ORIGINAL PAPER
Empirical Determination of an
ECG Compression Ratio for Mobile
Telecardiology Applications
Sun Kook Yoo, Ph.D.,2,4,5,6,7 Kwanghyun Lee,1 and Moon H. Lee, M.D.3
1
Yonsei University College of Medicine, Seoul, South Korea.
2
Medical Engineering, 3Internal Medicine, Yonsei University,
College of Medicine, Seoul, South Korea.
4
Brain Korea 21 Projects for Medical Science, 5Center for Emergency
Medical Informatics, 6Human Identification Research Center, 7Center
for Signal Processing Research, Yonsei University, Seoul, South Korea.
Abstract
It is important to consider electrocardiogram (ECG) data compression
that does not sacrifice diagnostic quality significantly before apply-
ing ECG data compression to mobile telecardiology applications. In
this paper, we assessed the reconstructed ECG quality after compres-
sion with a wavelet-based low-delay algorithm, using both subjective
and objective indices. We included diverse ECG databases including
both normal and abnormal ECG data, and evaluated the relationship
between the subjective and objective indices, paying close attention to
specific cases in which there was a large discrepancy between the objec-
tive and subjective quality. Based on our observations, an empirically
determined compression ratio can be applied to compress continuous
ECG signals in limited-bandwidth mobile telecardiology applications.
Key words: electrocardiogram, data compression, telecardiology,
compression ratio
Introduction
T
he electrocardiogram (ECG) is essential in diagnosing patients
with heart problems. In remote or underserved areas without
access to cardiologists, an ECG can be monitored through
a telecardiology system, which transmits the ECG signal
from the patient to a remote site through a network. In addition,
telecardiology may also be used to achieve continuous, instanta-
neous transmission of an ECG signal in real-time during movement,
156 TELEMEDICINE and e-HEALTH MA R CH 2008
for example, allowing the monitoring of emergency patients during
ambulance dispatch.1 Despite advances in network technology, it is
not always possible to secure sufficient network bandwidth for real-
time continuous ECG transmission,2 and bandwidth is even more
limited in a wireless network.
In order to work within limited bandwidth, it becomes necessary
to compress the ECG signal. Many ECG compression techniques have
been studied to maximize compression performance and comply
with bandwidth limitation,3,4 but few studies have focused on the
permissible amount of ECG compression associated with mobile
communication. Compression performance is generally measured
using an objective index, which calculates the squared difference
between the original and the reconstructed ECG signals.5 However,
that objective index using rate-distortion criteria alone cannot reveal
the diagnostic relevance of the compression ratio (CR). Alternatively,
cardiologists can evaluate compression performance subjectively by
visually ranking the reconstructed ECG quality. That visual ranking
can be used as a subjective index describing the clinical adequacy of
the reconstructed ECG quality from the cardiologist’s point of view.6
However, visual inspection is constrained by the limited availability
of cardiologists.
The use of both objective and subjective indices can overcome
the limitations inherent in each index alone. Thus, a permissible CR
that does not sacrifice diagnostic performance significantly can be
determined before applying the ECG compression technique to a tele-
cardiology application. In this paper, ECG compression performance
is analyzed in terms of CR by using the wavelet-based low-delay ECG
compression algorithm,2 which was chosen as a typical technique
suitable for continuous ECG transmission over a mobile network.
Measurements were obtained using various ECG databases includ-
ing both normal and abnormal ECG data. In addition, extraordinary
cases, in which the objective index does not match the subjective
index, were comprehensively inspected to determine the permissible
CR before transmitting the compressed ECG over a mobile network.
DOI : 1 0 .1 0 8 9 /tm j . 2 0 0 7 . 0 0 3 0
 ECG COMPRESSION RATIO FOR MOBILE TELECARDIOLOGY

Table 1. Changing Ratio of Objective Performance (CROP) for Testing Data

TYP VTA ARR MVE NOR SVA MOR SCD
Ratio 0.1041 0.635 0.3188 0.4159 1.7973 1.5992 0.4175 0.5197
TYP, typical testing data set; VTA, ventricular tachyarrhythmia; ARR, arrhythmia; MVE, malignant ventricular ectopy; NOR, normal testing data set; SVA, supraventricular
arrhythmia; MOR, morphological testing data set; SCD, sudden cardiac death.

Materials and Methods ELECTROENCEPHALOGRAM COMPRESSION ALGORITHM

TESTING DATA
Testing data were obtained from various ECG databases: the
Association for the Advancement of Medical Instrumentation (AAMI)
database, the Massachusetts Institute of Technology and Beth Israel
Hospital (MIT-BIH) database, the Creighton University Ventricular
Tachyarrhythmia (CU) database, the Holter database, and the QT
database. Testing data were functionally classified into 4 data sets:
a typical testing data set, a morphological testing data set, a normal
testing data set, and an abnormal testing data set. The typical testing
data set (TYP) came from the AAMI database, because it is used often
in ECG compression research.7,8 The morphological testing data set
(MOR) came from the QT database, which particularly emphasized P-
and T-waves. The normal testing data set (NOR) came from the MIT-
BIH database. The abnormal testing data set came from the MIT-BIH
database, which supplied Arrhythmia (ARR), Malignant Ventricular
Ectopy (MVE), and Supraventricular Arrhythmia (SVA) testing data;
from the CU database, which supplied Ventricular Tachyarrhythmia
(VTA) testing data; and from the Holter database, which supplied
Sudden Cardiac Death (SCD) testing data.7–10 The digitization resolu-
tion for the testing data sets was 12 bits over a range of 10 V.
Wavelet-transform ECG compression algorithms provide good
compression performance because of their straightforward imple-
mentation and efficiency, and their good time and frequency
localization capacity.2 In addition to compression performance, the
amount of delay imparted by wavelet transformation is also an
important factor to be considered, particularly in mobile telecar-
diology, in which the continuous transmission of ECG in real-time
is emphasized for time-critical patient care. To meet these require-
ments, the wavelet-based, low-delay ECG compression algorithm
was chosen from many ECG compression algorithms to analyze ECG
compression performance.
PERFORMANCE MEASUREMENT
Reconstructed ECG quality for different CR levels was evalu-
ated using both objective and subjective indices. The objective
index measured the amount of distortion using a quantitative error
measurement calculation, while the subjective index measured the
diagnostic quality (clinical acceptability) qualitatively by visual
inspection. The objective index was assessed using the root mean
squared error (RMSE), which calculates the difference between the

(a) Uncompressed TYP (b) Compressed TYP (c) Uncompressed SVA (d) Compressed SVA

FIG. 1. Compressed electrocardiogram (ECG) signals with lowest changing ratio of objective performance (CROP) typical testing data set (TYP)
and highest CROP supraventricular arrhythmia (SVA) are compared with uncompressed ECG signals.

© M ARY ANN LI E BE RT, I NC. • VOL. 14 NO. 2 • M ARCH 2008 TELEMEDICINE and e-HEALTH 157
YOO ET AL.

CR vs. RMSE
60
50
40
RMSE

30
20
10
0
0 5 10 15 20 25 30
CR
FIG. 2. The relationship between root mean squared error (RMSE) and compression ratio (CR) for all testing data.

original ECG and the reconstructed ECG. For N ECG samples, the reconstructed ECG, respectively. The subjective index used the Mean
RMSE is defined as Opinion Score (MOS) to quantify cardiologist assessment of reconstructed
ECG quality.6 The MOS provides a numerical interpretation of the per-
RMSE =  N1 Σf(i) ⫺ fˆ(i)
Ν
2
ceived quality of the reconstructed ECG after compression. In the MOS
measurement, 3 cardiologists were blinded and asked to score the clinical
where f(i) and fˆ(i) are the ith samples of the original ECG and the acceptability of a reconstructed ECG signal in comparison to the original

CR vs. MOS
TYP
6
VTA
5
ARR
4
MOS

MVE
3
NOR
2
SVA
1
MOR
0
SCD
0 5 10 15 20 25 30
CR
FIG. 3. The relationship between mean opinion score (MOS) and compression ratio (CR) for all testing data.

158 TELEMEDICINE and e-HEALTH MA R CH 2008

 ECG COMPRESSION RATIO FOR MOBILE TELECARDIOLOGY

ing on the type of testing data. The impact of an increase in CR on

Table 2. Root Mean Squared Error Performance for Testing
Data with Different Levels of Compression the amount of distortion is greater with higher values of CROP. In
other words, testing data types with high CROP are more sensitive
4:1 8:1 10:1 14:1
to changes in CR than those with low CROP. Figure 1 demonstrates
TYP 1.379091 1.887604 2.186796 2.651603
this relationship, with the SVA data set (high CROP) and the TYP
VTA 2.001757 2.75228 3.708174 6.188316 set (low CROP) shown with a CR of 28. The distortion in the com-
ARR 2.017741 2.529496 3.04267 4.102147 pressed SVA data set is considerably more recognizable than in
MVE 2.29047 3.790658 4.649682 6.186647 the compressed TYP set. This reveals that the same CR cannot be
NOR 3.335228 5.967294 8.930523 15.42904 applied to all types of ECG signals irrespective of signal type. In
general, ECG signals with low CROP can be compressed more than
SVA 3.72066 7.161333 9.717564 15.09806
those with high CROP.
MOR 1.598485 1.992633 2.513523 3.944556
Figure 2 shows the relationship between the RMSE and the CR for all
SCD 1.770871 2.287198 3.009873 4.950515 testing data. The CROP is 0.6319 for all types of testing data. The RMSE
TYP, typical testing data set; VTA, ventricular tachyarrhythmia; ARR, arrhythmia; MVE, (distortion) increases as the CR increases as a whole, but shows a lot of
malignant ventricular ectopy; NOR, normal testing data set; SVA, supraventricular variation. The variation becomes greater as the CR increases.
arrhythmia; MOR, morphological testing data set; SCD, sudden cardiac death. The values of RMSE for each CR vary considerably, depending on
the type of testing data, as shown in Table 2. Therefore, the CR for
ECG signal. The scores of 1, 2, 3, 4, and 5 correspond to “Bad (Completely a given distortion boundary cannot be determined independently of
Different),” “Poor (Different),” “Fair (Slightly Different),” “Good (Nearly the type of ECG signal. Instead, the worst condition, taken from the
Unnoticeable),” and “Excellent (Unnoticeable),” respectively. highest RMSE signal, should be selected as the threshold for satisfy-
ing objective criteria. When a RMSE less than 10 is chosen as the
Results permissible objective threshold without disturbing the original ECG
OBJECTIVE MEASUREMENT signal, then a CR of 10:1 is found to be the permissible objective
As shown in Table 1, the changing ratio of objective performance threshold, using the SVA data set, whose CROP at a CR of 10:1 is the
(CROP), defined as RMSE/CR, varies from 0.1041 to 1.7973 depend- worst among all testing data.

RMSE vs. MOS

6
5
4
MOS

3
2
1
0
0 10 20 30 40 50 60
RMSE

FIG. 4. The relationship between root mean squared error (RMSE) and mean opinion score (MOS).

© M ARY ANN LI E BE RT, I NC. • VOL. 14 NO. 2 • M ARCH 2008 TELEMEDICINE and e-HEALTH 159
YOO ET AL.
SUBJECTIVE MEASUREMENT
As shown in Figure 3, for all testing data except the TYP set, the
MOS decreases as the CR increases. Generally, an increase in CR
causes deterioration in the visual quality of the reconstructed ECG.
However, for the TYP data set, an increase in CR less than 30 has little
effect on the reconstructed visual quality, because the TYP data set
represents typical testing ECG with little variation. The TYP data set is
relatively insensitive to changes in CR. When an MOS of 4 (a nearly
unnoticeable difference in ECG quality) is chosen as the subjective
threshold, the permissible subjective CR is found to be 13:1, which is
the lowest among CRs for all testing data. The CRs corresponding to
an MOS of 4 are 13:1, 13:1, 19:1, 23:1, 26:1, 281:1, and 29:1, for the
NOR, SVA, MOR, SCD, MVE, VTA, and ARR data sets, respectively.
RELATIONSHIP BETWEEN SUBJECTIVE AND OBJECTIVE
MEASUREMENTS
As shown in Figure 4, the MOS (subjective index) correlates with
the RMSE (objective index) with the relationship MOS = ⫺0.0999 *
RMSE ⫹5.168. However, there is large variation (0.525), even though
the t-test (reliability of 95%) reveals a correlation (r = 0.725) between
the RMSE and MOS. This large variation between the subjective and
objective indices indicates that there is a discrepancy between subjec-
(a) Original ECG
(b) Reconstructed ECG
FIG. 5. Arrhythmia 104 for compression ratio of 26:1 (a) original and
(b) reconstructed electrocardiogram (ECGs).
160 TELEMEDICINE and e-HEALTH MA R CH 2008
tive and objective qualities. Therefore, the permissible CR cannot be
determined from either the MOS or RMSE criteria alone. Instead, the
lower CR (10:1), between the subjective permissible CR (13:1) and the
objective permissible CR (10:1), should be selected as the permissible
CR for mobile telecardiology.
CASES WITH GOOD OBJECTIVE QUALITY BUT BAD
SUBJECTIVE QUALITY
From a clinical point of view, the maintenance of subjective
quality (diagnostic quality) is more stringent than that of objective
quality.5 Some ECG data show good objective quality (low RMSE),
but bad subjective quality (low MOS). Even though the mean differ-
ence between the original ECG and the reconstructed ECG is small,
clinically important data along the time-sequenced ECG trace can be
significantly distorted, which can result in a low MOS despite the low
RMSE. In order to clarify the relationship between the RMSE and the
MOS, the cases with an RMSE less than 10 and an MOS of 2 were
examined.
Case 1 (Fig. 5): ARR 104: CR of 26:1, RMSE of 8.84, and MOS
of 2. The reconstructed signal shows a shorter PR interval and some
P waves that were absent on the original. Thus, patients with heart
problems may be misdiagnosed as normal.
(a) Original ECG
(b) Reconstructed ECG
FIG. 6. Malignant ventricular ectopy 602 for compression ratio of 26:1
(a) original and (b) reconstructed electroencephalograms (ECGs).
 ECG COMPRESSION RATIO FOR MOBILE TELECARDIOLOGY
Case 2 (Fig. 6): MVE 602: CR of 26:1, RMSE of 9.85, and MOS
of 2. Although there are peaking P waves on the original, the recon-
structed P waves appear normal. A peaking P wave is one of the
diagnostic indicators of an abnormal right heart. Therefore, because
of the altered P-wave appearance, patients with a right cardiac prob-
lem may be missed.
Case 3 (Fig. 7): NOR for CR of 28:1, RMSE of 7.22, and MOS of 2.
PR intervals are longer than in the original. Additionally, the T waves
appear variable in the reconstructed signal, although all the T waves
in the original share the same features.
Case 4 (Fig. 8): MOR 114 for CR of 26:1, RMSE of 9.77, and MOS
of 2. There are some shortened PR intervals on the reconstructed
signal. The energy of the P wave is smaller and the intervals are lon-
ger than in the original. Also, there is an initial blurring in the QRS
complex. These changes may lead to the misdiagnosis of a healthy
person as abnormal.
Case 5 (Fig. 9): MOR 116 for CR of 19:1, RMSE of 6.04, and MOS
of 2. The P wave on the original has disappeared and an additional P
wave has been generated in the PR interval. These factors may lead
to incorrect diagnosis in patients with heart problems.
Case 6 (Fig. 10): MOR 232 for CR of 19:1, RMSE of 8.88, and
MOS of 2. A newly created abnormal P wave on the reconstructed
(a) Original ECG
(b) Reconstructed ECG
FIG. 7. Normal testing dataset for compression ratio of 28:1 (a) origi-
nal and (b) reconstructed electroencephalograms (ECGs).
© M ARY ANN LI E BE RT, I NC. • VOL. 14 NO. 2 • M ARCH 2008 TELEMEDICINE and e-HEALTH 161
signal would still indicate a conduction disorder, but the cause of the
problem may be misdiagnosed. Altered P-wave appearance is also
found.
Case 7 (Fig. 11): MOR 15814 for CR of 26:1, RMSE of 10.00, and
MOS of 2. The energy of the P wave is lower and the appearance has
also widened on the reconstructed image. The reconstructed PR inter-
val is shorter than the original. These could lead to the misdiagnosis
of a healthy individual.
Discussion
In order to cope with bandwidth limitations, compression of the
ECG signal is inevitable. Because lossy compression can achieve
a higher CR than can lossless compression, the lossy compression
is more suitable for low-bandwidth telecardiology despite the
distortion it produces.11 In general, the greater the CR applied to
the original ECG, the lower the quality of the reconstructed image.
Because of this tradeoff, the CR should be carefully selected when
allocating the ECG transmission bitrates for a given network band-
width. When allocating bits, the most commonly used measure is
the RMSE (objective measure). However, this measure does not
reflect the diagnostic value of the compressed ECG. The diagnostic
quality can be validated precisely by expert cardiologists (subjec-
(a) Original
(b) Reconstructed ECG
FIG. 8. Morphological testing data set 114 for compression ratio of
26:1 (a) original and (b) reconstructed electroencephalograms (ECGs).
YOO ET AL.
tive measure). However, cardiologists are not always available.
As discussed in the Results section, the RMSE mostly correlates
with diagnostic quality. However, the subjective measure cannot
be consistently substituted by the objective measure. Each portion
of the ECG waveform, such as the P-wave, QRS-complex, ST seg-
ment, T-wave, and QT-interval, has its own diagnostic meaning.
RMSE, as a mean performance measure, considers the error in all
portions of the ECG waveform equally, and does not especially
reflect the quality of those portions that are most diagnostically
salient. Nonstationary ECG waveforms, missing beats, and noise-
to-signal problems can also affect the reconstructed ECG quality,
which can lead to a large variation in the correlation between
objective and subjective measures.
The permissible CR should be determined as the lowest value
that satisfies both the objective and subjective criteria. The great-
est obstacle in this determination is the variability of ECG data,
both in the case of the individual patient over a period of time
and with respect to different patients. Accordingly, both criteria
cannot be adequately evaluated using only specific patient data
with a limited observation interval. Instead, in our study, the CR
was empirically determined using diverse ECG databases, includ-
ing both normal and abnormal ECG data. Importantly, the ECG
(a) Original ECG
(b) Reconstructed ECG
FIG. 9. Morphological testing data set 116 for compression ratio of
19:1 (a) original and (b) reconstructed electroencephalograms (ECGs).
162 TELEMEDICINE and e-HEALTH MARCH 2008
databases used should contain diverse P waves and ST segments,
since P wave recognition (generally the smallest wave in the ECG
signal) is a difficult task in the diagnosis of ECG abnormalities,
and the ST segment is an important feature identifying myocardial
infarction.12,13
The subjective and objective quality of the reconstructed ECG
can be affected by the compression technique employed.3,4 In this
study, a wavelet-transform low-delay ECG compression algorithm
was selected for a telecardiology application over a mobile network
with limited bandwidth, which required a continuous ECG transmis-
sion with little delay. These same procedures should be applicable to
different ECG compression algorithms. Nevertheless, further research
is required, including more patterns of ECG data and more cardiolo-
gist involvement to reduce bias and to generalize quality evaluation
procedures.
In this study, we used both subjective and objective indices
to determine the permissible CR for telecardiology applications
over a mobile network that required continuous ECG transmission
with little delay. The combination of both indices can overcome
the limitations inherent in each index alone. By using diverse
patterns of ECG testing data and a wavelet-transform-based low-
delay ECG compression algorithm, the reconstructed ECG qual-
(a) Original
(b) Reconstructed ECG
FIG. 10. Morphological testing data set 232 for compression ratio of
19:1 (a) original and (b) reconstructed electroencephalograms (ECGs).
 ECG COMPRESSION RATIO FOR MOBILE TELECARDIOLOGY
(a) Original
(b) Reconstructed ECG
FIG. 11. Morphological testing data set 15814 for compression ratio of
26:1 (a) original and (b) reconstructed electroencephalograms (ECGs).
ity was inspected by clinicians, allowing a comparison between
subjective and objective measures of quality. We observed that
the reconstructed ECG quality generally correlated inversely with
the CR, with some variation. However, some extraordinary cases
demonstrated good objective quality but bad subjective quality at
a high CR. Based on our observations, the empirically determined
CR of 10:1 is the maximum compression ratio that should be used
to compress continuous ECG signals in limited bandwidth applica-
tions with little delay.
Acknowledgments
This study was supported by a grant of the Korea Health 21
R&D Project, Ministry of Health and Welfare, Republic of Korea
(A020608).
© M ARY ANN LI E BE RT, I NC. • VOL. 14 NO. 2 • M ARCH 2008 TELEMEDICINE and e-HEALTH 163
REFERENCES
1. Alesanco A, Olmos S, Istepanian RS, Garcia J. Enhanced real-time ECG coder for pack-
etized telecardiology applications. IEEE Trans Inf Technol Biomed 2006;10:229–236.
2. Kim BS, Yoo SK, Lee MH. Wavelet-based low-delay ECG compression algorithm
for continuous ECG transmission. IEEE Trans Inf Technol Biomed 2006;10:77–83.
3. Rajoub BA. An efficient coding algorithm for the compression of ECG signals
using the wavelet transform. IEEE Trans Biomed Eng 2002;49:355–362.
4. Hilton ML. Wavelet and wavelet packet compression of electrocardiograms. IEEE
Trans Biomed Eng 1997;44:394–402.
5. Al-Fahoum AS. Quality assessment of ECG compression techniques using a wave-
let-based diagnostic measure. IEEE Trans Inf Technol Biomed 2006;10:182–191.
6. Al-Fahoum AS, Reza AM. Combined edge crispiness and statistical differ-
encing for deblocking JPEG compressed images. IEEE Trans Image Process
2001;10:1288–1298.
7. Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, Mietus
JE, Moody GB, Peng CK, Stanley HE. PhysioBank, PhysioToolkit, and Physionet:
Components of a new research resource for complex physiologic signals.
Circulation 2000;101:215–220.
8. Taddei A, Distante G, Emdin M, Pisani P, Moody GB, Zeelenberg C, Marchesi C. The
European ST-T Database: Standard for evaluating systems for the analysis of ST-T
changes in ambulatory electrocardiography. Eur Heart J 1992;13:1164–1172.
9. Laguna P, Mark RG, Goldberger A, Moody GB. A database for evaluation of the
algorithms for measurement of the QT and other waveform intervals in the ECG.
Comput Cardiol 1997;24:673–676.
10. PhysioNet: Research Resource for Complex Physiologic Signals. Available at:
http://www.physionet.org (Last accessed 2005).
11. Abo-Zahhad M, Rajoub BA. An effective coding technique for the compres-
sion of one-dimensional signals using wavelet transforms. Med Eng Phys
2002;24:185–199.
12. Khadra L, Al-Fahoum AS, Al-Nashash H. Detection of life threatening car-
diac arrhythmias using the wavelet transformation. Med Biol Eng Comput
1997;35:626–632.
13. Al-Fahoum AS, Howitt I. Combined wavelet transformation and radial basis
neural networks for classifying life-threatening cardiac arrhythmias. Med Biol
Eng Comput 1999;37:566–573.
Address reprint requests to:
Sun K. Yoo
Medical Engineering
Yonsei University College of Medicine
Sudaemoon-gu Shinchon-dong
Seoul 120-752, South Korea
E-mail: sunkyoo@yuhs.ac
Received: April 2, 2007
Accepted: May 9, 2007