International Journal Of Electrical, Electronics And Data Communication, ISSN: 2320-2084 Volume-4, Issue-9, Sep.

-2016

DE-NOISING AND STATISTICAL FEATURE EXTRACTION OF THE

ECG SIGNAL USING WAVELET ANALYSIS
1
TANYA KAMBO, 2RAM AVTARJASWAL
1
M.tech Scholar, 2Assistant Professor,
Dept. of electrical engineering, U.I.E.T, Kurukshetra
E-mail: 1tanyakamboj08@gmail.com, 2ramavtar.jaswal@gmail.com

Abstract— The electrocardiogram is a technique of recording bioelectric currents generated by the heart which is useful for
diagnosing many cardiac diseases. The feature extraction and de-noising of ECG are highly useful in cardiology. And
baseline wander elimination is considered as a traditional problem. So this paper deals with the noise removal and feature
extraction of the signal. This baseline wander noise is removed by Butterworth filter. We present a wavelet-transform (WT)
based search algorithm. The algorithm computes wavelet packet coefficients and then in each scale the different nine features
of the signal is calculated. Comparison is made between the characteristics of signals and the branch of the wavelet binary
tree corresponding to minimum entropy wavelet spaces is chosen. This algorithm is tested using the data record from
MIT/BIH database for arrhythmia disease categorized into normal and abnormal and excellent results are obtained.

Keywords—ECG signal, de-noising, feature extraction, MIT database, wavelet packet decomposition.

I. INTRODUCTION removal of baseline wander. And the estimated this

Electrocardiogram (ECG) is nearly a periodic signal
that imitates the activity of the heart. Much
information on the normal and abnormal condition of
heart can be obtained from ECG. As the ECG signal
being non-stationary in nature, soit is very difficult to
examine the signal visually. Thus the need arises for
computer based methods for ECG signal analysis.
Observation of ECG signal in hospitals can take long
hours and can result in tiresome. And so we cannot be
relied upon visual analysis. This calls need for
computer-based techniques for signal analysis. At
every beat, the heart is depolarized to trigger its
contraction [1]. This electrical activity is transmitted
throughout the body and can be picked up on the skin
which is the principle behind the ECG. An ECG
machine records this activity via electrodes on the
skin and displays it graphically on the screen. An
ECG involves attaching of 10 electrical cables to the
body: one to each limb and six across the chest [6].
ECG is a wave that represents an electrical event in
the heart such as atria depolarization, ventricular
depolarization and repolarization. One cardiac cycle
of ECG signal consists of the P-QRS-T wave. The
major useful information in the ECG is taken from
the intervals and amplitudes called as its features. As
for the analysis of the signal it must be noise free [5].
As when we record the signal it contains many noises
like power line interference and baseline wandering.
So to removethese by using appropriate digital filters
like fir filter and Butterworth filter [9,11].There are
different type of noise is present in the signal such as
powerline interference, motion artifacts and baseline
wandering. Motion artifacts are due to motion of the
patient while recording the signal [11]. These can be
filtered using different digital filters like FIR and IIR
filters [11] but here we have chosen Butterworth filter
due to its simplicity. And the main consideration is of
De-noising and Statistical Feature Extraction of the ECG Signal Using Wavelet Analysis
noise can detect the diseases [2]. And then extracting
of features is an important tool for the analysis. There
are several methods to determine non-linear i.e.
statistical features like wavelet transform, FFT, STFT
[5]. But we have first decomposed the signal using
wavelet packet decomposition [4].These coefficients
are used for the non-statistical features [5, 8] which
tells about the original features of the signal. The
implementation of wavelet decomposition is easy as
compared to and it generates the wavelet tree [4]. So
after having coefficients and extracted features we
can differentiate the signals according to their
diseases [2]. Biomedical research in signal processing
of these biomedical signals are in very trend [4].
These are very efficient and interesting to read and
study.
II. DATA SELECTION
The collection of data is the one of the most
important chore of signal processing. In this work,
MIT–BIH Arrhythmia database directory of ECG
signals from physio.net ATM BANK is used. The
ECG signal of MIT–BIH Arrhythmia was obtained
from the Beth Israel Hospital Arrhythmia Laboratory
[8]. This database contains 48 files partitioned into
two parts first one is of 23 files (numbered from 100
to 124 inclusive with some numbers missing) selected
at random from this set, and another one contains 25
files. Each of the 48 records is slightly over 30 min
long [2]. The ECG signals from MIT–BIH database
are illustrated by – a text header file (.hea). The
header files describes the detailed information about
the signal such as number of samples, sampling
frequency and interval, format of ECG signal, type
and number of ECG leads. The sampling frequency
of the signal is 360Hz. Two records have been taken
105 and 118 from this database. These records have
analyzed to do comparison between them in their

38
International Journal Of Electrical, Electronics And Data Communication, ISSN: 2320-2084
features. According to [2]one is taken as normal and
another is abnormal.
III. PREPROCESSING
This process is the first and very vital stage of signal
processing, in which it is necessary to eliminate
noises from the signal by using appropriate filters.
For pre-processing of the ECG signal, noise
elimination comprisesvarious approaches for many
noises [11]. This preprocessing of ECG signal is
ended before extracting the features, which results in
accurate features to increase the efficiency of the
system. Preprocessing of ECG signal contains of de-
noising of ECG signal and baseline wander removal
using Butterworth filter.
3.1. Baseline wander removal
This type of noise artifacts that generally affects ECG
signals is baseline wandering. Generally it appears
from respiration and deceits between 0.15Hz and
0.3Hz [3]. It is necessary to eliminate this noise in
order to do ECG signal processing and to remove the
irregularities from the signal.Electrode and
respiration factors varies due to patient are vital
sources of this baseline wander in most of the ECGs
recordings. In this, the baseline wander of ECG is
removed by first packingthe signal taken from
database then smoothen the data by using
Butterworth filter of order 10[9]. This is a digital
filter which has good efficiency and simple to use.
Also it gives better result. It generates the coefficients
known as Butterworth coefficients which helps in
removing the noise [7].
3.2 Wavelet packet decomposition
Wavelet packet analysis is used to determine features
of the non-stationary likeECG signal and this is the
best method for finding features from ECG signal as
it decomposed the signal at various level in different
frequency bands. Wavelet transform gives time-
frequency representation of a signal and it gives
better resolution [7]. Wavelet packet decomposition
analysis is a generalor effective form of the DWT, in
which signal is riven into approximation and detailed
coefficients. The approximation coefficients is again
riven up into a second level approximation and
detailed coefficients, and this process is repeated till
we want get result. There are k+1 possible ways to
encrypt the signal for the k-level decomposition [3].
The signal is de-noised bylow and high pass filter
called quadrature mirror filters. In wavelet packet
analysis approximation and detailed coefficients can
be formed like a complete binary tree structure. This
decomposition process applies at both low and high
frequency sub – bandsto create next level in the
binary tree, it produces more than 22k-1 different ways
to encode the signal. For each level k, there are 2k-1
number of nodes remained after wavelet packet
decomposition [3]. Fig 1shows the wavelet packet
De-noising and Statistical Feature Extraction of the ECG Signal Using Wavelet Analysis
39
Volume-4, Issue-9, Sep.-2016
decomposition tree obtained at the 4th level of the
decomposition and data for node (4,15) means data at
4th level and 15th coefficient of decomposition.
Shannon entropy based criterion is used to find this
feature. Fourth level decomposition coefficients are
used to extract the features.
IV. FEATURE EXTRACTION
The pre-processed ECG signal is highly non-linear
and requires statistical feature extraction methods to
extract the features from the ECG signal. The Short
Time Fourier Transform (SFT) and Fast Fourier
Transform (FFT) has been used by numerous
researchers for ECG signal processing. However,
there are some limitations are still existing in FFT
due to its dearth of response of time information's
while analyzing it in frequency domain and vice versa
[5]. In order to come over this limitation, the time
frequency domain (TFD) analysis is required, which
is carried out by using the discrete wavelet transform,
empirical mode decomposition (EMD), Smoothed
Pseudo Wigner-Ville Distribution (SPWVD) and etc,
are used to derive the time and frequency domain
features. In this, we have used DWT with "db4" one
of the wavelet function which calculates statistical
features. Because the wavelet characteristics of “db4”
looks like ECG signal and it has been used by some
researchers on ECG signal analysis and for de-
noising[10]. In order to extract features, some
statistical properties are used.
4.1 Features
As signal is decomposed to determine its features.
Statistical and non- statistical properties such as
energy, mean, median, entropy, standard deviation,
skewness, kurtosis, covariance is calculated from the
coefficients which is generated after decomposition to
create the feature set [4]. All algorithm are developed
in the MATLAB environment with wavelet. The
coefficients Cik at 4th level of decomposition are used
to calculate the features by using following standard
equations;
Energy, = (1)
Standard deviation, σi =
( − i)2 (2)
Mean I = ∑ (3)
( )
Kurtosis,KURi= (4)
( )
Skewness, SKi= (5)
Entropy, ENi=∑ log ( ) (6)
Median, MEDi=∑ (7)
Covariance, COV= E[(Cik – ik)…(Cn– n )] (8)
Variance, VAR = (σi2 ) (9)
The coefficients Cikis the decomposition coefficient.
Here i=1,2,…,l is the node number at 4th level of
decomposition. N is the number of coefficients at
International Journal Of Electrical, Electronics And Data Communication, ISSN: 2320-2084 Volume-4, Issue-9, Sep.-2016

each decomposed level [4]. Energy, standard standard deviation plot

deviation, entropy, median are computed for each 1000
node at 4th level of decomposition tree. Total nine 500
properties, as defined in above equations, are used for
0
creating the feature set. Thus at 4th level of 0 2 4 6 8 10 12 14 16
5
decomposition, there are 16 coefficients (24=16) x 10 covariance plot
4
coefficients. This is called wavelet tree.
2

0
0 2 4 6 8 10 12 14 16
(e)
11
x 10 entropy plot
1
0
-1
0 2 4 6 8 10 12 14 16
mean plot
1000
0
-1000
0 2 4 6 8 10 12 14 16
5
x 10 variance plot
2
Fig1. Wavelet tree 1
0
0 2 4 6 8 10 12 14 16
V. RESULTS
(f)
Original and noisy signal Figure2: signal of record 105 (normal) database and its
1400
characteristics
Amplitude

1200
Original and noisy signal
1300
1000
1200
Amplitude

800 1100
0 500 1000 1500 2000 2500 3000 3500 4000
No. of samples
1000
(a) 900
Denoised signal 0 500 1000 1500 2000 2500 3000 3500 4000
3000 No. of samples

(a)
A m plitude

2000 Denoised signal

2000
1000
Amplitude

1500
0
0 500 1000 1500 2000 2500 3000 3500 4000 1000
No. of samples
500
(b) 0 500 1000 1500 2000 2500 3000 3500 4000
Wavelet packet decomposition No. of samples
180.00
(b)
WWavelet packet indices - Frequential order

23
24
26
Wavelet packet decomposition
WWav elet pack et indic es - Frequential order

144.00 25 180.00
23
29 162.00 24
30 26
108.00 28 144.00
Freq (Hz)

25
27 29
126.00
19 30
72.00 20 108.00 28
Freq (Hz)

22 27
21 90.00
19
36.00 17 72.00 20
18 22
16 54.00 21
15 17
0.00 36.00
1.389 2.778 4.167 5.556 6.944 8.333 9.722 18
Time (secs) 18.00 16
15
(c) 0.00
1.389 2.778 4.167 5.556 6.944 8.333 9.722
kurtosis plot
Time (secs)
20
(c)
10 standard deviation plot
400
0
0 50 100 150 200 250
skewness plot 200
5

0 0
0 2 4 6 8 10 12 14 16
5
-5 x 10 covariance plot
0 50 100 150 200 250
2
median plot
10
1
0
0
-10 0 2 4 6 8 10 12 14 16
0 50 100 150 200 250

(d)
(d)

De-noising and Statistical Feature Extraction of the ECG Signal Using Wavelet Analysis

40
International Journal Of Electrical, Electronics And Data Communication, ISSN: 2320-2084 Volume-4, Issue-9, Sep.-2016

kurtosis plot sensitivity can be obtained using the features of these

20
signals.
10

0
0 50 100 150 200 250 REFERENCES
skewness plot
5

0
[1] AsutoshKar, Leena Das “A Technical Review on Statistical
Feature Extraction of ECG signal” ,IJCA Special Issue on
-5
0 50 100 150 200 250 “2nd National Conference- Computing, Communication
median plot and Sensor Network” CCSN, 2011
10
[2] Mohammad Ali Tinati and BehzadMozaffary, “A Wavelet
0 Packets Approach to Electrocardiograph Baseline Drift
-10 Cancellation”, Hindawi Publishing Corporation
0 50 100 150 200 250
International Journal of Biomedical Imaging Volume
2006.
(e) [3] Harimohanrai, Anurag Trivedi, Shailja Shukla, “ECG
11
x 10 entropy plot signal processing for abnormalities using multi-resolution
1
0 wavelet transform and artificial neural network classifier”,
-1 Deptt of Electrical Engg, Jabalpur Engineering College,
0 2 4 6 8 10 12 14 16
mean plot Jabalpur, MP, India.
1000
[4] RohtashDhiman, J S Saini, Priyanka,”genetic algorithm
0
-1000 tuned expert model for detection of epileptic seizures from
0 2 4 6 8 10 12 14 16
5
variance plot
EEG signatures”, Electrical EnggDeptt,
x 10
2 Deenbandhuchhotu Ram University of science &
1
0
technology, Murthal, District Sonipat, Haryana, India.
0 2 4 6 8 10 12 14 16 [5] P karthikeyan, M Murugappan, S Yaacob, “ECG Signals
Based Mental Stress Assessment Using Wavelet
(f) Transform”, 2011 IEEE International Conference on
Figure 3: signal of record 118(abnormal) database and its Control System, Computing and Engineering.
characteristics [6] Eduardo José da S. Luz, William Robson Schwartz,
Guillermo Cámara Chávez, David Menotti, “ECG-based
CONCLUSION Heartbeat Classification for Arrhythmia Detection: A
Survey”, Computer Methods and Programs in
Biomedicine.
In this paper, we have extracted nine features of the [7] ShaoyuQiao, Ping Zhou “Wavelet and wavelet packet
signal which determines the characteristics of the transform analysis in the ECG signals ofAtrial
signal. These features tells the difference between the Fibrillation” 2007 IEEE/ICME international conference
two signals. As these features gives information on complex medical engineering.
[8] Fatin A. ElhajNaomieSalimArief R. Harris Tan
which helps us to find the condition of the signal i.e TianSweeTaqwa Ahmed “Arrhythmia Recognition and
normal or abnormal. These features have the different Classification using Combined Linear and Nonlinear
values in both the signals which tells us about the Features of ECG Signals”, Computer Methods and
condition. So these non-statistical have great Programs in Biomedicine (2016).
[9] R. Rodrígueza, A. Mexicanob, J. Bilac, S. Cervantesd, R.
significance in this biomedical research. Ponceb “Feature extraction of electrocardiogram signals
by applying adaptive threshold and principal component
FUTURE SCOPE analysis”, journal of Applied Research and Technology 13
(2015) 261-269.
[10] G. Ranganathan, et al., "ECG Signal Processing Using
The data is taken from the MIT database for analysis. Dyadic Wavelet for MentalStress Assessment," in
So after extracting the features of the signal they can Bioinformatics and Biomedical Engineering (iCBBE),
be classified using non-linear classifier like SVM, 2010 4th International Conference on, 2010, pp. 1-4.
KNN. Two different types of signal are taken normal [11] B. Jagadiswara Rao, A. Usharani “Enhancement of ECG
Signal by using Digital FIR Filter”,Department of
and abnormal in their conditions. These signals can Electronics and Communication, Vignan‟s Institute of
be classified using classifier and accuracy and Information Technology, Visakhapatnam, India.



De-noising and Statistical Feature Extraction of the ECG Signal Using Wavelet Analysis

41