See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/288750153

Fuzzy C-means clustering, neural network, WT, and HRV for classification of
cardiac arrhythmia

Article in ARPN Journal of Engineering and Applied Sciences · October 2011

CITATIONS READS

33 321

3 authors:

Adel Dallali Abdennaceur Kachouri

Faculty of Sciences of Gafsa. University of Sfax
17 PUBLICATIONS 181 CITATIONS 348 PUBLICATIONS 3,501 CITATIONS

SEE PROFILE SEE PROFILE

Mounir Samet
University of Sfax
206 PUBLICATIONS 2,389 CITATIONS

SEE PROFILE

All content following this page was uploaded by Adel Dallali on 05 March 2020.

The user has requested enhancement of the downloaded file.

 VOL. 6, NO. 10, OCTOBER 2011 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2011 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

FUZZY C-MEANS CLUSTERING, NEURAL NETWORK, WT, AND HRV

FOR CLASSIFICATION OF CARDIAC ARRHYTHMIA
A. Dallali, A. Kachouri and M. Samet
Laboratory of Electronics and Technology of Information (LETI), National School of Engineers of Sfax, Sfax, Tunisia
E-Mail: [email protected]

ABSTRACT
The classification of the electrocardiogram registration into different pathologies diseases devises is a complex
pattern recognition task. The traditional methods of diagnosis and classification present some inconveniences; seen that the
precision of credit note one diagnosis exact depends on the cardiologist experience and the rate concentration. Due to the
high mortality rate of heart diseases, early detection and precise discrimination of ECG arrhythmia is essential for the
treatment of patients. In this paper, a new cardiology system has been proposed for diagnosis, consultation, and treatment.
The aim of this method is to help to practitioner doctor. During the recording of ECG signal, different forms of noise can
be superimposed in the useful signal. This model consists of three subsystems. The first subsystem divides into suppression
of base line and filtering the ECG recorded from different forms of noise that can be superimposed in the useful signal. The
second subsystem realizes the extraction of RR interval using wavelet transform, and pre-classification based on FCMC
technique. The third subsystem classifies the output clusters centers of the second using artificial neural network (ANN). In
addition, FCMC-HRV is a new method proposed for classification of ECG. In this study, a combined classification system
has been designed using fuzzy c-means clustering (FCMC) algorithm and neural networks. FCMC was used to improve
performance of neural networks which was obtained very high performance accuracy to classify RR intervals of ECG
signals. The ECG signals taken from MIT-BIH ECG database are used in training and testing data to classify four different
arrhythmias (Atrial Fibrillation Termination). The test results suggest that HRV-FCMCNN structure can generalize better
and is faster than other structures. Correct classification rate was found as 99.99% using proposed combination of Fuzzy C-
Means Clustering Neural Networks (FCMCNN) method.

Keywords: cardiac arrhythmia, fuzzy C-means clustering, WT, HRV, MCN, neural network, classification.

1. INTRODUCTION
Electrocardiography deals with the electrical
activity of the central of the blood circulatory system, i.e.,
the heart. Monitored by placing sensors at the limb
extremities of the subject, electrocardiogram (ECG) is a
record of the origin and the propagation of the electrical
potential through cardiac muscles [1]. Thus, ECG offers
cardiologists with useful information about the rhythm and
functioning of the heart. Therefore, its analysis represents
an efficient way to detect and treat different kinds of
cardiac diseases. [2].
The state of cardiac heart is generally reflected in Figure-1. Structure of the ECG signal.
the shape of ECG waveform and heart rate. A typical
structure of the ECG signal is shown in Figure-1. It may Up to now; numerous algorithms have been
contain important pointers to the nature of diseases developed in the literature for the recognition and
afflicting the heart. Early and quick detection and classification of ECG signal. Some of them use time, some
classification of ECG arrhythmia are important, especially use frequency, and other use frequency-level domain for
for the treatment of patients in the intensive care unit. In depiction. The ECG waveforms may be different for the
the last four decades, computer-aided diagnostic (CAD) same patient to such extent that they are unlike each other
systems have been applied to the classification of the ECG and at the same time alike for different types of beats [1].
resulting in several techniques [1-3]. This is mainly due to Artificial neural network [4-5], and fuzzy-logic based
the fact that ECG signal provides cardiologists with useful technique [6], were also employed to exploit their natural
and important information concerning the dysfunctions ability in pattern recognition task for successful
and physical condition of human heart. In designing of classification of ECG beats. The hybrid system of neural
CAD system, the most important is the integration of network and fuzzy logic has been widely accepted for
suitable features extractor and pattern classifier such that pattern recognition tasks. Yu et al., have implemented the
they can operate in coordination to make an effective and integration of independent component analysis and neural
efficient system [2]. network classifier (ICA - NN) along with R-R intervals to
discriminate eight types of ECG beats [7].

112
 VOL. 6, NO. 10, OCTOBER 2011 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2011 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

In [8], Ozbay et al., had combined principal shifted from zero line. The baseline of the ECG signal was
component analysis with neural network (PCA -NN) and adjusted at zero line by subtracting the median of the ECG
compared it with wavelet transform technique for ECG signal.
signal classification. In [9], T.M. Nazmy had combined
ICA and hybrid system (ICA - ANFIS) for ECG signal 2. LITERATURE STUDY
classification. In the literature survey, several methods have
In this paper, the approach to ECG beat been proposed for the automatic classification of ECG
classification presented thorough experimental exploration signals.
of the FCMCNN capabilities for ECG classification. Among the most recently published works are
Further the performances of the FCMCNN approach in those presented as follows. CUIWEI Li et al., (1995)
terms of classification accuracy are evaluated. In our showed that it is easy with wavelets transform
previous works, we had showed clearly that neural decomposition to characterize the ECG waves. The
network and fuzzy system with feature extraction methods difference between QRS, P and T waves, noise, baseline
had better performances than the traditional clustering and drift and interference were recognized. Using a raised
statistical methods. cosine wavelet transform, Khadra et al., (1997) have
undertaken a preliminary investigation of three
arrhythmias-ventricular fibrillation (VF), ventricular
tachycardia (VT) and atrial fibrillation (AF). They
developed an algorithm based on the scale-dependent
energy content of the wavelet decomposition to classify
the arrhythmias, distinguishing them from each other and
normal sinus rhythm. Again this study involved low
numbers of data: 13 VF, 12 VT, 13 AF and 8 normal sinus
rhythms. A. Kachouri, M. Ben Messaoud and A.Dallali
(2003) compared wavelet transform for recognizing
cardiac patterns. The choices of the wavelet family as well
as the selection of the analyzing function into these
families have been discussed; this study comprises four
wavelet families: Daubechies, coiflet, haar, and
symmetric. The choice of level decomposition has the
same importance as the selection of wavelet function.
Using a support vector machines methodology,
MG Tsipouras and al (2004), had classify cardiac
arrhythmia. The results indicate high classification ability
of the proposed method. The largest misclassification rates
are between the PVC and NSR classes. This is due to the
fact that PVC and NSR classes present high similarity.
The method utilizes only the intervals of the ECG;
therefore it is faster and more unaffected by the presence
of noise than other proposed methods but also classifies a
relatively small number of cardiac beats. (2007), Ceylan,
R. and Ozbay, Y. proposed a classification of ECG
arrhythmias using neural network architecture based on
techniques of FCM, PCA and WT. Accurate and
Figure-2. Block diagram of proposed computationally efficient means of classifying
arrhythmia classifier. electrocardiography (ECG) arrhythmias has been the
subject of considerable research effort in recent years. This
All the samples must be normalized in order to study presents a comparative study of the classification
have the features at the same level. ECG signals can be accuracy of ECG signals using a well-known neural
contaminated with several kinds of noise, such as power network architecture named multi-layered perceptron
line interference (A/C), baseline wandering (BW), and (MLP) with backpropagation training algorithm, and a
electromyography noise (EMG), which can affect the new fuzzy clustering NN architecture (FCNN) for early
extraction of parameters used for classification, so we diagnosis. The results suggest that proposed FCNN
want to filter the signal. The unwanted noise of the signal architecture can generalize better than ordinary MLP
must be removed. ECG was filtered using Low pass filter architecture and also learn better and faster. The advantage
and high pass filter. The pre-treatment of ECG signals of the proposed structure is a result of decreasing the
imposes the suppression of each perturbation signals, the number of segments by grouping similar segments in
noise high frequency electromyography and the low training data with fuzzy c-means clustering.
frequency drift. After that, the signal baseline may be

113
 VOL. 6, NO. 10, OCTOBER 2011 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2011 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

(2010), S. Karpagachelvi et al., have developed a high frequency and low frequency for wavelet
combined support vector machine and relevance vector decomposition as the level increases (j = 1… 6).
machine analysis method for the automatic classification
of electrocardiogram signals. The obtained experimental
results show that the use of the RVM approach for
classifying ECG signals on account of their superior
generalization capability as compared to traditional
classification techniques. The results confirm that the
RVM classification system substantially boosts the
generalization capability achievable with the SVM
classifier, and its robustness against the problem of limited
training beat availability, which may characterize
pathologies of rare occurrence.

3. MATERIALS AND METHODS

This paper presents using fuzzy c-means
clustering algorithm to develop the performance of neural
network classifier. In many pattern recognition
applications, the task of partitioning a pattern set can be
considered to be the result of clustering algorithms in
which the cluster prototypes are estimated from the
information of the pattern set. In many cases, it may be
impossible to obtain exact knowledge from a given pattern
set. For recognition of the ECG arrhythmias, different
methods were presented in the literature, such as the MLP
approach, LVQ. In this paper, we present the combination
of different forms of fuzzy c-means clustering, neural
network, and wavelet transform; named as WT - FCMC -
NN.

3.1 Wavelet transform

Physiological signal used for diagnosis are
frequently characterized by a non-stationary time
behavior. For such patterns, time and frequency
representations are desirable. The ECG signals are
considered as representative signals of cardiac physiology,
which are useful in diagnosing cardiac disorders. The
wavelet transform (WT) provides very general and
powerful techniques, which can be applied to many tasks
in signal processing. WT can represent signals in different
resolutions. The most important application is the ability
to compute and manipulate data in compressed
parameters. Thus, the ECG records can be compressed
into a few useful parameters. These parameters can be
used for recognition and diagnosis. The most favorable
choice of types of wavelet functions and level for pre-
processing is problem dependent. The smoothing feature
of the Daubechies wavelet of order 3 made it more
suitable to detect variation on the ECG signals [8].
Therefore, the wavelet coefficients were computed using
the Daubechies wavelet of level 3 in the present work.

(1)
where Dj,k represents the detailed signal at j level. Note
that j controls the dilation or contraction of the scale
function and k denotes the position of the wavelet
function , and n represents the sample number of the x Figure-3. Top: Normal signal (a), and histogram of
(n). The frequency spectrum of the signal is classified into approximation after wavelet transform (b). Bottom:
Same (c), but for signal with atrial fibrillation (d).

114
 VOL. 6, NO. 10, OCTOBER 2011 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2011 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

3.2 The fuzzy c-means clustering This iteration will stop when
The FCM algorithm has successfully been error , where is a termination
applied to a wide variety of clustering problems. The FCM criterion between 0 and 1; whereas k are the iteration
algorithm attempts to partition a finite collection of steps. This procedure converges to a local minimum or a
elements X = {x1, x2 … xN} Rh where N represents the saddle point of Jm. The algorithm is composed of the
number of data vectors and h the dimension of each data following steps:
vector, into a collection of C fuzzy clusters. C - Partition
of X constitutes sets of (c. N) {uij} member ship values The steps are as follows.
can be conveniently arranged as a (c. N) matrix u = [uij].
The objective of fuzzy clustering is to find the optimum I. Initialize the number of clusters (c), weighting
member ship matrix U. the most widely used objective exponent (m), iteration limit, termination criterion
function for fuzzy clustering is the weight within - groups ( ) and matrix, U (0).
sum of squared errors Jm, which is used to define the II. Guess initial position of cluster centers.
following constrained optimization problem [13].
III. At k step calculate the center vectors

IV. Update U(k) to U(k+1) (2)

Where 1≤m≤ .., i.e., m is any real number greater than 1,
uij is the degree of member ship of xi in the cluster j, xi is
the ith component of d-dimensional measured data, cj is the
d - dimension center of the cluster, and is any If , then stop; otherwise to step (i)
norm expressing the similarity between any measured data
and the center. Fuzzy partition is carried out through an
iterative optimization of the objective function shown
above, with the update of member ship uij. 3.3 Artificial neural network (ANN)
Artificial Neural Network is biologically inspired
network that are suitable for classification of biomedical
signal. A combination of wavelets transform, FCMC and
NNs is proposed to classify cardiac arrhythmias. The
precision of classification results of the anomalies depends
and the cluster cj by: on the number of parameters selected; the number of
neurons of input layer is equals to the numbers of FCMC
clusters centers.

Figure-4. Structure of WT-FCMCNN classifier.

In this study, a three-layered feed-forward (MLP) b) Presentation of input and desired outputs: Present
neural network architecture was utilized and trained with the input vector x(1), x(2),…,x(N) and corresponding
the error back propagation algorithm with Log-sigmoid desired response d(1),d(2),…,d(N), one pair at a time,
activation function that using the generalized back where N is the number of training patterns.
propagation for training which minimize the squared error
c) Calculation of actual outputs: Use Equation (8) to
between the desired outputs and the actual outputs of the
calculate the output signals y1, y2,…, yN
ANNs. The desired output is being a real number in the
interval [0-1]. The Back propagation algorithm is
(8)
described step by step as [18]:
a) Initialization: Set all the weights and biases to small d) Adaptation of weights (wij) and biases (bi):
real random values. (9)

115
 VOL. 6, NO. 10, OCTOBER 2011 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2011 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

(10) training and testing of the ANNs classifier and the results
obtained are tabulated in Table-2. The parameters
And extracted are used as inputs to system and the
classification is done. The output of the classifier is a
graphical representation.
(11)
4. TEST RESULTS OF NUMERICAL
EXPERIMENTS
where xj(n)= output of node j at iteration n, l is layer, k is
the number of output nodes of neural network, M is output 4.1 Dataset description
layer, φ is activation function. The learning rate is The experiment conducted on the basis of ECG
demonstrated by µ. It may be noted here that a large value data from the MIT-BIH arrhythmia database. ECG signal
of the learning rate may lead to faster convergence but is filtered with low pass and high pass filters. After that
may also result in oscillation. With the purpose of signal is preprocessed to detect RR intervals; which is
achieving faster convergence with minimum oscillation, a calculated from the location of the R points, of ECG
momentum term may be added to the basic weight signals which were sampled at 360 Hz. The detailed
updating equation. The training of the artificial neural numbers of training and test beats are reported for each
network ends if the sum of the square errors for all class in Table-1.
segments is less than 0.01. The number of data set used for

Table-1.
Numbers of training and test beats used in the experiments.
Type MIT - BIH data base Training file Testing file
1 n01, n02, n03, n04 20 18
2 S01, s02, s03, s04 20 18
3 A01, a02, a03, a04 20 18
4 B01, b02, b03, b04 20 18
Total 80 72

4.2 Test results Table-2. Pre-classification results for each arrhythmia

In this paper, a fuzzy c-means clustering neural in test.
network was proposed to diagnose an abnormal heart beat
Arrhythmia types of Number FCMC
on electrocardiogram. The improved FCMCNN was
t t tt fb t
obtained by combination of fuzzy clustering layer and RMC
MCN
neural network classifier (Figure-2) we proposed an (%)
approach that the number of HRV segment in original 1 20 0 0
training set was reduced by fuzzy c-means clustering
algorithm and process of reducing was performed on each 2 20 0 0
arrhythmia type individually. 3 20 0 0
Training max error was obtained as 1.3878e-016.
4 20 1 5
Mis - Classification Number (MCN) represents number of
misclassification ECG in testing. Rate of misclassification Total 80 1 1.25
(RMC) is calculated using: Average test error (%) 0.0118

(4) The performance of WT - FCMC - NN technique

is depicted as shown in Figure-6. It is observed that the %
of recognition rates average is 99.99%.
The obtained results can be seen in Table-2. As it
can be seen in this table, we found training and test errors
calculated from tables according to the equations given in
[8, 9].

116
 VOL. 6, NO. 10, OCTOBER 2011 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2011 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

réponse réelle: Sortie obtenue aprés apprentissage de réseau1001iterations

Obtained Output
Wanted Output
0.4

0.3

0.2

0.1

-0.1
2 4 6 8 10 12 14 16 18

Figure-5. Classification results for different arrhythmias.

-17
x 10 erreur d'apprentissage d'apres1001iterations
20

15

10

2 4 6 8 10 12 14 16 18

Figure-6. Error of classification for different arrhythmias.

5. CONCLUSIONS FCMCNN was to achieve more optimum cluster centers

In this piece of work, a novel ECG beat locations and to reduce the time of training of the neural
classification system using FCMC and NN are proposed network classifier.
and applied to MIT/BIH data base. The WT-FCMCNN This technique is obtained by incorporating the
was proposed and developed to classify preprocessing method of ECG signal, extraction of RR
electrocardiography signals. The wavelet transforms and intervals, fuzzy c-means clustering method, and neural
HRV model parameters have been used for the features network technique and combining their advantages for the
selection. For the conventional FCMC all patterns in the classification of ECG arrhythmias. So, it can be said that
pattern space are assigned membership values, which are the structure, which is a widely beneficial structure than
based on the Euclidean distance between the patterns to conventional WT - NN to recognize and classify ECG
each cluster. The cluster centers obtained by FCMC are signals, is obtained.
classified by neural network. The aim in developing WT -

117
 VOL. 6, NO. 10, OCTOBER 2011 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2011 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

The results obtained confirm that the NN [10] M. I. Owis, A. B. M. Youssef and Y. M. Kadah. 2002.
classification system substantially boosts the Characterisation of electrocardiogram signals based
generalization capability achievable with the FCMC on blind source separation. Medical and Biological
classifier. It can also be seen that NN accomplishes better Engineering and Computing, Cairo, Egypt.
and more balanced classification for individual categories
as well in very less training time comparative to FCMC. [11] Yu S.N. and Chou K.T. 2008. Integration of
independent component analysis and neural networks
REFERENCES for ECG beat classification. Expert Systems with
Applications. 34: 2841-2846.
[1] Owis M. I., Youssef A.-B. M. and Kadah Y. M. 2002.
Characterization of ECG signals based on blind source [12] Guler I. and Ubeyli E. D. 2005. ECG beat classifier
separation. Medical and Biological Engineering and designed by combined neural network model. Pattern
Computing. 40: 557-564. Recognition. 38: 199-208.

[2] Rahime Ceylan, Yuksel Ozbay and Bekir Karlik. [13] Dipti Patra and Smita Pradhan. 2009. Integration of
2009. A novel approach for classification of ECG FCM, PCA and Neural networks for classification of
arrhythmias: Type-2 fuzzy clustering neural network. ECG Arrhythmias. IAENG International Journal of
Expert Systems with Applications. pp. 6721-6726. Computer Science, IJCS_36_3_05, 01 August. 36:3.

[3] Meau Y.P., Ibrahim F., Naroinasamy S.A.L. and [14] E. M. Tamil, N. H. Kamarudin, R. Salleh, M. Yamani
Omar R. 2006. Intelligent Classification of Idna Idris, M. N. Noorzaily and A. M. Tamil. 2008.
Electrocardiogram (ECG) Signal Using Extended Heartbeat electrocardiogram (ECG) signal feature
Kalman Filter (EKF) Based Neuro Fuzzy System. extraction using discrete wavelet transforms (DWT).
Elsevier Science Computer Methods and Programs in In: Proceedings of CSPA. pp. 1112-1117.
Biomedicine. 82: 157-168.
[15] B.U. Kohler, C. Hennig and R. Orglmeister. 2002.
[4] A. Kachouri, M. Ben Messaoud and A. Dallali. 2003. The principles of software QRS detection. IEEE
Wavelet based on Electrocardiogram Signal Analyses Engineering in Medicine and Biology Magazine.
for Classification and Diagnosis by Neural Networks. 21(1): 42-57. January-February.
2nd International Conference on Signals Systems
Decision and Information Technology SSD’03. March [16] Yu S. N. and Chen Y. H. 2007. Electrocardiogram
26-28, Sousse, Tunisia. beat classification based on wavelet transformation
and probabilistic neural network. Pattern Recognition
[5] S.Y. Foo, G. Stuart, B. Harvey and A. Meyer-Baese. Letters. 28: 1142-1150.
2002. Neural network-based ECG pattern recognition.
Eng. Appl. Artif. Intell. 15: 253-260. [17] B. Anuradha and V.C. Veera Reddy June 2008. ANN
for Classification of Cardiac Arrhythmias. ARPN
[6] Hu Y.H., Palreddy S. and Tompkins W. 1997. A Journal of Engineering and Applied Sciences. 3(3): 1-
patient adaptable ECG beat classifier using a mixture 6. June.
of experts approach. IEEE Trans. Biomed. Eng. 44(9):
891-900. [18] Srinivasca K. G., Amrinder Singh, A. O. Thomas,
Venugopal K. R. and L. M Patnaik. Generic Feature
[7] Yu S. N. and Chou K.T. 2006. Combining Extraction for Classification using Fuzzy C - Means
Independent Component Analysis and Back Clustering. 0- 7803 - 9588 - 3/05/$20.00 ©2005 IEEE.
propagation Neural Network for ECG Beat
Classification. Proceedings of the 28th IEEE EMBS [19] S. Z. Mahmoodabadi, A. Ahmadian and M. D.
Annual International Conference, New York City, Abolhasani. 2005. ECG Feature Extraction using
USA, Aug.30-Sept.3. pp. 3090-3093. Daubechies Wavelets. Proceedings of the 5th IASTED
International conference on Visualization, Imaging
[8] Ceylan R. and Ozbay Y. 2007. Comparison of FCM, and Image Processing. pp. 343-348.
PCA and WT techniques for classification ECG
arrhythmias using artificial neural network. Expert
Systems with Applications. 33: 286-295.

[9] T. M. Nazmy, H. El-Messiry and B. Al-Bokhity.

2010. Adaptive Neuro-Fuzzy Inference System for
Classification of ECG Signals. Journal of Theoretical
and Applied Information Technology, JATIT. 12(2):
71-76.

118
View publication stats