ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-Complex Estimation

Mohammed Abo-Zahhad, Sabah M. Ahmed & Ahmed Zakaria
Signal Processing – An International Journal (SPIJ), Volume (4) : Issue (2) 138
ECG Signal Compression Technique Based on Discrete Wavelet
Transform and QRS-Complex Estimation
Mohammed Abo-Zahhad zahhad@yahoo.com
Electrical and Electronics Engineering Department,
Faculty of Engineering, Assiut University,
Assiut, 71515, Egypt.
Sabah Mohamed Ahmed sabahma@yahoo.com
Ahmed Zakaria azakaria@yahoo.com
Abstract
In this paper, an Electrocardiogram (ECG) signal is compressed based on
discrete wavelet transform (DWT) and QRS-complex estimation. The ECG signal
is preprocessed by normalization and mean removal. Then, an error signal is
formed as the difference between the preprocessed ECG signal and the
estimated QRS-complex waveform. This error signal is wavelet transformed and
the resulting wavelet coefficients are threshold by setting to zero all coefficients
that are smaller than certain threshold levels. The threshold levels of all
subbands are calculated based on Energy Packing Efficiency (EPE) such that
minimum percentage root mean square difference (PRD) and maximum
compression ratio (CR) are obtained. The resulted threshold DWT coefficients
are coded using the coding technique given in [1], [21]. The compression
algorithm was implemented and tested upon records selected from the MIT - BIH
arrhythmia database [2]. Simulation results show that the proposed algorithm
leads to high CR associated with low distortion level relative to previously
reported compression algorithms [1], [15] and [19]. For example, the
compression of record 100 using the proposed algorithm yields to CR=25.15
associated with PRD=0.7% and PSNR=45 dB. This achieves compression rate of
nearly 128 bit/sec. The main features of this compression algorithm are the high
efficiency and high speed.
Keywords: ECG Signals Compression, QRS-complex estimation, Energy Packing Efficiency, Discrete
Wavelet Transform.

1. INTRODUCTION
An ECG signal is a graphical representation produced by an electrocardiograph, which records
the electrical activity of the heart over time. The ambulatory monitoring system usually requires
continuous 12 or 24-hours ambulatory recording for good diagnostic quality. For example, with
the sampling rate of 360 Hz, 11 bit/sample data resolution, a 24-h record requires about 43 M-
Byte per channel. So, 12-channel system requires nearly 513.216 M-Byte of storage disks daily.
Because of the tremendous amount of ECG data generated each year, an effective data
compression schemes for ECG signals are required in many practical applications including ECG
data storage or transmission over telephone line or digital telecommunication network. ECG data
compression techniques are typically classified into three major categories; namely direct data
compression [3]-[4], transform coding [5]-[8], and parameter extraction methods [9]-[11]. The
direct data compression methods attempt to reduce redundancy in the data sequence by
examining a successive number of neighboring samples. These techniques generally eliminate
samples that can be implied by examining preceding and succeeding samples. Examples of this
approach include amplitude zone epoch coding AZTEC [3], coordinate reduction time encoding
system(CORTES), delta coding algorithms, the SLOPE and the approximate Ziv-Lempel
algorithm (ALZ77) [4]. Transform coding of ECG signals is one of the most widely used
compression techniques. In these techniques a linear transformation is applied to the signal and
then compression via redundancy reduction is applied in the transform domain rather than in the
time domain. Typically, the transformation process produces a sequence of coefficients which
reduces the amount of data needed to adequately represent the original signal. Many different
transformations have been employed: Karhunen–Loeve transform (KLT), Fourier transform (FT),
Cosine transform (CT), Walsh transform (WT), Legendre transform (LT), the optimally warped
transform and subband coding [5]-[7]. In recent years the wavelet transform (WT) [13]-[14] has
received great attention. The wavelet transform techniques are based on consideration of the
hierarchical relationship among subband coefficients of the pyramidal wavelet decomposition as
the algorithms proposed in [15]-[17]. Finally, in parameter extraction methods, a set of
parameters is extracted from the original signal which is used in the reconstruction process. The
idea is to quantize a small set of extracted signal features. The methods that can be classified in
this group are: peak-peaking methods [9], cycle-pool-based compression (CPBC) algorithms,
neural network methods [10] and linear prediction methods [11].
In this paper a new compression technique based on wavelet transform and QRS-complex
estimation is proposed. There are two motivations in this work. The first motivation is the QRS-
complex estimation using the extraction of significant features of ECG waveform. The second
motivation is the selection of the threshold levels in each subband such that high CR and low
PRD are obtained. The significant features of ECG waveform are extracted to estimate the QRS-
complex. Then, the estimated QRS-complex is subtracted from the original ECG signal. After
that, the resulting error signal is wavelet transformed and the DWT coefficients are threshold
based on the energy packing efficiency. Finally the significant coefficients are coded and stored
or transmitted.
The structure of this paper is as follows: Section 2 is a review of the discrete wavelet transform.
Section 3 presents the QRS-complex detection system. Section 4 is an overview of the energy
packing efficiency principle and coefficients thresholding. Section 5 shows how the coding
technique works. In section 6 the algorithm is tested on selected records from the MIT - BIH
arrhythmia database and compared with other coders in the literature [1], [15] and [19]. Finally
the conclusion of the paper is presented in section 7.
2. Discrete Wavelet Transform

The continuous wavelet transform (CWT) transforms a continuous signal into highly redundant
signal of two continuous variables — translation and scale. The resulting transformed signal is
easy to interpret and valuable for time-frequency analysis. The continuous wavelet transform of
continuous function, )(xf relative to real-valued wavelet, )(x is described by:



 dxxxfsW s )()(),( ,  (1)
where,
)
-
(
1
)(,
s
x
s
xs

   (2)
s and are called scale and translation parameters, respectively. ),(  sW denotes the wavelet
transform coefficients and  is the fundamental mother wavelet. If ),(  sW is given, )(xf can
be obtained using the inverse continuous wavelet transform (ICWT) that is described by:
dsd
s
x
sW
C
xf
s






2
,
0
)(
),(
1
)(  
 

 (3)
where, )(u is the Fourier transform of )(x and




 du
u
u
C
||
|)(| 2
 (4)
The discrete wavelet transform can be written on the same form as Equation (1), which
emphasizes the close relationship between CWT and DWT. The most obvious difference is that
the DWT uses scale and position values based on powers of two. The values of s and τ are:
jj
ks 2*,2   and
2
),( Zkj  as shown in Equation (5).
)
-
(
1
)(, j
o
j
oo
j
o
kj
s
skx
s
x

  (5)
The key issues in DWT and inverse DWT are signal decomposition and reconstruction,
respectively. The basic idea behind decomposition and reconstruction is low-pass and high-pass
filtering with the use of down sampling and up sampling respectively. The result of wavelet
decomposition is hierarchically organized decompositions. One can choose the level of
decomposition j based on a desired cutoff frequency. Figure (1-a) shows an implementation of a
three-level forward DWT based on a two-channel recursive filter bank, where )(0 nh and
)(1 nh are low-pass and high-pass analysis filters, respectively, and the block 2 represents
the down sampling operator by a factor of 2. The input signal )(nx is recursively decomposed
into a total of four subband signals: a coarse signal
)(3 nC
, and three detail

signals,
)(),( 23 nDnD
, and
)(1 nD , of three resolutions. Figure (1-b) shows an implementation
of a three-level inverse DWT based on a two-channel recursive filter bank, where
)(
~
0 nh
and
)(
~
1 nh are low-pass and high-pass synthesis filters, respectively, and the block 2 represents
the up sampling operator by a factor of 2. The four subband signals
)(),(),( 233 nDnDnC
and
)(1 nD , are recursively combined to reconstruct the output signal
)(~ nx . The four finite
impulse response filters satisfy the following relationships:
)()1-()( 01 nhnh n
 (6)
)-1()(
~
00 nhnh  (7)
)-1()1-()(
~
01 nhnh n
 (8)
so that the output of the inverse DWT is identical to the input of the forward DWT.
(a)
(b)
Figure (1): A three-level two-channel iterative filter bank (a) forward DWT (b) inverse DWT
)(3 nC
)(3 nd
)(2 nd
)(1 nd
)(nx
↓2)(nho
↓2)(1 nh
↓2)(nho
↓2)(1 nh
↓2)(nho
↓2)(1 nh
)(3 nC
)(3 nd
)(2 nd
)(1 nd
)(~ nx
2↑
2↑
⊕
)(
~
0 nh
)(
~
1 nh
2↑
2↑
⊕
)(
~
0 nh
)(
~
1 nh
2↑
2↑
⊕
)(
~
0 nh
)(
~
1 nh

3. The QRS-Complex Detection
A typical scalar ECG heartbeat is shown in Figure (2). The significant features of the ECG
waveform are the SRQP ,,, and T waves and the duration of each wave.
Figure (2): Typical ECG signal
A typical ECG tracing of electrocardiogram baseline voltage is known as the isoelectric line. It is
measured as the portion of the tracing following the T wave and preceding the next P wave.
The aim of the QRS-complex estimation is to produce the typical QRS-complex waveform using
the parameters extracted from the original ECG signal. The estimation algorithm is a Mat lab
based estimator and is able to produce normal QRS waveform. A single heartbeat of ECG signal
is a mixture of triangular and sinusoidal wave forms. The QRS-complex wave can be represented
by shifted and scaled versions of these waveforms. The ECG waveform contains, in addition to
the QRS-complex, P and T waves, 60-Hz noise from power line interference, EMG signal from
muscles, motion artifact from the electrode and skin interface, and possibly other interference
from electro surgery equipments.
The power spectrum of the ECG signal can provide useful information about the QRS-complex
estimation. Figure (3) summarizes the relative power spectra (based on the FFT) of the ECG, QRS-
complex, P andT waves, motion artifact, and muscle noise taken for a set of 512 sample points
that contain approximately two heartbeats [18]. It is visible that QRS-complex power spectrum
involves the major part of the ECG heartbeat. Normal QRS-complex is 0.06 to 0.1 sec in duration
and not every QRS-complex contains a Q wave, R wave, and S wave. By convention, any
combination of these waves can be referred to as a QRS-complex. This portion can be represented

by
R,Q and S values, the
SRandRQ -- durations and the event time of R as shown in
Figure (2). These values can be extracted from the original ECG signal.
Figure (3): Relative power spectra of QRS-complex, P and T waves, muscle noise and motion artifacts.
4. Optimal Energy Packing Efficiency and Thresholding
The QRS-complex estimator is tested on the first 1000 sample of record 100 from the MIT-BIH
arrhythmia database. Figure (4) illustrates the original signal, the resulting estimated QRS-
complex signal and the difference between them. After applying the DWT on the error signal, the
resulted wavelet coefficients are divided into the following subbands:
]D.........DDDA[ 12L1LLL -- (9)
Where A refers to the approximation coefficients, D refer to the details coefficients and L
denotes the decomposition level. To demonstrate the optimal energy packing efficiency principle,
the first 4096 samples of record 103 is considered as a test signal. The percentage subband
energy of all subbands of this signal are decomposed up to the sixth level using "bior4.4" wavelet
filter are illustrated in Table (1). It illustrates also, the percentage energy and the number of
coefficients in each subband for the original signal (S_org), the normalized signal (S_norm) and
the error signal (S_diff).

0 100 200 300 400 500 600 700 800 900 1000
-0.1
0
0.1
0.2
(a)
0 100 200 300 400 500 600 700 800 900 1000
-0.1
0
0.1
0.2
(b)
0 100 200 300 400 500 600 700 800 900 1000
-0.06
-0.04
-0.02
0
0.02
(c)
Figure (4): The first 1000 sample of record 100. a) The original signal. b) The estimated QRS -
complex signal. b) Difference between the original and the estimated QRS-complex
signal.
Percentage Subband Energy
Subbands The original signal
(S_org)
The normalized
signal (S_norm)
The error signal
(S_diff)
Number of
coefficients
in each subband
99.26284 % 13.80015 % 66.43408938 % 82
0.197928 % 23.14474 % 28.42895567 % 82
0.268063 % 31.34588 % 2.722341084 % 146
0.241415 % 28.22983 % 1.333762044 % 273
0.028619 % 3.346545 % 0.878562618 % 528
1.09E-03 % 0.127727 % 0.181650989 % 1038
4.38 E-05 % 0.005124 % 0.020638213 % 2057
Table (1): Record 103 percentage subbands energy, number of coefficients in each subband
(S_org), the normalized signal (S_norm) and the error signal (S_diff).
Thresholding of a certain subband coefficients is done by eliminating all coefficients that are
smaller than a certain threshold level L . This process introduces distortion in a certain aspect in
the reconstructed signal. To decrease the effect of thresholding, threshold levels in all subbands
are defined according to the energy contents of each subband. For this purpose, a percentage
quantity (EPE) represents a measure of the total preserved energy of a certain subband after
thresholding compared to the total energy in that subband before thresholding is defined as [21]:

%100*=
i
i
C
C
i
E
E
EPE (10)
where, iCE is the total energy of the coefficients in the ith subband after thresholding and iCE is
the total energy of the coefficients in this subband before thresholding.
An optimization routine has been developed to find the threshold level of each subband that
yields to the highest CR and the lowest PRD. This has been achieved through the minimization of
the function P= PRD + 1/CR. However, since the value of 1/CR is small relatively to the values of
PRD, a weighting factor W is introduced to increase the percentage of sharing of 1/CR. So, P is
rewritten in the form:
P= PRD + W/CR (11)
The selection of W is based on which is more important: high CR or low PRD. The Mat lab
optimization toolbox is adopted to perform the minimizing the objective function with threshold
level T is a parameter. In a certain subband, the threshold level is calculated by carrying out the
following steps for a predefined preserved energy E':
1- Calculate the total energy E of the DWT coefficients C in each subband using E =

M
n
C
1
2
, where M is the number of subband coefficients .
2- Calculate the probability distribution function f; [f, V] = hist (abs(C), 100).
3- Calculate the energy E (L) = )(*)(
0
2
iffV
l
i

.
4- Threshold level T is the coefficient at which, l = k, where E (k) ≤ E'.
5. The Coding Technique
As stated in section 3, QRS-complex contains the most energy of the ECG signal. According to
this observation, extracting the QRS-complex data and dealing with it in an accurate manner
leads to low PRD and enhanced compression ratio. Moreover, compressing the difference
between the original signal and the estimated QRS-complex one improves the overall CR. Figure
(5) illustrates the block diagram of the proposed compression algorithm. The following steps
detail the proposed algorithm.
A. Preprocessing
Firstly, the ECG signal ].......[ 4321 Nxxxxxx  is preprocessed by normalization
and mean removal using the following relation:
Nnm
A
nx
ny x
m
,....,2,1,-
)(
)(  (12)
Where, )(nx and )(ny are the original and normalized signal samples respectively and N
denotes the length of the original signal. Am and mx are the maximum value of the original ECG

signal and the average of the normalized ECG signal respectively. The ECG signal is normalized
by dividing the original signal by its maximum value Am. Consequently, all DWT coefficients will
be less than one. Mean removal is done by subtracting from the normalized ECG signal its mean
mx to reduce the number of the significant wavelet coefficients.
B. QRS-Complex Estimation.
Detecting QRS-complex and the extraction of its significant features are performed before the
transformation process. The features extracted include the locations and the amplitudes of the Q,
R and S peaks and their values.
Figure (5): The block diagram of the ECG compression algorithm
C. Discrete Wavelet Transformation of the Error Signal
The error signal is discrete wavelet transformed up to decomposition level L. To obtain perfect
reconstruction, the selected mother wavelet must be compactly supported. The selection of
wavelet filter, decomposition level and signal length have great influence in the determination of
the algorithm performance [22]-[24]. Here, the 'bior 4.4' wavelet filter is adopted.

D. Thresholding Process
The wavelet coefficients representing the error signal are threshold according to the energy
packing efficiency principle explained in section 4. The intent of this part is to investigate the
optimal values of EPE that achieve maximum CR and minimum PRD. To encounter this, the error
signal is coded without thresholding for all subbands. Consequently, the data stream is decoded
to obtain the reconstructed ECG signal. Then the CR, PRD and the objective function P are
calculated. As a result the threshold level in each subband is calculated.
E. Coding of the Wavelet Coefficients
The coded stream consists of two parts:
1- The header part.
2- The significant and non significant coefficients part.
The header consists of two sections. The first section has 50 bits: 20 bits are dedicated for storing
the total number of wavelet coefficients, 12 bits is dedicated for storing the maximum value in the
original signal, 12 bits is dedicated for storing the mean of the normalized signal and the last 6
bits are dedicated for the number of beats contained in the signal. The second section is
constructed from 66 bits: 36 bits to represent the RQ, and S values and 30 bits to represent
SR,RQ -- durations and the event time of R . Figures (6-a) and (6-b) illustrate the coding
stream that represents the header part.
mx A*]m)n(y[)n(X += (13)
Number of beatsMaximum ValueMean ValueNumber of coefficients
6 Bits12 Bits12 Bits20 Bits
(a) The first section of the header.
event time of RR–S durationQ–R durationS ValueR ValueQ Value
10 Bits10 Bits10 Bits12 Bits12 Bits12 Bits
(b) The second section of the header.
Figure (6): The coding stream of the header part.
The significant and insignificant coefficients are coded separately. The runs of significant
coefficients are coded as follows:
 One bit of value '1' identifies the run of significant coefficients.
 A sign bit to encode the sign of the significant coefficient.
 Eight bits to encode the value of the significant coefficient.

Figure (7-a) illustrates the coding stream that represents runs of significant coefficients. The runs
of insignificant coefficients are coded with variable-length coding (VLC) based on run length
encoding as follows:
 One bit of value '0' identifies the run of insignificant coefficients.
 Four bits to represent the number of bits needed to code the run length.
 Variable in length code (from 1 to 16 bits) to represent the run length.
Figure (7-b) illustrates the coding stream that represents runs of insignificant coefficients.
CoefficientSign'1'
8 Bits1 Bit1 Bit
(a) Representation of runs of significant coefficients.
Run lengthNumber of VLC bits'0'
1 < VLC < 164 Bits1 Bit
(b) Representation of insignificant coefficients.
Figure (7): The coding stream of the significant and insignificant coefficients.
The compression ratio CR, the percent RMS difference PRD and the peak signal to noise ratio
PSNR, are used as a performance measure. The three measures are defined by:
streamoutputoflength
nxofLength
CR
11*)(
 (14)




 N
n
N
n
nx
nxnx
PRD
1
2
1
2
])([
)](-)([
(15)


N
n
nxnx
N
nx
PSNR
1
2
10
)](-)([
1
)]([max
log20 (16)
where, )(nx and )(nx represent the original and the reconstructed signals respectively.
F. Post Processing
The reconstructed ECG signal is obtained from decoded signal )(y n by adding the decoded
mean value and multiply it by the maximum value.
6. Experimental Results

MIT-BIH arrhythmia database has been adopted for evaluating the performance of the proposed
compression technique. The ECG signals of the database were sampled at 360 Hz and each
sample was represented by 11 bit/sample (total bit-rate of 3960 bit/s). Two datasets formed by
taking certain records from the MIT-BIH database were used for the evaluation process. These
datasets were used in the evaluation of other coders in earlier studies [1], [15] and [19]. The first
dataset consists of the first 10 minutes from records 100, 101, 102, 103, 111, 115, 117, and 118.
The second dataset consists of the first 1 minute from records 104, 111, 112, 115, 119, 201, 207,
208, 214 and 232. The wavelet decomposition is carried out using the 'bior4.4' filter up to the sixth
level. Here, the compression ratio, percent RMS difference and peak signal to noise ratio
performance measures of the proposed method are compared with other coders in the literature
[1], [15] and [19].
In the first experiment, the proposed algorithm is tested on records 100 and 103 in order to
explore the effect of compression on the clinical information of the ECG records. Figures (8) and
(9) show the two records before and after compression together with the difference between them
(error signal). The optimal EPE values for all subbands are listed in Table (2).
Parameters in SubbandsRecord
Number
Parameters
A6 D6 D5 D4 D3 D2 D1
Maximum
Coefficient
Values
0.2347 0.0510 0.0323 0.0694 0.0484 0.0174 0.0058
Optimal
Threshold
Levels
0.0190 0.0244 0.0159 0.0207 0.0305 0.0174 0.0058
100
EPE at Optimal
Threshold Levels
99.70 86.47 81.32 69.12 59.94 1.999 0.2612
Maximum
Coefficient
Values
0.2527 0.1603 0.0619 0.0660 0.0433 0.0166 0.0054
Optimal
Threshold
Levels
0.0238 0.0201 0.0240 0.0246 0.0258 0.0166 0.0054
103
EPE at Optimal
Threshold Levels
99.92 99.58 91.12 62.86 63.24 2.228 0.868
Table (2): Maximum coefficient values, optimal threshold levels and EPE values of records 100 and 103.

0 200 400 600 800 1000 1200 1400 1600 1800 2000
900
1000
1100
1200
The original signal
0 200 400 600 800 1000 1200 1400 1600 1800 2000
900
1000
1100
1200
The reconstructed signal
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-20
0
20
The error signal
Figure (8): The original, reconstructed and error signals of the first 4096 samples of record 100
(CR= 15.4, PRD= 0.43 and PSNR= 49.4).
0 200 400 600 800 1000 1200 1400 1600 1800 2000
1000
1200
1400
The original signal
0 200 400 600 800 1000 1200 1400 1600 1800 2000
1000
1200
1400
The reconstructed signal
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-20
0
20
The error signal
Figure (9): The original, reconstructed and error signal the first 4096 sample of Record 103
Record 103 (CR= 14.2, PRD= 0.474 and PSNR= 49.6).

To test the quality of the reconstructed ECG signals, a Mean Opinion Score (MOS) test was
adopted on the reconstructed signals [20]. In this test both the original and reconstructed ECG
signals of a certain record are printed in a paper form and the cardiologists evaluators are asked
to see the signals. For every tested signal, the evaluators were asked to answer some questions
about the similarity between the printed signals. These questions are listed in Table 3. The
percentage MOS error for any tester k is given by:
100)D1()factor1(100
5
C5
factorMOSk ×-×-+×
-
×= (17)
where,
C is a five scale that measures the similarity between the original signal and the reconstructed
one (1 for completely different signals and 5 for identical signals).
D is the answer to the Boolean question about the diagnosis (0—YES, 1—NO).
and factor is a weighting coefficient between the measure of similarity and the Boolean question (0
to less than 1).
Table (3): MOS Test Questionnaire
The mean percentage MOS error is determined as follows:
MOS% =
V
N
k
k
N
MOS
V
∑
1
x 100 (18)
Comparison of ECG signal N0. (-----) with its original signal
1- Details of tester
Name: _________________ Date: __________________
2- The measure of similarity between the original and reconstructed one (circle
one number).
1 2 3 4 5
Completely different Identical
3- Would you give a different diagnosis with the tested signal if you hadn’t seen
the original signal? (circle YES or NO).
YES NO
4- Comments:
__________________________________________________________________
__________________________________________________________________

where, VN is the number of evaluators. The lower the value of the MOS error, the evaluation quality of
the reconstructed signal is better. A rough classification of signal quality to be good if the percentage MOS
error is less than 35 % [20]. The ECG signal of record 100 are printed at many CR and PRD and brought
forward to the evaluator cardiologists, then the percentage MOS is calculated using equation (18). Table
(4) lists the resulted MOS error of all testers for Record 100 compressed at different values of CR and
PRD. It is clear form the table that the MOS error values is less than 35%, which mean that all tested
signals are acceptable from the point view of the cardiologists' evaluators and there are no loss in the
clinical information of the ECG signal.
kMOS %
CR PRD PSNR
1'st evaluator 2'nd evaluator 3'rd evaluator
MOS %
3.4683 0.2069 55.8949 0 0 0 0
7.4032 0.2374 54.6985 0 0 0 0
10.0654 0.2948 52.8183 0 0 0 0
14.2823 0.4429 49.2838 0 0 0 0
17.7061 0.5297 47.7294 0 0 0 0
19.2492 0.6042 46.5855 4 4 2 3.33
22.7517 0.6637 45.7702 4 6 2 4
24.2497 0.7074 45.2159 4 6 2 4
24.8425 0.7451 44.7653 4 6 2 4
27.2626 0.802 44.1255 4 6 2 4
28.2247 0.837 43.7551 4 6 4 4.66
31.4125 0.8603 43.5165 6 6 4 5.33
31.1017 0.8753 43.3662 6 6 4 5.33
32.0227 0.8859 43.2611 6 6 4 5.33
30.2728 0.8707 43.4121 6 6 4 5.33
Table (4): The MOS error of the three evaluators for Record 100.
The second experiment discusses the effect of the weighting factor W on the performance measures
(CR, PRD, and PSNR) of the compression algorithm. The experiment is executed upon the first 4096 of
record 100 while, the weighting factor W varies from 0 up to 100. Figure (10) shows the CR, PRD, and
PSNR at the optimal vales of threshold levels versus the weighting factor W. It is cleared from the
results of this experiment that the CR gets rise versus the increase in the weighting factor W. The
performance results of this experiment versus the total percentage EPE of the thresholded coefficients
is shown in Figure (11). Relying on the results shown in Figure (10), when W = 0 only PRD affect the

optimization parameter P which results in CR = 3.406, PRD = 0.202 and PSNR = 55.9. On the other
hand, when W = 100, only CR affect the optimization parameter P, which results in CR = 36.3, PRD =
0.99 and PSNR = 42.11. It is cleared from the obtained results that the CR and PRD increase with the
increase in the weighting factor W, while PSNR decreases with the increase in W.
0 10 20 30 40 50 60 70 80 90 100
0
20
40
CR
weighting factor W
0 10 20 30 40 50 60 70 80 90 100
0
0.5
1
1.5
PRD
weighting factor W
0 10 20 30 40 50 60 70 80 90 100
40
50
60
PSNR
weighting factor W
Figure (10): The CR, PRD and PSNR results versus the weighting factor W for Record 100.

55 60 65 70 75 80 85 90 95 100
0
20
40
EPEt %
CR
55 60 65 70 75 80 85 90 95 100
0
0.5
1
1.5
EPEt %
PRD
55 60 65 70 75 80 85 90 95 100
40
50
60
EPEt %
PSNR
Figure (11): The performance results of record 100 versus the total EPE %.
The third experiment studies the performance of the proposed algorithm in compressing the first
and second data sets. Figure (12 a, b and c) and Figure (13 a, b and c) show the results of this
experiment. The results indicate that, the performance results are dependant on the compressed
ECG signal. For the first data set the highest CR achieved is for record 101, and the performance
measure are CR = 40, PRD = 2.7% and PSNR = 32.5 dB. On the other hand, the smallest PRD
achieved is for record 100, and the performance measures are CR = 3.4, PRD = 0.2% and PSNR
= 55 dB. For the second data set the highest CR achieved is for record 232, and the performance
measures are CR = 40.5, PRD = 1.5% and PSNR = 37.5 dB. On the other hand, the smallest
PRD achieved is for record 232, and the performance measures are CR = 2.5, PRD = 0.2% and
PSNR = 53 dB.
7. Conclusion
In this paper, a new method for compressing ECG signal based on wavelet transform has been
proposed. The key idea lies in the estimation of QRS-complex signal from a given ECG signal.
The QRS-complex is estimated using parameters extracted from the original ECG signal. This
method is applied to many ECG records selected from the MIT-BIH arrhythmia database. It
results in CR higher than previously published results [1], [15], [19] with less PRD as shown in
Table (5).

Coding scheme Record CR PRD
117 22.19 : 1 1.06%
117 10.80 : 1 0.48%
232 4.314 : 1 0.30%
210 11.55 : 1 0.44%
Reference [1]
119 23.0 : 1 1.95%
Reference [15] 117 8.00 : 1 1.18%
Reference [19] 101 26.64 : 1 9.14%
119 23.00:1 1.95%
232 4.314 : 1 0.25%
210 11.55 : 1 0.49%
Proposed algorithm
101 26.70:1 1.77%
Table 5: Summary of CR and PRD results for some MIT-BIH arrhythmia database
records using different algorithms versus the proposed algorithm.

60 65 70 75 80 85 90 95 100
10
15
20
25
30
35
EPEt (%)
CR
100
101
102
103
111
115
117
118
( a )
60 65 70 75 80 85 90 95 100
0.5
1
1.5
2
2.5
3
3.5
4
EPEt (%)
PRD(%)
100
101
102
103
111
115
117
118
( b )

60 65 70 75 80 85 90 95 100
30
35
40
45
50
55
EPEt (%)
PSNR(dB)
100
101
102
103
111
115
117
118
( c )
Figure (12): The performance results for compressing the first data set.
60 65 70 75 80 85 90 95 100
5
10
15
20
25
30
35
EPEt (%)
CR
104
111
112
115
119
201
207
208
214
232
( a )

40 50 60 70 80 90 100
0
0.5
1
1.5
2
2.5
3
EPEt (%)
PRD(%)
104
111
112
115
119
201
207
208
214
232
( b )
60 65 70 75 80 85 90 95 100
30
35
40
45
50
55
EPEt (%)
PSNR(dB)
104
111
112
115
119
201
207
208
214
232
( c )
Figure (13): The performance results for compressing the second data set.

8. References
[1] B. A. Rajoub, “An efficient coding algorithm for the compression of ECG signals using the
wavelet transform,” IEEE Transactions on Biomedical Engineering, 49 (4): 355–362, 2002.
[2] MIT-BIH Arrhythmia Database, www.physionet.org/physiobank/database/mitdb.
[3] J. Cox, F. Nulle, H. Fozzard, and G. Oliver, “AZTEC, a preprocessing program for real-time
ECG rhythm analysis,” IEEE. Trans. Biomedical Eng., BME-15: 128–129, 1968.
[4] R.N. Horspool and W.J. Windels, “ECG compression using Ziv-Lempel techniques, Comput”
Biomed. Res., 28: 67–86, 1995.
[5] B. R. S. Reddy and I. S. N. Murthy, “ECG data compression using Fourier descriptors,” IEEE
Trans. Biomed. Eng., BME-33 (4): 428–434, 1986.
[6] H. A. M. Al-Nashash, “ECG data compression using adaptive Fourier coefficients estimation,”
Med. Eng. Phys., 16: 62–66, 1994.
[7] S. C. Tai, “Improving the performance of electrocardiogram sub-band coder by extensive
Markov system,” Med. Biol. Eng. And Computers, 33: 471–475, 1995.
[8] J. Chen, S. Itoh, and T. Hashimoto, “ECG data compression by using wavelet transform,”
IEICE Trans. Inform. Syst., E76-D (12): 1454–1461, 1993.
[9] A. Cohen, P. M. Poluta, and R. Scott-Millar, “Compression of ECG signals using vector
quantization,” in Proc. IEEE-90 S. A. Symp. Commun. Signal Processing COMSIG-90,
Johannesburg, South Africa, pp. 45–54, 1990.
[10]G. Nave and A. Cohen, “ECG compression using long-term prediction,” IEEE. Trans. Biomed.
Eng., 40: 877–885, 1993.
[11]A. Iwata, Y. Nagasaka, and N. Suzumura, “Data compression of the ECG using neural
network for digital Holter monitor,” IEEE Eng. Med. Biol., Mag, pp. 53–57, 1990.
[12]M. Abo-Zahhad, S. M. Ahmed, and A. Al-Shrouf, “Electrocardiogram data compression
algorithm based on the linear prediction of the wavelet coefficients," in Proc.7th IEEE Int.
Conf., Electronics, Circuits and Systems, vol. 1, Lebanon, pp. 599–603, 2000.
[13]O. O. Khalifa, S. H. Harding, A. A. Hashim, “Compression Using Wavelet Transform” Signal
Processing: An International Journal (SPIJ), pp. 17 – 26, 2008.
[14]M. Zia Ur Rahman , R. A. Shaik, D V Rama Koti Reddy, “Noise Cancellation in ECG Signals
using Computationally Simplified Adaptive Filtering Techniques: Application to Biotelemetry”
Signal Processing: An International Journal (SPIJ), pp. 120 – 131, 2009.
[15]Z. Lu, D. Y. Kim, and W. A. Pearlman, “Wavelet compression of ECG signals by the set
partitioning in hierarchical trees algorithm,” IEEE Trans. on Biomedical Engineering, 47(7):
849–856, 2000.
[16]Shen-Chuan Tai, Chia-Chun Sun, and Wen-Chien Yan, “A 2-D ECG Compression Method
Based on Wavelet Transform and Modified SPIHT.” IEEE Transactions on Biomedical
Engineering, 52 (6), 2005.

[17]M. Okada, “A digital filter for the QRS complex detection,”IEEET trans. Biomed. Eng., BME-26
(12): 700–703, 1979.
[18]Thakor, N. V., Webster, J. G., and Tompkins, W. J., Optimal QRS detector. Medical and
Biological Engineering, pp. 343–50, 1983.
[19]Yaniv Zigel , Arnon Cohen, and Amos Katz,” ECG Signal Compression Using Analysis by
Synthesis Coding”, IEEE Transactions on Biomedical Engineering, 47 (10), 2000.
[20]Y. Zigel, A. Cohen, and A. Katz, “The weighted diagnostic distortion measure for ECG signal
compression,” IEEE Trans. Biomed. Eng., 2000.
[21]Abo-Zahhad, M. and Rajoub, B.A., An effective coding technique for the compression of one-
dimensional signals using wavelet transform. Med. Eng. Phys. 24: 185-199, 2001.
[22]Ahmed, S.M., Al-Zoubi, Q. and Abo-Zahhad, M., "A hybrid ECG compression algorithm based
on singular value decomposition and discrete wavelet transform," J. Med. Eng. Technology
31: 54-61, 2007.
[23]S.M. Ahmed, A. Al-Shrouf and M. Abo-Zahhad, "ECG data compression using optimal non-
orthogonal wavelet transform," Medical Engineering & Physics, 22 (1): 39-46, 2000.
[24]R. Javaid, R. Besar, F. S. Abas, “Performance Evaluation of Percent Root Mean Square
Difference for ECG Signals Compression” Signal Processing: An International Journal
(SPIJ): 1–9, 2008.

ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-Complex Estimation

More Related Content

What's hot (20)

Similar to ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-Complex Estimation (20)

Recently uploaded (20)

ECG Signal Compression Technique Based on Discrete Wavelet Transform and QRS-Complex Estimation