Content uploaded by Oliver Faust
Author content
All content in this area was uploaded by Oliver Faust on Aug 02, 2017
Content may be subject to copyright.
Automated classification of normal and premature
ventricular contractions in electrocardiogram
signals
Nam Zheng Ning Jenny1, Oliver Faust2, Wenwei Yu3
1Department of Electronics and Computer Engineering, Ngee Ann Polytechnic, Singapore 599489
2School of Science and Engineering, Habib University, Karachi-Pakistan
3Department of Medical System Engineering, Chiba University, Chiba, 263-8522 Japan
Abstract
The objective of this project was to improve the accuracy of cardiac arrhythmia
detection by using advanced signal processing and machine learning methods.
The proposed Computer-Aided Diagnosis (CAD) system classified Premature Ven-
tricular Contraction (PVC) and normal Electrocardiogram (ECG) signals using un-
supervised machine learning algorithms. The classification quality was measured
and expressed as accuracy, Positive Predictive Value (PPV), sensitivity and speci-
ficity. The ECG records, which were used to establish the CAD system quality, were
obtained from the MIT-BIH arrhythmia database. These signals were analyzed in
four stages. The pre-processing stage standardized and improved the ECG signals
by subjecting them to Discrete Wavelet Transform (DWT) based noise reduction.
The second stage used Independent Component Analysis (ICA) for dimension re-
duction. The third stage assessed the extracted features with Student’s t-test to
determine if the features were discriminative enough to serve as classifier input. At
the last stage, two unsupervised classifiers, k-means and Fuzzy C-Means (FCM),
were used to find clusters. The proposed system achieved: accuracy = 80.94%,
sensitivity = 81.10% and specificity = 80.1%.
Keywords: Premature ventricular contraction, Electrocardiogram, Computer aided
diagnosis, Discrete wavelet transform, Independent component analysis, Fuzzy C-
means
1
1 Introduction
Cardiovascular Disease (CVD) is an important medical problem, because of its high
incidence and prevalence [1]. One of the complications of CVD, among many others,
is atrial and ventricular arrhythmias which occur due to cardiac rhythm disturbances.
Arrhythmia is a collective term for a heterogeneous group of conditions where the heart
exhibits abnormal electrical activity. Heart arrhythmia is the medical term for irregular
heartbeat or palpitations, caused by deficiencies in the heart’s electrical system [2]. Ar-
rhythmias can lead to complications, like stroke, heart failure and Alzheimer’s disease
[3]. In medical terms, the disease can be classified into one of two groups: Bradycar-
dia (abnormally slow heartbeat) and Tachycardia (abnormally fast heartbeat) [3]. The
four main types of arrhythmia include premature beats, supra-ventricular arrhythmias,
ventricular arrhythmias, and brady-arrhythmias. Arrhythmias, like Ventricular Tachy-
cardia (VT) and Atrial Flutter (AFL), are life threatening medical emergencies which
result in cardiac arrest, hemodynamic collapse and sudden cardiac death [4]. There
are many causes for arrhythmias, but the majority of them is related to CVD. Cardiac
arrhythmias can endanger life when a normal ventricular ectopic, with an underlying
heart diseases, gradually leads to VT or ventricular fibrillation. Arrhythmias can occur
at anytime, during exercise, working or even when sleeping. A person, with this dis-
ease, may just collapse. The need to understand and treat patients with this condition
has encouraged studies which aim to find diagnosing methods that pick up early signs
of potentially deadly cardiac arrhythmias [5].
Electrocardiogram (ECG) analysis, with the aim of detecting different heartbeat
types, is of major importance for the diagnosis of cardiac dysfunctions [6]. The main
feature of ECG signals is the P-QRS-T wave which contains information about the
heart condition [7]. Computer-Aided Diagnosis (CAD) systems use signal processing
algorithms to extract relevant information from ECG signals [8]. This information forms
the basis for an automated diagnosis which is established through machine learning
algorithms [9]. To ensure practicality and usefulness, the quality of such CAD systems
must be evaluated with statistical and classification tests [10]. These tests instill trust
into the systems.
In this paper, we propose an ECG based CAD system for Premature Ventricu-
lar Contraction (PVC). The proposed system uses Independent Component Analy-
sis (ICA) to extract features for k-means and Fuzzy C-Means (FCM) classification. The
results show that ICA together with FCM outperforms the combination of ICA and k-
means. The proposed system can be used effectively as a non-invasive tool for the
classification of cardiac arrhythmias. It can be applied to ECG monitoring systems and
electronic pacemakers. The CAD system can help doctors to pick up early signs of
arrhythmias quickly, which can lead to an immediate lifesaving treatment. Through its
digital nature, the system is readily deployable and very versatile, therefore it can be
used for mass screening programs. It saves time and effort for both doctors and health-
care professionals, since they do not need to go through the ECG recordings of every
patient.
The paper is organized as follows: Section 2 provides the medical background
on arrhythmias. Section 3 introduces the materials and methods used to construct
the proposed CAD system. Section 4 provides the performance assessment results.
The discussion section relates the work done in this project to studies from the wider
research community. Section 6 concludes this work.
2
2 Background
This section provides a brief overview of the electrical activity of the heart. This electri-
cal activity is captured in ECG signals which can be used for medical diagnosis [11, 12].
One disease class, that can be diagnosed with ECG signal analysis, are arrhythmias.
Section 2.2 introduces arrhythmias with a special focus an PVC.
2.1 Electrical activity of the heart
Ventricular ectopic beats are also called PVC, because the majority of these beats
occur just before the Normal Sinus Rhythm (NSR). These beats are triggered by the
heart itself and they are caused by abnormalities in the electrical conductance system
of the heart ventricles [13]. In contrast Atrial Premature Contractions (APCs) are char-
acterized by premature heartbeats originating in the atria. The Sino Atrial (SA) node,
together with the Atrioventricular (AV) node and the Purkinje fibers, execute the main
electrical activity of the heart. The SA node is located at the upper portion of the right
atrium. The AV node is located near the bottom of the right atrium, around the septum
region. The Purkinje fibers are located at the walls of the ventricles [14]. The SA node
is the natural pacemaker of the heart. Every time it triggers, electrical signals are sent
out to the atria and they stimulate the muscles to contract. At this instance, blood is be-
ing pumped to the ventricles. This electrical activity is recorded as ‘P’ wave in the ECG
[15]. Subsequently, the signal arrives at the AV node. It slows down or delays for an in-
stant so that the heart’s right and left ventricles can fill with blood. This delay is crucial,
because it allows blood to flow effectively from the atria to the ventricles. Without this
delay, both atria and ventricles would contract at the same time. This delay produces
the so called ‘PR’ segment in the ECG signal. Once the signal is released, it moves
along a pathway called the bundle of His, which is located in the walls of the heart
ventricles. The bundle of His divides into two branches and these branches spread
further out to form numerous Purkinje fibers [16]. The Purkinje fibers connect directly
to the myocardial cells and stimulate them to contract. The spread of electrical activity
through the ventricles wall shapes the ‘QRS’ complex in the ECG. The last event of
one cardiac cycle is de- and repolarization of the ventricles. This is represented by
the ‘ST’ segment and the ‘T’ wave respectively. The impulses travel from the atrium to
the center of the heart, then to the ventricles [17]. In a healthy heart, this process will
go smoothly, however interruptions or other problems with the signal distribution can
cause arrhythmias [18].
2.2 Arrhythmias
VT is characterized by a fast heart rate, which can go up to more than 100 beats per
minute [19]. It occurs when the heart beats with rapid, erratic electrical impulses. This
causes the ventricles to quiver uselessly instead of pumping blood. It usually indicates
a serious underlying heart disease and in many cases it requires prompt or emergency
treatment, because without an effective heartbeat, blood pressure plummets, cutting
off blood supply to vital organs. Ventricular Fibrillation (VF) is also one of the deadly
arrhythmias. A person with VF will collapse within seconds and soon won’t be breathing
or have a pulse [20]. During Right Bundle Branch Block (RBBB) the right ventricle is not
directly activated by the electrical impulses traveling through the right bundle branch.
Left Bundle Branch Block (LBBB) describes a similar condition for the left ventricle. In
3
Figure 1: Block diagram of the PVC detection system.
ECG signal
Noise reduction with DWT
Feature Extraction with ICA
Statistical Analysis with the t-test
Classification
Normal PVC
Atrial Fibrillation (AF), the electrical impulses from the SA node are overwhelmed by
disorganized electrical impulses which usually come from the roots of the pulmonary
veins, leading to irregular conduction of the ventricles. A Nodal Escape Beat (NEB) is a
delayed and a Nodal Premature Beat (NPB) is an early heartbeat. A Paced Beat (PB)
is induced by a pace maker. In Atrial Tachycardia (AT) the electrical impulse originates
from an ectopic atrial pacemaker rather than from the SA node .
PVC is easily seen in the time domain representation of the ECG signal [21]. Dur-
ing PVC, the electrical impulses in the ventricles travel eccentrically or backwards.
Therefore, the ECG will show wild and bizarre shapes. PVC is a common abnormality
and in most cases it does not cause a problem, unless the patient has a history of
heart diseases. Although PVC may not cause any serious problems, regular signs and
symptoms should not be ignored.
3 Materials and Methods
This section introduces the materials and methods used to construct the proposed ar-
rhythmia CAD system. The blockdiagram, in Figure 1, shows the algorithm structure
used to construct the system. The ECG signal flows through a sequence of algorithms
and at the end of this process, a classifier decides whether the signal is normal or
PVC. In the first step, the Discrete Wavelet Transform (DWT) is used for noise re-
duction. Subsequently, the ICA algorithm extracts features from these noise reduced
signals. The t-test analysis provides statistical information about the feature quality.
Only the most discriminative features were used to assemble a feature vector for the
classification step. In this classification step, the feature vectors, from all ECG signals,
form training and test sets which were used to assess the classification algorithms.
3.1 Data Used
The ECG signals, used for this project, were taken from the MIT-BIH arrhythmias
database. They were obtained from 47 subjects. The MIT-BIH arrhythmia database
contains 48 half-hour excerpts of two-channel ambulatory ECG-Holter recordings, which
4
were digitized at 360 samples per second per channel with 11-bit resolution over a
10 mV range [22]. The subjects were 25 men aged 32 to 89 years. In this analysis,
we have chosen several recordings with ventricular ectopic beats (PVC) and normal
ECG beats. Lead II data were taken, because it is the most sensitive and accurate
lead. 1000 normal ECG and 1000 PVC data samples were randomly selected to run
the test.
3.2 Denoising
Different methods have been developed for ECG signal conditioning [23]. The general
concept of denoising ECG signals in a transform domain is to find a set of coefficients
which represent the noise and a set of coefficients which represent the signal. Once
these sets are established, only the coefficients that represent the signal are used in
subsequent processing steps [24].
The wavelet transform has been widely used in signal processing applications, be-
cause this method can separate signal information from noise [25, 26]. Unlike the
Fourier transform, the wavelet transform can be used to analyze non-stationary sig-
nals. The DWT algorithm is a fast implementation of the wavelet transform. It has
a low computational complexity, thus it reduces both processing time and processing
resource requirements. In DWT, the signal is analyzed by passing it through a series
of filters with different cut-off frequencies at different scales. This technique is called
multi-resolution analysis. Multi-resolution analysis enables us to analyze the signal in
different frequency bands; therefore we can observe any movement or change in both
domains [27]. As a consequence, it offers perfect resolution in both time and frequency
domain.
One way of thinking about ECG is that the signal is composed from wavelets that are
non-stationary and non-periodic, hence we can use DWT to analyze and subsequently
to reduce noise in ECG signals [28]. Scientists demonstrated that DWT can be used
as a tool to isolate relevant properties of the waveform morphology from the noise
and other unwanted signal components, such as baseline drift, and amplitude variance
of the original ECG signal [29, 30]. They achieved high classification accuracy by
using the downsampled wavelet coefficients, that were extracted from ECG signals, as
their feature set, rather than the original waveform itself. Their findings were based
on multi-resolution techniques, which were used to find average features and specific
signal details. For example, the sharp spikes in an ECG signal, which make up the
‘QRS’ wave have a high frequency. Therefore, they can be separated from noise by
decomposing the ECG signal into different DWT levels. The multi-resolution technique
comprises of low- and high-pass filtering [31].
3.3 Independent component analysis (ICA)
ICA is a nonlinear dimension reduction method. It is good for finding underlying fea-
tures or components from multi-dimensional data. The need for these abilities is en-
countered often in signal processing applications, where the signal needs to be decom-
posed into its independent components. This decomposition process is also known
as blind source separation [32]. ICA can be applied to many medical and biological
applications like the Electroencephalogrphy (EEG) and ECG, because these signals
originate from different sources, i.e. they are composed out of multiple independent
signals [33].
5
The algorithm description starts by defining a random observation vector X.Xhas
mixtures [x1, x2, ..., xn], let sbe the source vector [s1, s2, ..., sn]and let Adenote the
weight matrix with elements aij . The ICA model assumes that the signal x, in our case
the DWT coefficients in a specific sub-band, was linearly mixed with the source signals
[34]. Another assumption, which underpins the ICA method, is that the original signals
are non-Gaussian. The entropy of a random variable is the measure of the information
that can be obtained by observing that variable. The more unlikely it is to observe a
random variable the more information is gained by observing it. A random variable
with a normal distribution has maximum entropy. That means, entropy can be used
as a measure of non-gaussianity [34]. One of the most commonly used measures for
entropy is known as Negentropy. Using the Negentropy concept, the FastICA algorithm
uses the following formula to measure non-gaussianity:
J(x)=[G(x)−G(v)]2(1)
The ICA model itself is given by:
x=As =
n
X
i=1
aisi(2)
We need to solve for the elements of Ato solve the ICA problem. The source signal is
expressed in terms of mixed signals as:
s=W x (3)
The goal of ICA is to find the unmixing matrix W; where Aand Ware the inverse of
each other, which will give Y, the best possible approximation of S:
Y=W X ≈S(4)
The ICA method assumes that the original signals are statistically independent. These
assumptions might not always hold for practical applications, but they are important for
finding the principal components s1and s2.
ICA uses centering (zero mean), whitening and dimensionality reduction as pre-
processing steps in order to simplify and reduce the complexity of the problem and
transform the observed signals into a suitable numeric format [34].
3.4 Student’s t-test
The t-test assesses whether or not the means of two groups are statistically different
from each other [35]. This analysis is appropriate whenever it is necessary to compare
the means of two groups. Like other tests, the t-test has various assumptions that must
be met to ensure the validity of the test. For the t-test there are four assumptions:
1. One variable is continuous, and the other is dichotomous.
2. The two distributions have equal variances.
3. The observations are independent.
4. The two groups are normally distributed.
6
For this project, independent two-sample t-test was used, because it is suitable for
unequal and equal sample sizes as well as unequal variances. This test is known as
Welch’s t-test, and it is particularly applicable when the two population variances are
assumed to be different (the two sample sizes may or may not be equal), hence they
must be estimated separately [36]. The tstatistic, to test whether or not the population
means are different, is calculated as:
t=¯
X1−¯
X2
s¯
X1−¯
X2
(5)
Where,
s¯
X1−¯
X2=ss2
1
n1
+s2
2
n2
(6)
Here, s2is an unbiased estimator of the variance of the two samples, niindicates the
number of participants in group iand iis equal to 1 or 2. Note that in this case s2
¯
X1−¯
X2is
not a pooled variance. For use in significance testing, the distribution of the test statistic
is approximated as an ordinary Student’s tdistribution with the degrees of freedom (df)
calculated using:
df =(s2
1/n1+s2
2/n2)2
(s2
1)2/(n1−1) + (s2
2/n2)2/(n2−1) (7)
Advantages of the t-test:
• Works well with two means.
• Excellent for ratio data.
• Detects significant differences.
• More economical, because only a small number of samples is needed.
Disadvantages of the t-test:
• Does not work well if the means are too large, because any difference may appear
insignificant.
• Subjected to practice effects – tests needed to run a few times and results will
tend to improve.
3.5 K-means clustering
This algorithm aims to classify or group objects, based on attributes or features, into
k groups, where k is a positive integer [37]. K-means clustering features in a large
number of applications, such as unsupervised learning neural networks, pattern recog-
nition, classification analysis, artificial intelligence, image processing, machine vision,
etc. [38].
The grouping is done by minimizing the squared sum of distances between a data
point and the corresponding cluster centroid. Thus, the k-means clustering has the
ability to classify data without the intervention of a trainer [38, 39].
The k-means algorithms executes three basic steps:
1. Determine the centroid coordinates.
7
2. Determine the distance of each object to the centroids.
3. Group the object based on minimum distance (find the closest centroid).
The k-means algorithm will repeat these three steps until convergence. Each object,
represented by one attribute point, is an example to the algorithm and it is assigned
automatically to a cluster.
3.6 Fuzzy C-means (FCM) clustering
FCM clustering is an unsupervised method, which is derived from fuzzy logic. It is
suitable for solving multiclass and ambiguous clustering problems. Therefore, it has
been a very important tool for image processing, especially for clustering objects in
an image. Furthermore, it is frequently used in pattern recognition [40]. FCM gives
excellent results for overlapping data sets and for some applications it is better then
the k-means algorithm. Unlike k-means, where a data point exclusively belongs to one
cluster center, in FCM a data point is assigned membership to each cluster center, as
a result data points may belong to more than one cluster center [41].
The FCM algorithm assigns membership to each data point corresponding to each
cluster center based on the distance between cluster center and data point [42]. The
closer a data point is to a particular cluster center the higher ranks the membership
of this cluster [43]. The main objective of the algorithm is to minimize the following
expression:
J(U, V ) =
N
X
i=1
C
X
j=1
µij ||xi−vj||2(8)
where, ||xi−vj|| is the Euclidean distance between the ith data point and the jth
cluster center. Fuzzy partitioning is carried out through an iterative optimization of the
objective function, shown in Equation 8. The membership update µij and the cluster
centers vjare defined in the following two equations [36]:
µij =1
PC
k=1 ||xi−cj||
||xi−ck||
2
m−1
(9)
vj=PN
i=1(µij )mxi
PN
i=1(µij )m,∀j= 1,2, ..., C (10)
where, Nis the number of data points, vjrepresents the jth cluster center, mis the
fuzziness index m∈[1,∞],µij represents the membership of ith data to jth cluster
center, dij represents the Euclidean distance between ith data and jth cluster center.
3.7 Classification assessment
Classification performance measures include True Negative (TN), True Positive (TP),
False Negative (FN), False Positive (FP), Positive Predictive Value (PPV), specificity,
sensitivity and accuracy. These measures are defined as follows:
• TP – defines the number of correct positive results, that occurred among the
entire positive sample during the test.
8
• FP – defines the number of incorrect positive results, which occurred among the
entire negative sample during the test.
• TN – defines the number of correct negative results, that occurred among the
entire negative sample during the test.
• FN – defines the number of incorrect negative results, which occurred among the
entire positive sample during the test.
Accuracy is defined as the ratio of the number of correctly classified samples in
each class to the total number of samples:
Accuracy =TP +TN
TP +FP +TN +FN ×100% (11)
Sensitivity is another name for the true positive rate. It measures the number of
actual positives that are identified correctly. It is also the probability that indicates
which test will produce a positive result when it is used on a diseased population.
Therefore, tests with a high sensitivity are reliable indicators when a result is negative.
The sensitivity is defined as:
Sensitivity =TP
TP +FN ×100% (12)
Specificity is another name for the true negative rate that measures the number of
negative results that were identified correctly. It states the probability, with which a test
will produce a negative result, when used on a disease free population. Therefore,
higher specificity results in lower false positive rate. Specificity is defined as:
Specificity =TN
TN +FP ×100% (13)
PPV specifies the number of patients who are correctly diagnosed with a positive
result. It shows the probability, with which a positive test indicates an underlying con-
dition. The PPV is defined as:
PPV =TP
TP +FP ×100% (14)
4 Results
12 lead ECG was used to measure the electrical signals from the heart. The DWT
based noise reduction played an important role in eliminating unwanted signal compo-
nents and the denoised data improved accuracy. 10 features were extracted from the
ECG signals with the ICA method. Due to the random and nonlinear characteristic of
heart signals, statistical analysis was used to assess the features. Only the features
with the highest statistical relevance were selected to form feature vectors. These fea-
ture vectors were fed to the two classifiers. Table 1 shows that all extracted features
were distinctive. However, only 8 out of 10 measurements were clinically significant
with a p-value below 0.05.
K-means and FCM classifiers were tested with the following performance param-
eter: accuracy, PPV, sensitivity and specificity. The results, presented in Table 2,
indicate an average accuracy of 80.60%, average PPV of 80.94%, average sensitivity
of 81.10% and specificity of 80.10% for the FCM algorithm. Similarly, the k-means al-
gorithm achieved an average accuracy of 60.20%, average PPV of 60.18%, average
sensitivity of 60.00%, and specificity of 60.40%.
9
Table 1: Student’s t-test results. †clinically significant features with a p-value below
0.05, which were used as input to the classifiers.
Normal PVC
Mean σMean σ p-value
ICA1 -0.2859 0.4486 -0.1998 1.3405 0.0542
ICA2 -0.1686 0.2814 0.0508 1.3780 0.0000†
ICA3 -0.2915 0.5533 -0.0903 1.2944 0.0000†
ICA4 0.2500 0.4455 -0.2588 1.2938 0.0000†
ICA5 -0.2265 0.3618 0.3677 1.3017 0.0000†
ICA6 0.8304 0.8364 -0.1623 0.8996 0.0000†
ICA7 0.4964 0.3635 0.5299 1.3672 0.4543
ICA8 -0.2492 0.3151 -1.2286 1.1927 0.0000†
ICA9 -0.8310 1.2232 -0.9252 0.7081 0.0351
ICA10 1.0588 0.9225 0.5351 1.0068 0.0000†
Table 2: Classifier performance.
No of features Accuracy PPV Sensitivity Specificity
FCM 8 80.60% 80.94% 81.10% 80.10%
K-means 8 60.20% 60.18% 60.00% 60.40%
5 Discussion
Table 3 puts the performance results of the FCM classifier alongside quality measures
from other studies. The other studies were also concerned with computer based ar-
rhythmia detection for medical diagnosis. The results we report are not the highest, but
unlike all the other PVC detection work, we have used unsupervised learning classi-
fiers. That means, our results are inherently objective, only the classification assess-
ment was based on human experience.
Unsupervised learning develops classification labels automatically. It seeks out sim-
ilarities between data in order to determine whether or not they can be characterized
as forming a group. These groups are called clusters. Apparently, unsupervised learn-
ing results may vary widely and they may be completely off if the first steps are wrong.
However, cluster analysis has the ability to introduce grouping at a higher interpretative
level. Thus cluster analysis is a very promising tool.
The performance measures fur this project and indeed all the performance mea-
sures reported in Table 3 were obtained by testing models. In order to build physical
problem solutions, these models must be implemented [62]. This implementation is
the result of a design process which turns the theoretical model into a practical CAD
system, which benefits patients [63]. To achieve the promise of a certain quality, estab-
lished through model performance measures, the design process must be formal and
systematic [64]. This process is cost intensive and to follow through with this process
requires even stronger evidence. The current study was based on data from 47 sub-
jects. However, this sample size is insufficient to represent all current and future PVC
patients. To build a stronger case, which justifies the high development cost, the CAD
system must be tested with more cases that reflect all known facets of the disease.
10
Table 3: Arrhythmia detection systems. – indicates not reported, No S. indicates num-
ber of subjects taking part in the study, No F. indicates the number of features.
Year Arrhythmia No S. No F. Acc Sn Sp
This study 2013 NSR, PVC 47 8 80.94 81.10 80.10
Martis et al. [44] 2013 NSR, AF 47 10 99.33 99.32 99.33
Martis et al. [45] 2013 NSR, AF,
AFL
48 10 97.65 98.16 98.75
Martis et al. [46] 2013 NSR, RBBB,
LBBB, APC,
PVC
– 12 93.48 99.27 98.31
Sufi and Khalil
[47]
2011 NSR, APC,
PVC, AF
50 11 97 – –
Huang et al.
[48]
2011 NSR, AF 48 4 – 96.1 98.1
Lim [49] 2009 NSR, PVC 3 48 99.80 – –
Sarkar et al.
[50]
2008 NSR, AF, AT 307 – 96 94 –
Inan et al. [51] 2006 NSR, PVC,
LBBB, RBBB
47 42 95.16 – –
Logan and
Healey [52]
2005 AF – – – 96 89
Christov et al.
[6]
2005 NSR, PVC 48 11 – 75.4 80.9
Christov et al.
[53]
2004 NSR, PVC 48 26 – – –
Jekova et al.
[54]
2004 NSR, PVC 48 6 – 94.6 98
Zhou [55] 2003 NSR, PVC 25 10 – – –
Tateno and
Glass [56]
2001 NSR, AF 48 2 – 86.6 84.3
Wang et al. [57] 2001 AF, VF, VT 180 6 96 97 98
Al-Nashash [58] 2000 NSR, PB,
LBBB,
RBBB, NEB,
NPB
14 20 – 98.1 94.7
Cerutti et al.
[59]
1997 NSR, AF 7 5 – 96 81
Slocum et al.
[60]
1992 NSR, AF 73 – – 87.8 68.3
Charles Oliver
[61]
1971 NSR, PVC 34 3 78 – –
11
6 Conclusion
This paper introduces a processing structure for an ECG based PVC diagnosis sup-
port system. The first processing step subjects the ECG signals to DWT based noise
reduction. Processing step number two employed ICA to reduce the data dimensional-
ity. Students t-test was used to determine the most discriminative ICA results. These
results were used in the classification step to determine whether or not the ECG signal
shows signs of PVC. We have studied a large number of pre-classified ECG signals
to assess the proposed system and indeed we have used these signals to answer vital
questions, such as the ones revolving around feature selection. During the assess-
ment, we found that eight features were clinically significant and these features were
used as input to the classifiers. With these features the FCM algorithm achieved an
accuracy of 80.94%. Sensitivity and specificity of the same classifier were measured
as 81.10% and 80.10% respectively.
This technique of recognizing cardiac arrhythmias in ECG recordings is very useful
for patients who live in countries where the level of medical expertise is vastly different
between metropolitan centers and rural areas. The reason for this beneficial effect
comes from the fact that an initial assessment of ECG signals can be done by less
skilled personal. This also reduces the work load for experienced practitioners. As a
direct consequence, this reduces the need to travel from an underdeveloped country
side to a metropolitan area with medical expertise. This safes travel cost and it reduces
patient discomfort. Hence, this automated screening method can help to bring quality
diagnosis to the less fortunate.
7 Acronyms
AF Atrial Fibrillation
AFL Atrial Flutter
APC Atrial Premature Contraction
AT Atrial Tachycardia
AV Atrioventricular
CAD Computer-Aided Diagnosis
CVD Cardiovascular Disease
DWT Discrete Wavelet Transform
ECG Electrocardiogram
EEG Electroencephalogrphy
FCM Fuzzy C-Means
FN False Negative
FP False Positive
ICA Independent Component Analysis
LBBB Left Bundle Branch Block
NEB Nodal Escape Beat
NPB Nodal Premature Beat
NSR Normal Sinus Rhythm
PB Paced Beat
PPV Positive Predictive Value
PVC Premature Ventricular Contraction
RBBB Right Bundle Branch Block
12
SA Sino Atrial
TN True Negative
TP True Positive
VF Ventricular Fibrillation
VT Ventricular Tachycardia
References
[1] S. S. Anand, S. Yusuf, Stemming the global tsunami of cardiovascular disease,
The Lancet 377 (9765) (2011) 529–532.
[2] H. Calkins, J. Brugada, D. L. Packer, R. Cappato, S.-A. Chen, H. J. G. Crijns,
R. J. Damiano, D. W. Davies, D. E. Haines, M. Haissaguerre, Hrs/ehra/ecas ex-
pert consensus statement on catheter and surgical ablation of atrial fibrillation:
Recommendations for personnel, policy, procedures and follow-up a report of the
heart rhythm society (hrs) task force on catheter and surgical ablation of atrial
fibrillation developed in partnership with the european heart rhythm association
(ehra) and the european cardiac arrhythmia society (ecas); in collaboration with
the american college of cardiology (acc), american heart association (aha), and
the society of thoracic surgeons (sts). endorsed and approved by the governing
bodies of the american college of cardiology, the american heart association, the
european cardiac arrhythmia society, the european heart rhythm association, the
society of thoracic surgeons, and the heart rhythm society., Europace 9 (6) (2007)
335–379.
[3] P. J. Podrid, P. R. Kowey, Cardiac Arrhythmia: Mechanisms, Diagnosis, and Man-
agement, Lippincott Williams & Wilkins, 2001.
[4] H. V. Huikuri, A. Castellanos, R. J. Myerburg, Sudden death due to cardiac ar-
rhythmias, New England Journal of Medicine 345 (20) (2001) 1473–1482.
[5] U. Acharya, D. Ghista, Z. KuanYi, L. C. Min, E. Ng, S. Sree, O. Faust, L. Weidong,
A. Alvin, Integrated index for cardiac arrythmias diagnosis using entropies as fea-
tures of heart rate variability signal, in: Biomedical Engineering (MECBME), 2011
1st Middle East Conference on, IEEE, 2011, pp. 371–374.
[6] I. Christov, I. Jekova, G. Bortolan, Premature ventricular contraction classifica-
tion by the kth nearest-neighbours rule, Physiological measurement 26 (1) (2005)
123–130.
[7] R. J. Martis, U. R. Acharya, H. Prasad, C. K. Chua, C. M. Lim, J. S. Suri, Applica-
tion of higher order statistics for atrial arrhythmia classification, Biomedical Signal
Processing and Control 8 (6) (2013) 888–900.
[8] M. M. A. Hadhoud, M. I. Eladawy, A. Farag, Computer aided diagnosis of car-
diac arrhythmias, in: Computer Engineering and Systems, The 2006 International
Conference on, IEEE, 2006, pp. 262–265.
[9] U. R. Acharya, O. Faust, S. V. Sree, D. N. Ghista, S. Dua, P. Joseph, V. T. Ahamed,
N. Janarthanan, T. Tamura, An integrated diabetic index using heart rate variability
signal features for diagnosis of diabetes, Computer methods in biomechanics and
biomedical engineering 16 (2) (2013) 222–234.
13
[10] U. R. Acharya, O. Faust, S. V. Sree, F. Molinari, L. Saba, A. Nicolaides, J. S. Suri,
An accurate and generalized approach to plaque characterization in 346 carotid
ultrasound scans, Instrumentation and Measurement, IEEE Transactions on 61 (4)
(2012) 1045–1053.
[11] U. R. Acharya, E. C.-P. Chua, O. Faust, T.-C. Lim, L. F. B. Lim, Automated detec-
tion of sleep apnea from electrocardiogram signals using nonlinear parameters,
Physiological measurement 32 (3) (2011) 287.
[12] R. J. Martis, U. R. Acharya, K. 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.
[13] T. F. Sanam, Determination of Premature Ventricular Contraction Beat in Ecg Sig-
nal: Of Heart with Automated Methods Using the Empirical Mode Decomposi-
tion (EMD), Morphological Transform and Correlation Coefficient, VDM Publish-
ing, 2011.
[14] L. B. Rowell, Human cardiovascular control, Oxford University Press, 1993.
[15] C. P. Cannon, G. Khan, Rapid ECG Interpretation, Contemporary Cardiology, Hu-
man Press, 2008.
[16] D. H. Singer, R. Lazzara, B. F. Hoffman, Interrelationships between automaticity
and conduction in purkinje fibers, Circulation research 21 (4) (1967) 537–558.
[17] O. N. Tripathi, U. Ravens, M. C. Sanguinetti, Heart Rate and Rhythm: Molecu-
lar Basis, Pharmacological Modulation and Clinical Implications, SpringerLink :
B¨
ucher, Springer, 2011.
[18] M. E. Silverman, D. Grove, C. B. Upshaw, Why does the heart beat? the discovery
of the electrical system of the heart, Circulation 113 (23) (2006) 2775–2781.
[19] A. Natale, A. Raviele, Ventricular Tachycardia / Fibrillation Ablation: The state
of the Art based on the VeniceChart International Consensus Document, Wiley,
2009.
[20] J. Jalife, R. A. Gray, G. E. Morley, J. M. Davidenko, Self-organization and the
dynamical nature of ventricular fibrillation, Chaos: An Interdisciplinary Journal of
Nonlinear Science 8 (1) (1998) 79–93.
[21] J. Huff, ECG Workout: Exercises in Arrhythmia Interpretation, Huff, ECG Workout,
Lippincott Williams & Wilkins, 2006.
[22] G. B. Moody, R. G. Mark, The mit-bih arrhythmia database on cd-rom and software
for use with it, in: Computers in Cardiology 1990, Proceedings., IEEE, 1990, pp.
185–188.
[23] N. V. Thakor, Y.-S. Zhu, Applications of adaptive filtering to ecg analysis: noise
cancellation and arrhythmia detection, Biomedical Engineering, IEEE Transac-
tions on 38 (8) (1991) 785–794.
[24] P. S. Addison, Wavelet transforms and the ecg: a review, Physiological measure-
ment 26 (5) (2005) R155.
14
[25] U. R. Acharya, O. Faust, S. V. Sree, F. Molinari, J. S. Suri, Thyroscreen system:
High resolution ultrasound thyroid image characterization into benign and malig-
nant classes using novel combination of texture and discrete wavelet transform,
Computer methods and programs in biomedicine 107 (2) (2012) 233–241.
[26] U. R. Acharya, O. Faust, S. V. Sree, A. P. C. Alvin, G. Krishnamurthi, J. C. R.
Seabra, J. Sanches, J. S. Suri, Atheromatictm : Symptomatic vs. asymptomatic
classification of carotid ultrasound plaque using a combination of hos, dwt & tex-
ture, in: Engineering in Medicine and Biology Society, EMBC, 2011 Annual Inter-
national Conference of the IEEE, IEEE, 2011, pp. 4489–4492.
[27] O. Faust, W. Yu, N. A. Kadri, Computer-based identification of normal and alco-
holic eeg signals using wavelet packets and energy measures, Journal of Me-
chanics in Medicine and Biology 13 (03).
[28] M. Alfaouri, K. Daqrouq, Ecg signal denoising by wavelet transform thresholding,
American Journal of Applied Sciences 5 (3) (2008) 276.
[29] L. Senhadji, G. Carrault, J. J. Bellanger, G. Passariello, Comparing wavelet trans-
forms for recognizing cardiac patterns, Engineering in Medicine and Biology Mag-
azine, IEEE 14 (2) (1995) 167–173.
[30] D. Wu, Z. Bai, An improved method for ecg signal feature point detection based
on wavelet transform, in: Industrial Electronics and Applications (ICIEA), 2012 7th
IEEE Conference on, 2012, pp. 1836–1841.
[31] B. N. Singh, A. K. Tiwari, Optimal selection of wavelet basis function applied to
ecg signal denoising, Digital Signal Processing 16 (3) (2006) 275–287.
[32] D. Langlois, S. Chartier, D. Gosselin, An introduction to independent component
analysis: Infomax and fastica algorithms, Tutorials in Quantitative Methods for
Psychology 6 (1) (2010) 31–38.
[33] S. Amari, A new learning algorithm for blind source separation, Advances in Neu-
ral Information Processing Systems 8 (1996) 757–763.
[34] A. Hyv¨
arinen, E. Oja, Independent component analysis: algorithms and applica-
tions, Neural networks 13 (4) (2000) 411–430.
[35] M. P. Fay, M. A. Proschan, Wilcoxon-mann-whitney or t-test? on assumptions for
hypothesis tests and multiple interpretations of decision rules, Statistics surveys
4 (2010) 1.
[36] J. H. McDonald, University of Delaware, Handbook of Biological Statistics, Sparky
House Publishing, 2009.
[37] J. MacQueen, Some methods for classification and analysis of multivariate obser-
vations, in: Proceedings of the fifth Berkeley symposium on mathematical statis-
tics and probability, Vol. 1, California, USA, 1967, pp. 281–297.
[38] T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silverman, A. Y. Wu,
An efficient k-means clustering algorithm: Analysis and implementation, Pattern
Analysis and Machine Intelligence, IEEE Transactions on 24 (7) (2002) 881–892.
15
[39] D. T. Pham, S. S. Dimov, C. D. Nguyen, Selection of k in k-means clustering, Pro-
ceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical
Engineering Science 219 (1) (2005) 103–119.
[40] A. Dallali, A. Kachouri, M. Samet, Fuzzy c-means clustering, neural network, wt
and hrv for classification of cardiac arrhythmia, ARPN Journal of Engineering and
Applied Sciences 6 (10).
[41] A. d. Amo, J. Montero, G. Biging, V. Cutello, Fuzzy classification systems, Euro-
pean Journal of Operational Research 156 (2) (2004) 495–507.
[42] G. H. Ball, D. J. Hall, Isodata, a novel method of data analysis and pattern classi-
fication, Tech. rep., DTIC Document (1965).
[43] J. C. Bezdek, Pattern recognition with fuzzy objective function algorithms, Kluwer
Academic Publishers, 1981.
[44] R. J. Martis, U. R. Acharya, H. Prasad, C. K. Chua, C. M. Lim, Automated detec-
tion of atrial fibrillation using bayesian paradigm, Knowledge-Based Systems 54
(2013) 269–275.
[45] R. J. Martis, U. R. Acharya, C. M. LIM, K. M. Mandana, A. K. Ray, C. Chakraborty,
Application of higher order cumulant features for cardiac health diagnosis using
ecg signals, International journal of neural systems 23 (04).
[46] R. J. Martis, U. R. Acharya, K. Mandana, A. Ray, C. Chakraborty, Cardiac decision
making using higher order spectra, Biomedical Signal Processing and Control
8 (2) (2013) 193–203.
[47] F. Sufi, I. Khalil, Diagnosis of cardiovascular abnormalities from compressed ecg:
A data mining-based approach, Information Technology in Biomedicine, IEEE
Transactions on 15 (1) (2011) 33–39.
[48] C. Huang, S. Ye, H. Chen, D. Li, F. He, Y. Tu, A novel method for detection of
the transition between atrial fibrillation and sinus rhythm, Biomedical Engineering,
IEEE Transactions on 58 (4) (2011) 1113–1119.
[49] J. S. Lim, Finding features for real-time premature ventricular contraction detection
using a fuzzy neural network system, Neural Networks, IEEE Transactions on
20 (3) (2009) 522–527.
[50] S. Sarkar, D. Ritscher, R. Mehra, A detector for a chronic implantable atrial tach-
yarrhythmia monitor, Biomedical Engineering, IEEE Transactions on 55 (3) (2008)
1219–1224.
[51] O. T. Inan, L. Giovangrandi, G. T. A. Kovacs, Robust neural-network-based classi-
fication of premature ventricular contractions using wavelet transform and timing
interval features, Biomedical Engineering, IEEE Transactions on 53 (12) (2006)
2507–2515.
[52] B. Logan, J. H. Healey, Robust detection of atrial fibrillation for a long term tele-
monitoring system, in: Computers in Cardiology, 2005, IEEE, 2005, pp. 619–622.
16
[53] I. Christov, G. Bortolan, Ranking of pattern recognition parameters for premature
ventricular contractions classification by neural networks, Physiological Measure-
ment 25 (5) (2004) 1281.
[54] I. Jekova, G. Bortolan, I. Christov, Pattern recognition and optimal parameter se-
lection in premature ventricular contraction classification, in: Computers in Cardi-
ology, 2004, IEEE, 2004, pp. 357–360.
[55] J. Zhou, Automatic detection of premature ventricular contraction using quantum
neural networks, in: Bioinformatics and Bioengineering, 2003. Proceedings. Third
IEEE Symposium on, IEEE, 2003, pp. 169–173.
[56] K. Tateno, L. Glass, Automatic detection of atrial fibrillation using the coefficient
of variation and density histograms of rr and δrr intervals, Medical and Biological
Engineering and Computing 39 (6) (2001) 664–671.
[57] Y. Wang, Y.-S. Zhu, N. V. Thakor, Y.-H. Xu, A short-time multifractal approach
for arrhythmia detection based on fuzzy neural network, Biomedical Engineering,
IEEE Transactions on 48 (9) (2001) 989–995.
[58] H. Al-Nashash, Cardiac arrhythmia classification using neural networks, Technol-
ogy and Health Care 8 (6) (2000) 363–372.
[59] S. Cerutti, L. T. Mainardi, A. Porta, A. M. Bianchi, Analysis of the dynamics of
rr interval series for the detection of atrial fibrillation episodes, in: Computers in
Cardiology 1997, IEEE, 1997, pp. 77–80.
[60] J. Slocum, A. Sahakian, S. Swiryn, Diagnosis of atrial fibrillation from surface
electrocardiograms based on computer-detected atrial activity, Journal of Electro-
cardiology 25 (1) (1992) 1–8.
[61] G. Charles Oliver, F. M. Nolle, G. A. Wolff, J. R. Cox Jr, H. D. Ambos, Detection of
premature ventricular contractions with a clinical system for monitoring electrocar-
diographic rhythms, Computers and Biomedical Research 4 (5) (1971) 523–541.
[62] O. Faust, R. Acharya U, B. H. C. Sputh, L. C. Min, Systems engineering principles
for the design of biomedical signal processing systems, Computer methods and
programs in biomedicine 102 (3) (2011) 267–276.
[63] O. Faust, C. W. Yan, M. R. K. Mookiah, U. R. Acharya, E. Y. K. Ng, W. Yu, Formal
design and development of an anterior segment eye disease classification system,
Image Analysis and Modeling in Ophthalmology (2014) 245.
[64] O. Faust, U. R. Acharya, T. Tamura, Formal design methods for reliable computer-
aided diagnosis: a review, Biomedical Engineering, IEEE Reviews in 5 (2012)
15–28.
17
