
frequencies, in Wagner GS, Scherlag BJ, Bailey JJ (eds):
Computerized Interpretation of the Electrocardiogram. New
York, Engineering Foundation, 1986, pp 130-133
57. Garson A Jr: Clinically significant differences between the
"old" analog and the "new" digital electrocardiograms. Am
Heart J 1987;114:194-197
58. Yamamoto H, Miyahara H, Domae A: Is a higher sampling
rate desirable in the computer processing of the pediatric
electrocardiogram? J Electro Cardiol 1987;10:321-328
59. Berson AS, Wojick JM, Pipberger HV: Precision requirements
for electrocardiographic measurements computed automatically. IEEE Trans Biomed Eng 1977;24:382-385
60. McRae DD: Interpolation errors, in Advanced Telemetry Study,
Technical Report #1, Parts 1 and 2. Melbourne, Fla, Radiation,
Inc, 1961
61. Barr RC: Adaptive sampling of cardiac waveforms (review). J
Electrocardiol 1988;21:S57-S60
62. Pollard AE, Barr RC: Adaptive sampling of intracellular and
extracellular cardiac potentials with the fan method. Med Biol
Eng Comput 1987;25:261-268
63. Berson AS: Bandwidth, sampling, and quantizing for ECG, in
Wagner GS, Scherlag BJ, Bailey JJ (eds): Computerized Interpretation of the Electrocardiogram. New York, Engineering
Foundation, 1986, pp 118-121
64. American College of Cardiology: Tenth Bethesda Conference
on Optimal Electrocardiology. Am J Cardiol 1978;41:111-184
65. Greenwald SD, Patil RS, Mark RG: Improved arrhythmia
detection in noisy ECGs through the use of expert systems, in
Computers in Cardiology. Washington, DC, IEEE Computer
Society, 1989, pp 163-165
66. Moody GB, Mark RG, Goldberger AL: Evaluation of the
"TRIM" ECG data compressor, in Computers in Cardiology.
Washington, DC, IEEE Computer Society, 1989, pp 167-170
Downloaded from http://circ.ahajournals.org/ by guest on June 15, 2017
Recommendations for standardization and specifications in automated
electrocardiography: bandwidth and digital signal processing. A report for health
professionals by an ad hoc writing group of the Committee on Electrocardiography and
Cardiac Electrophysiology of the Council on Clinical Cardiology, American Heart
Association.
J J Bailey, A S Berson, A Garson, Jr, L G Horan, P W Macfarlane, D W Mortara and C
Zywietz
Circulation. 1990;81:730-739
doi: 10.1161/01.CIR.81.2.730
Circulation is published by the American Heart Association, 7272 Greenville Avenue, Dallas, TX 75231
Copyright © 1990 American Heart Association, Inc. All rights reserved.
Print ISSN: 0009-7322. Online ISSN: 1524-4539
The online version of this article, along with updated information and services, is located on
the World Wide Web at:
http://circ.ahajournals.org/content/81/2/730.citation
Permissions: Requests for permissions to reproduce figures, tables, or portions of articles originally
published in Circulation 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 is online at:
http://circ.ahajournals.org//subscriptions/