Electrocardiogram Beat Classification using Discrete Wavelet Transform, Higher Order Statistics and Multivariate Analysis

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072
© 2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 854
Electrocardiogram beat classification using Discrete Wavelet Transform,
higher order statistics and multivariate analysis
Thripurna Thatipelli1, Padmavathi Kora2
1Assistant Professor, Department of ECE, GRIET, Hyderabad, Telangana, India
2Associate Professor, Department of ECE, GRIET, Hyderabad, Telangana, India
--------------------------------------------------------------------------------------***--------------------------------------------------------------------------------------
Abstract—Arrhythmia is a cardiovascular condition
caused by abnormal activities of the heart,
Electrocardiogram(ECG) is used to detect heart
irregularities. The development of many existing systems has
depended on linear features such as Discrete Wavelet
Transform(DWT) on ECG data which accomplish high
performance on noise-free data. However, higher order
statistics and multivariate analysis illustrate the ECG signal
more efficiently and achieve good performance under noisy
conditions. This paper investigates the representation of
DWT and Higher order statistics and multivariate analysis
to improve the classification of ECG data. Five types of beat
classes of arrhythmia as recommended by the Association
for Advancement of Medical Instrumentation are analyzed,:
non-ectopic beats (N), supraventricular ectopic beats (S),
ventricular ectopic beats (V), fusion beats (F) and
unclassifiable and paced beats (U). The representation
capability of nonlinear features such as high order
statistics(HOS) and cumulants and nonlinear feature
reduction methods such as independent component
analysis(ICA) are collective with linear features, namely, the
principal component analysis(PCA) of discrete wavelet
transform coefficients. The obtained features are applied to
the classifier, namely, the support vector machine(SVM) .
The proposed method is able to perform ECG beat
classification using DWT,PCA,HOS and ICA with high
accuracy 98.91% percent.
Keywords—ECG, DWT,HOS,ICA,PCA.
1. Introduction
An electrocardiogram (ECG) is the non-invasive method
used to identify arrhythmias or heart abnormalities.
Cardiovascular disease (CVD) has turned out to the main
origins of death in the World. The American Heart
Association expressed that, in 2006, more than 70 million
individuals around the globe were determined to have
CVD. The basic reasons for CVD are hypertension,
lacking physical exercise, ineffectively adjusted eating
regimen, smoking and unusual glucose levels.
Due to the existence of noise and the abnormality of the
heartbeat, physicians face complications in the analysis of
Arrhythmias[2]. Moreover, visual inspection alone may
lead to misdiagnosis or irrelevant detection of
arrhythmias. Therefore, the computer aided analysis of
ECG data supports physicians to proficiently detect
arrhythmia.
There are three main processes in an ECG arrhythmia
detection system, namely, feature extraction, feature
selection, and classifier construction. The ECG beat
classification [1]as per ANSI/AAMI EC57:1998 standard
database shown in Table -1.Feature extraction transforms
the input data into a set of features and plays an important
role in detecting most heart diseases.
The proposed method uses Principal Component Analysis
with Wavelet Transform coefficients and Higher order
statistics with Independent Component Analysis to achieve
high accuracy .
Table - 1: MIT-BIH arrhythmia beats classification per
ANSI/AAMI EC57:1998 standard database .

2. Pre Processing
The proposed system adopts different methods for
Preprocessing, Feature extraction. The excellence of the ECG
Beat classification depends on the accuracy of the
detection of each cardiac cycle. Proposed method MIT –
BIH arrhythmia database[5]. In this paper , the pre
processing module is decomposed into three components.
i. De noise
ii. QRS Detection
iii. Segmentation
The block diagram of proposed System is shown in fig 1.
Pre processing
ECG De noise(DWT)
Signal QRS detection
Segmentation
Feature Extraction
SVM DWT+PCA
HOS+ICA
Fig- 1: Block diagram of proposed method
2.1 De-noising
DWT is capable in analyzing non-stationary signals.
Wavelet transform method for de noising of ECG signal,
Selection of appropriate wavelet and number of
decomposition level is very significant in investigation of
signals.
The Daubechies D6 (db6) wavelet basis function is
used to de noise the data, with the ECG signals
decomposed to nine levels[3]. The inverse wavelet
transform is obtained from the third to the ninth level
detail sub-bands to achieve the de noised and smoothed
ECG signal.
2.2 QRS complex detection & Segmentation
QRS complex detected on denoised ECG signal by Pan-
Tompkins algorithm. It describes the slope, amplitude and
width. The algorithm classified to three steps. First step, the
low pass and high pass filters shapes a band pass filter,
which decreases noise in the ECG signal like muscle
noise[4]. In the second step, to recognize QRS complexes
from low-frequency ECG components for example the P and
T waves, the signal is gone through a differentiator to
highlight the high slopes.
The third step is the squaring operation, which places
stress on the higher values that are essentially present
because of QRS complexes.Smooth ECG signal is obtained
by passing squared signal through moving window integrator.
After identification of the QRS complex, 99 samples were
selected from the left side of the QRS mid-point and 100
samples behind QRS mid-point and the QRS midpoint itself
as a segment or beat of 200 samples.
3. Feature Extraction
Wavelet Transform is appropriate to analyze non
stationary signals. The feature vectors from the ECG data
set created by extracting the Wavelet transform , linear
dimensionality reduction technique PCA and Higher order
statistics cumulant features with Independent component
analysis(ICA)[8].
DWT was applied after ECG signal denoising, QRS complex
detection and segmentation, and DWT is used to extract
hidden information. The PCA was applied on both the sub-
band coefficients of the third level detail and fourth level
detail[6]. In total, 12 features (six each from the two sub-
bands) were used for subsequent pattern identification
using classifiers.

Fig -2: Five types of beat classes
(a) Smoothing signal beats using DWT
(b) Signal beats without using DWT
The ECG signal x(n) is passed through a low pass filter
h(n), and then down sampled by a factor of two to attain
the approximation coefficients at level one. The high pass
filter is resulting from the low pass filter as,
g (L-1-n)= h(n)------(1)
Fig -3:Third level detail and fourth level detail
of DWT for the five types of beats classes
where L is the length of the filter in number of points. The
detail coefficients were obtained by passing the signal
through g(n) and then down sampling by a factor of two.
The two filters h(n) and g(n) were called quadrature
mirror filters. The DWT filtering along with sub sampling
were given by,
= ------(2)
And
= -------(3)
PCA is one of the best techniques of linear dimensionality
reduction technique for extracting effective features from
high dimensions. PCA consists of following 3 phases.
Phase 1: Compute covariance matrix from the data as,
C= ( x- ) ----------(4)
where is the data matrix, and represents mean vector of x .
Phase 2: Compute the array of eigenvectors V and diagonal
matrix of eigen values D as
CV=D -----(5)

Phase 3: Sort the eigenvectors in V in descending order of
eigen values in D and project the data on these eigenvector
directions by taking the inner product between the data
matrix and the sorted eigenvector matrix as,
p= ---------(6)
PCA was computed on both sub-band coefficients of the
third level detail and fourth level detail. In total, 12
features (six each from the two sub-bands) were used for
subsequent pattern identification using SVM. Higher order
statistics removes Guassian noise[7].ICA involves a multi-
variant analysis to reduce a multi-source signal into
additive subcomponents[13].
The feature vector from ECG data is obtained by
combining linear and nonlinear features. The combined
feature vector is obtained by appending the twelve PCA of
DWT features, 16 ICA features and HOS cumulant
features[12].
4. Classification
4.1 Support Vector Machines
The SVM classifier for a solitary layer can manage
classification problems because of its ability for
generalization. It changes over the input vector patterns to
higher measurement highlight space through some
nonlinear mapping and acquires an ideal isolating hyper-
plane which is worked to isolate two classes of tests[9]. In
specific, the SVM classifier demonstrates an ideal
speculation ability when utilizing the maximal margin
principle.
A target work is defined in view of the separations of the
class isolating hyper-plane and the enhancement
procedure is done .Diverse kernel transformations are
utilized to delineate information into high dimensional
functions, for example, the quadratic, polynomial and
radial basis function (RBF). The performance of the SVM
can be influenced by the hyper-parameter (C parameter
and the kernel parameter), as these parameters decide the
quantity of support vectors and the maximization margin
of the SVM.
5. Results
The ECG beat classification using DWT,PCA and Higher
order statistics and ICA with the MIT-BIH arrhythmia
database. Feature vectors formed with combination of
linear and nonlinear methods consisted of 28 features. The
feature vectors inputted to the SVM based on Gaussian
kernel . The corresponding overall accuracy of the
proposed model were 98.91% respectively.
TABLE-2: Overall accuracy for all ECG beat classes
Featu
res
Overall
Accuracy
(%)
N S V F U
PCA-
DWT
88.04% 91.01
%
94.19
%
92.05
%
97.59
%
94.5
%
ICA-
HOS
97.83% 98.85
%
99.02
%
97.83
%
98.90
%
96.6
%
PCA-
DWT
+
HOS-
ICA
98.91% 98.91
%
100
%
98.91
%
100
%
100
%
For Calculating accuracy two parameters sensitivity and
specificity are calculated using the equations 7 and 8
Specificity = --------
-(7)
Sensitivity =
---------(8)
Accuracy = -------
--(9)
TABLE - 3: Classification Results
Classifier Sensitivity(%) Specificity(%) Accuracy(%)
SVM 98.91% 97.85% 98.91%
NN 98.51% 98.62% 98.7%
6. Conclusion
An effective ECG Beat classification system that consists of
the combination of PCA of DWT, ICA and HOS feature
extraction methods and classifier SVM –Gaussian kernel is
proposed. The experiments on the proposed system
showed that the combination of PCA-DWT, ICA and HOS
feature extraction methods with SVM-Guassian kernal and
provided average accuracy is 98.91%.

REFERENCES
[1] N. Maglaveras, T. Stamkapoulos, K. Diamantaras, C.
Pappas, M.Strintzis, “ECG pattern recognition and
classification using nonlinear transformations and neural
networks: A review”, Int. J.Med. Inform, vol 52, pp. 191–
208, 1998.
[2] . Padmavathi Kora, K. Sri Rama Krishna; ECG Based
Heart Arrhythmia Detection Using Wavelet Coherence and
Bat Algorithm" Sensing and Imaging, Springer, Vol
17, no. 1, Jun2016.
[3] S. Osowski, T.H. Linh, “ECG beat recognition using fuzzy
hybrid neural network”, IEEE Trans. Biomed. Eng, vol. 48,
pp. 1265–1271, 2001.
[4] Padmavathi Kora, and Sri Ramakrishna Kalva,
”Improved Bat algorithm for the detection of myocardial
infarction,”SpringerPlus, Springer, vol 4, no. 1, pp. 1-18,
Nov 2015.
[5] P. de Chazal, B. G. Celler, R. B. Rei, “Using Wavelet
Coefficients for the Classification of the
Electrocardiogram”, Proceedings of the 22nd Annual EMBS
International Conference, July 23-28,2000, Chicago IL.
[6] S.Banerjee, M.Mitra “Application of Cross Wavelet
Transform for ECG Pattern Analysis and Classification”,
IEEE transaction on
Instrumentation and Measurement, Vol.63 ,No.2, pp.326-
333,Feb 2014
[7] Padmavathi Kora, K. Sri Rama Krishna, ”Hybrid Firefly
and Particle Swarm Optimization algorithm for the
detection of Bundle Branch Block,” International Journal of
Cardiovascular Academy, Elsevier, Dec 2015.
[8] “The MIT-BIH Arrhythmia Database,”
http://physionet.ph.biu.ac.il/physiobank/database/mitdb
/
[9] Padmavathi Kora, K. Sri Rama Krishna, ”Adaptive
Bacterial Forging Optimization for the detection of Bundle
Branch Block,” Egyptian Informatics Journal, Elsevier, Vol
17, June 2016.
[10] R.J. Martis, U.R. Acharya, K.M. Mandana, A.K. Ray, C.
Chakraborty, Application of principal component analysis
to ECG signals for automated diagnosis of cardiac health,
Expert Systems with Applications 39 (14) (2012) 11792–
11800.
[11] Padmavathi Kora, and Sri Rama Krishna, “Hybrid
Bacterial Foraging and Particle Swarm Optimization for
detecting Bundle Branch Block,” SpringerPlus, Springer,
vol 4, no 1, 481,Sep 2015.
[12] R.J. Martis, C. Chakraborty, A.K. Ray, An integrated
ECG feature extraction scheme using PCA and wavelet
transform, in: IEEE INDICON-2009, 2009, ISBN: 978-1-
4244-4859-3/09.
[13]R.O. Duda, P.E. Hart, D.G. Stork, Pattern Classification,
2nd ed., Wiley, New York, 2001. A. Hyvärinen, E. Oja,
Independent component analysis: algorithms and
applications, Neural Networks 13 (June (4–5)) (2000)
411–430

Electrocardiogram Beat Classification using Discrete Wavelet Transform, Higher Order Statistics and Multivariate Analysis

More Related Content

Similar to Electrocardiogram Beat Classification using Discrete Wavelet Transform, Higher Order Statistics and Multivariate Analysis (20)

More from IRJET Journal (20)

Recently uploaded (20)

Electrocardiogram Beat Classification using Discrete Wavelet Transform, Higher Order Statistics and Multivariate Analysis