Information Sciences 377 (2017) 17–29

Contents lists available at ScienceDirect

Information Sciences
journal homepage: www.elsevier.com/locate/ins

Automated characterization and classification of coronary

artery disease and myocardial infarction by decomposition of
ECG signals: A comparative study
U Rajendra Acharya a,b,c, Hamido Fujita d,∗, Muhammad Adam a, Oh Shu Lih a,
Vidya K Sudarshan a, Tan Jen Hong a, Joel EW Koh a, Yuki Hagiwara a,
Chua K. Chua a, Chua Kok Poo a, Tan Ru San e
a
Department of Electronics and Computer Engineering, Ngee Ann Polytechnic, Singapore
b
Department of Biomedical Engineering, School of Science and Technology, SIM University, Singapore
c
Department of Biomedical Engineering, Faculty of Engineering, University of Malaya, Malaysia
d
Iwate Prefectural University (IPU), Faculty of Software and Information Science, Iwate 020-0693 Japan
e
Department of Cardiology, National Heart Centre, Singapore

a r t i c l e i n f o a b s t r a c t

Article history: Cardiovascular diseases (CVDs) are the main cause of cardiac death worldwide. The Coro-
Received 23 June 2016 nary Artery Disease (CAD) is one of the leading causes of these CVD deaths. CAD condition
Revised 13 August 2016
progresses rapidly, if not diagnosed and treated at an early stage may eventually lead to
Accepted 5 October 2016
an irreversible state of heart muscle death called Myocardial Infarction (MI). Normally, the
Available online 6 October 2016
presence of these cardiac conditions is primarily reflected on the electrocardiogram (ECG)
Keywords: signal. However, it is challenging and requires rich experience to manually interpret the
Coronary artery disease visual subtle changes occurring in the ECG waveforms. Thus, many automated diagnostic
Myocardial infarction systems are developed to overcome these limitations. In this study, the performance of an
Electrocardiogram automated diagnostic system developed for detection of CAD and MI using three meth-
Discrete cosine transform ods such as Discrete Wavelet Transform (DWT), Empirical Mode Decomposition (EMD) and
Discrete wavelet transform Discrete Cosine Transform (DCT) are compared. In this study, ECG signals are subjected
Empirical mode decomposition to DCT, DWT and EMD to obtain respective coefficients. These coefficients are reduced
using Locality Preserving Projection (LPP) data reduction method. Then, the LPP features
are ranked using F-value. Finally, the highly ranked coefficients are fed into the K-Nearest
Neighbor (KNN) classifier to achieve the best classification performance. Our proposed sys-
tem yielded highest classification results of 98.5% accuracy, 99.7% sensitivity and 98.5%
specificity using only seven features obtained using DCT technique. The screening system
can help the cardiologists in detecting the CAD and hence presents any possible MI by
prescribing suitable medications. It can be employed in routine community screening, old
age homes, polyclinics and hospitals.
© 2016 Elsevier Inc. All rights reserved.

∗
Corresponding author.
E-mail address: [email protected] (H. Fujita).

http://dx.doi.org/10.1016/j.ins.2016.10.013
0020-0255/© 2016 Elsevier Inc. All rights reserved.
18 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29

1. Introduction

Globally, cardiovascular diseases (CVDs) remain as the number one cause of cardiac death. In 2012, a total of 17.5 million
deaths due to CVDs are reported worldwide, which accounts for 31% of global death. Of these, an approximately 7.4 million
deaths are due to Coronary Artery Disease (CAD) [59]. CAD is primarily the result of atherosclerosis, in which fibrous plaques
begin to form thick regions in the inner wall of the arteries [22,24,75]. These atherosclerotic plaques are basically made up
of fats and fibrous tissues [23,25]. CAD is the leading cause of death in the United Kingdom (UK) with close to half (45%)
of CVDs death are due to it. As of 2012, 16% and 10% of male and female deaths respectively are due to CAD, approximately
74,0 0 0 deaths in total [73,74]. In 2013, US reported mortality of 370,213 due to CAD, which is equivalent to 1 in every 7
Americans [12]. In severe condition, rupturing of these atherosclerotic plaques leads to acute Ischemic Heart Disease (IHD)
that comes with spectrum of Acute Coronary Syndrome (ACS), in particularly Myocardial Infarction (MI).
MI irreversible muscle death occurs when there is a disruption or reduction of blood flow to the heart muscles (my-
ocardium) [27]. Evidently, the pathogenesis for majority of MI conditions is caused by atherosclerotic plaques that occlude
the local coronary artery [36]. With severe and prolonged MI, the dead cardiac muscle tissue is replaced with scar tissue
that has no myocardial contractility. In other words, myocardium permanently loses its contractility at the region of MI.
Every year, 660,0 0 0 Americans are experiencing fresh coronary attack, first MI, and 305,0 0 0 are having repeated episodes.
In addition, 160,0 0 0 silent MIs happens every year [59]. It is reported that, in UK, approximately 1% of men and 0.5% of
women are affected with MI. In 2012, a total of 175,0 0 0 people are hospitalized due to MI, which is around 1 in every
3 min [74].
These epidemiological studies on CAD and MI indicate the significance for the need of an early detection system. Clin-
ically, Electrocardiogram (ECG) is the most commonly preferred diagnostic tool due to its non-invasiveness, and low cost.
The ECG signal provides vital information with regard to the function and rhythm of the heart. The morphological changes
in the ECG signal are observed as the heart experiences episode of ischemia, infarction or arrhythmias due to CAD ([43];
Levin 2013; [27,39]). Thus, standard or resting ECG can be a convenient noninvasive diagnostic method for detecting the
CAD and MI. Within seconds of MI, the T waves appear abnormally high and the QT intervals are longer than normal [27].
Also, the elevation of ST segment is an indication that the heart is experiencing serious and extensive myocardial ischemia.
In contrast, sub endocardial ischemia is reflected as a ST segment depression [27]. However, about 70% of CAD patients do
not exhibit any significant variations in their ECG signals [70], thus, it is tedious to manually or visually examine and infer
these subtle morphological changes in long continuous ECG beats. Thus, this process is time consuming and prone to er-
rors due to fatigue. Therefore, an automated diagnostic system is essential which can overcome these limitations of manual
evaluation of ECG signals [58].
Over the years, automated classification of heart rate [2] or ECG signals [65] for the detection of CAD and MI have
been developed. Summary of various detection and diagnosis studies of CAD and MI is tabulated in Table 1. Many signal
processing algorithms such as linear (time and frequency domain) [9,15,72,67,15,72], nonlinear methods [4,45,48,49,53],
Discrete Wavelet Transform (DWT) [19,33,42,44], and Tunable Q Wavelet Transform (TQWT) [64] have been developed to
accurately extract features from Heart Rate (HR) or ECG signals for the classification of normal or abnormal (CAD or MI)
signals (Table 1).
Thus, it is evident from the literature review Table 1 that all the studies reported focuses on automated classification of
two classes (normal and CAD or MI). To best of our knowledge, this is the first study on the characterization and classifica-
tion of three classes (normal, CAD and MI) using ECG signals.
The proposed study compares the performance of three different techniques namely, DWT [31], Empirical Mode Decom-
position (EMD) [31] and Discrete Cosine Transform (DCT) [31], in detection of normal, CAD and MI classes using ECG beats.
The features (coefficients and IMFs) obtained from the respective transformation techniques (DWT, DCT and EMD) are ap-
plied to Locality Preserving Projection (LPP) reduction method. Then, the reduced features are ranked using Analysis of
Variance (ANOVA) statistical analysis technique. The highly ranked features are classified using K-Nearest Neighbor (KNN)
classifier. The robustness of the proposed system is justified with ten-fold cross validation technique. The proposed sys-
tem yielded a highest classification performance of 98.5% accuracy, 99.7% sensitivity and 98.5% specificity using only seven
features obtained from DCT method.

2. Methodology

2.1. Materials

For this study, the ECG signals were obtained from Physionet open access databases. The ECG signals of Normal and MI
were acquired from PTB Diagnostic ECG Database while the ECG signals for CAD were acquired from St.-Petersburg Institute
of Cardiological Technics 12-lead Arrhythmia Database [34]. A total of 7 CAD, 148 MI and 52 Normal subjects ECG records
were used.
 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29 19

Table 1
Overview of study using HRV and ECG signals in the diagnosis of CAD and MI.

Author
(Year) Database Techniques/Features Efficiency

Normal and CAD using HRV signals

[48] Lead: lead II Frequency domain, Time domain, Poincare plot, SVM:
Subjects: 99 CAD, 94 Normal Approximate entropy Accuracy = 90%
Source: supine, right lateral, left lateral positions Support vector machine (SVM)
Classification based on multiple association rules
(CMAR)
Naïve Bayesian (NB)
Decision tress (DT)
[45] Lead: 12 leads Frequency domain, Time domain, Poincare plots, Accuracy = 72.5–84.6%
Subjects: 20 Normal, 64 CAD (51 angina pectoris, Fractal scaling measures, Complexity estimations
13 acute coronary syndrome) Multiple discriminant analysis (MDA)
Source: supine, right lateral, left lateral positions
[49] Lead: lead II Frequency domain, Time domain, Ooincare plot, CPAR & SVM:
Subjects: 99 CAD, 94 Normal Hurst exponent, DFA, Approximate entropy Accuracy = 85–90%
Source: supine, right lateral, left lateral positions SVM, CMAR, MDA, NB, DT and Classification
based on predictive association rules (CPAR)
[33] Lead: Lead II DWT GMM classifier:
Subjects: 10 CAD, 15 Normal SVM, Gaussian mixture model (GMM) Accuracy = 96.8%
Source: Iqraa hospital, Calicut, Kerela, India K-Nearest Neighbors (KNN) Sensitivity = 100%
Probabilistic neural network (PNN) Specificity = 93.7%
[64] Lead: Lead II TQWT and CC Accuracy = 99.7%
Subjects: 10 CAD, 10 Normal Least squares SVM (LS-SVM) Sensitivity = 99.6%
Source: Iqraa hospital, Calicut, Kerela, India Specificity = 99.8%
[46] Lead: Lead II Flexible Analytic Wavelet Transform and Entropy Accuracy = 100%
Subjects: 10 CAD, 10 Normal Features, LS-SVM Sensitivity = 100%
Source: Iqraa hospital, Calicut, Kerela, India Specificity = 100%
[3] Lead: Lead II Time domain and frequency domain analysis, No accuracy reported.
Subjects: 10 normal 10 CAD ECG signals Nonlinear methods (Poincare geometry, Proposed range of parameters
Source: Iqraa Hospital, Calicut, India Recurrence quantification analysis, approximate for two classes with (p < 0.05)
entropy, sample entropy, DFA, correlation
dimension, HOS, EMD)
[71] Lead: Lead II EMD/second-order difference plot area, analytical No accuracy reported.
Subjects: 10 normal 10 CAD ECG signals signal representation area, amplitude modulation Proposed range of parameters
Source: Iqraa Hospital, Calicut, India bandwidth, frequency modulation bandwidth, for two classes with (p < 0.05)
Fourier-Bessel expansion
Normal and CAD using ECG signals
[69] Lead: Lead V3 to V4 BCT Men:
Subjects: 62 men, (51 CAD & 11 normal), 31 Blinded test Sensitivity = 84.3%
women (21 CAD & 10 normal) Fisher’s exact test Specificity = 81.8%
Source: resting ECGs Women:
Sensitivity = 76.2%
Specificity = 80%
[50] Lead: 9 leads Artificial neural network (ANN) Roc = 91.5%
Subjects: 127 CAD, 220 Normal Receiver operating characteristics (ROC) analysis
Source: Bicycle exercise ECG test
[51] Lead: 12 leads RBF neural networks Average specificity and
Subjects: 479 CAD, 297 Normal sensitivity of about 97%
Source: Exercise stress ECG
[13] Lead: lead v5 Fuzzy uncertainty Combined Uncertainty:
Subjects: 8 CAD, 7 Normal Probabilistic uncertainty 80% correct classification
Source: Exercise stress ECG Combined uncertainty percentage (CCP)
[16] Lead: 12 leads PCA, SVM Accuracy = 79.17%
Subjects: 480 patients
Source: Exercise stress ECG
[17] Lead: 12 leads Binary particle swarm Binary particle swarm
Subjects: 480 patients Genetic algorithm optimization – feature
Source: Exercise stress ECG SVM selection technique:
Accuracy = 81.46%
[76] Lead: 12 leads Wavelet decomposition Accuracy = 80%
Subjects: 46 R-wave peaks detection
Source: Long-term ST database (motion activity) ST segment detection
[44] Lead: Lead II DWT, PCA and SVM Accuracy = 88%
Beat(s): raw ECG waveform
Subjects: 43 CAD, 49 Normal
Source: Multi-parameter Intelligent Monitoring
in Intensive Care (MIMIC II)
(continued on next page)
20 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29

Table 1 (continued)

Author Database Techniques/Features Efficiency

(Year)

[5] Lead: 12 leads HOS bispectrum and cumulants Using 13 Bispectrum features:
Beat(s): 182,013 ECG beats PCA Accuracy = 98.17%
Subjects: 40 normal 7 CAD KNN and DT Sensitivity = 94.57%
Source: St. Petersburg Institute of Cardiological Specificity = 99.34%
Technics 12-lead Arrhythmia database; Fantasia Using 31 cumulant features:
open access database. Accuracy = 98.99%
Sensitivity = 97.75%
Specificity = 99.39%
Normal and MI using ECG signals
[53] Lead: 12 leads ANN Training set:
Beat(s): ST-segment & T wave Fuzzy logic system Sensitivity = 94.2%,
Subjects: 104 MI, 20 Normal Specificity = 100%
Testing set:
Sensitivity = 84.6%
Specificity = 90%
[47] Lead: 12 leads Detection of R-peaks Efficiency = 96%
Beat(s): 64,680 R-peaks Phase space fractal dimension (PSFD)
Subjects: 177 MI, 125 Normal Neuro-Genetic algorithm classifier
[26] Lead: lead V1, V2, V3, V4 ECG beat sampling Hidden markov models:
Beat(s): 582 MI, 547 Normal Hidden markov models Sensitivity = 79%
Subjects: 1129 samples of heartbeats from Gaussian mixture models Accuracy = 71.5%
clinical data
[42] Lead: lead II DWT, Energy-entropy Accuracy = more than 95%
Beat(s): 2282 Normal, 718 MI Energy and entropy plots
Subjects: 2 MI, 6 Normal
Banerjee PTB database Denoise –DWT PTB database
et al., Lead: leads V1-V4 QRS complex band selection Sensitivity = 99.6%
2010 Beat(s): 42,852 R peak detection MIT BIH database
Subjects: 5 patients Baseline detection Sensitivity = 99.8%
MIT BIH arrhythmia database QRS vector computation
Lead: leads V1-V4
Beat(s): 19,062
Subjects:
[14] Lead: 12 leads T wave amplitude, Q wave, ST level deviation Detection:
Beat(s): 10,580 MI, 1840 Normal PCA, Back propagation neural networks Sensitivity = 97.5%
Subjects: 148 MI, 52 Normal (594 records) Specificity = 99.1%
Localization:
Accuracy = 93.7%
[9] Lead: 12 leads R-complex time index detection Sensitivity = 85%
Beat(s): ST segment analysis Q & S time index detection Specificity = 100%
Subjects: 20 MI, 20 Normal J-point time index detection
Reference point detection (REF)
ST segment analysis
[19] PTB database DWT MIT BIH Arrhythmia database:
Beat(s): 42,852 QRS complex band selection & finding the QRS Sensitivity = 99.8%
Subjects: 5 patients window PTB database:
MIT BIH arrhythmia database R peak detection Sensitivity = 99.84%
Lead: leads V1-V4 Q & S point detection Accuracy: 96.4%
Beat(s): 19,062 T wave detection
Subjects: QRS vector computation
Mahalanobish distance based classification
[15] Lead: 12 leads Q wave amplitude MI detection:
Beat(s): 16,960 MI, 3200 Normal ST level deviation Sensitivity = 99.97%
Subjects: 10 types of MI, 1 Normal T wave amplitude Specificity = 99.9%
KNN classifier MI localization:
Sensitivity & specificity = more
than 99%
Accuracy = 98.8%
[20] Lead: Lead 3 R peak registration ECG patterns shows difference
Beat(s): 1 beat Time normalization of cardiac cycles characteristics over QRS and T
Subjects: 148 MI, 52 Normal Cross transform and Wavelet coherence of ECG waves regions.
beats
[72] Lead: 12 leads ST segment Sensitivity = 91%
Subjects: 369 MI, 79 normal Polynomial fitting Specificity = 85%
KNN ensemble
(continued on next page)
 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29 21

Table 1 (continued)

Author Database Techniques/Features Efficiency

(Year)

[67] Lead: lead 2 T-wave integral Localization:

Beat(s): 1 beat ANN, PNN, KNN, Multilayer perceptron (MLP), NB Accuracy = 76%
Subjects: 290 subjects (549 records) Detection:
Accuracy = 94%
[52] Lead: 12 leads ECG polynomial fitting algorithm (PolyFit) MI detection:
Beat(s): 1 beat PolyFit-based ECG parameterization algorithm Accuracy = 94.4%
Subjects: 148 MI, 52 Normal (PolyECG)
Akaike information criterion (AIC)
[68] Lead: 12 leads Wavelet transform of multi-lead ECG MI detection:
Beat(s): Multi-scale energy Accuracy = 96%
Subjects: 148 MI, 52 Normal (549 ECG records) Multiscale Eigen space analysis Sensitivity = 93%
Specificity = 99%
MI localization:
Accuracy = 99.58%
[4] Lead: 12 leads DWT on ECG beats MI detection:
Beat(s): 485,753 MI, 125,652 Normal signal energy, approximate, fuzzy, Accuracy = 98.8%
Subjects: 148 MI of 10 types, 52 Normal Kolmogrov-Sinai, permutation, Renyi, Shannon, Sensitivity = 99.45%
Tsallis, wavelet entropies, fractal dimension Specificity = 96.27%
Kolmogrov complexity MI localization:
Largest lyapunov exponent Accuracy = 98.74%
Sensitivity = 99.55%
Specificity = 99.16%
Normal, CAD and MI using ECG signals
In this Lead: lead 2 DCT coefficients, DWT coefficients and IMFs of DCT coefficients and KNN
work Beat(s): 41,545 CAD, 40,182 MI, 10,546 Normal EMD classifier:
Subjects: 7 CAD, 148 MI, 52 Normal Accuracy = 98.5%,
Sensitivity = 99.7%
Specificity = 98.5%

Fig. 1. System sketch of the proposed system.

2.2. Methods

System sketch of the proposed methodology is shown in Fig. 1. The features extracted using DWT, DCT and EMD from
the pre-processed ECG beats are reduced and ranked using LPP and F-value respectively. Subsequently, the highly ranked
features are fed to the KNN classifier independently for the characterization or normal, CAD and MI ECG signals.

2.2.1. Pre-processing
The ECG signals from the databases [34,41] are sampled at 257 Hz for CAD and 10 0 0 Hz for MI and normal classes. In
order to ensure uniformity across the databases, sampling frequency of 10 0 0 samples per second is chosen for this study.
ECG signals are preprocessed to remove the baseline wander and noise using Daubechies-6 (db6) mother wavelet [55].
22 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29

Table 2
Overview of the ECG beats used.

Class Number of beats

Normal 10,546
MI 40,182
CAD 41,545
Total beats 92,273

Fig. 2. Typical ECG beat: normal, MI and CAD.

2.2.2. Beats segmentation

The ECG beat segmentation of the pre-processed ECG signals is subjected to R-peak detection using Pan-Tompkins tech-
nique [54,63]. The R-peak is chosen as the distinctive point because of its visible high amplitude. In this study, all the ECG
signals were segmented, using the detected R-peaks, without the inclusion of the first and last beats. The segmentation of
the ECG beat is done by taking 250 and 400 samples before and after the R-peak respectively. A total of 92,273 ECG beats
were used consisting of 10,546 normal, 41,545 CAD and 40,182 MI. The summary of number of ECG beats used in each class
is shown in Table 2. Fig. 2 shows the normal, MI, and CAD beats segmented from continuous ECG signals.

2.2.3. Feature extraction

Each ECG beat is subjected to DCT, DWT and EMD methods separately to obtain their coefficients and IMFs respectively.
Three techniques used are briefly described in the following sections.
Discrete Cosine Transform (DCT)
The DCT technique has the ability of energy compaction within a restricted number of DCT coefficients. In general, the
DCT [7] of an ECG beat of length N is commonly defined as:

N−1  π 2n + 1 m 
( )
C (m ) = α (m ) f (n )cos for m = 0, 1, 2 . . . , N − 1. (1)
2N
n=0

The inverse DCT is defined as:


N−1  π 2n + 1 m 
( )
f (n ) = α (m )C (m )cos for n = 0, 1, 2 . . . , N − 1 (2)
2N
m=0

For both Eqs. (1) and (2), α (m) is defined as:


1
α (m ) = ,m=0
N

2
α (m ) = , m = 0
N


N−1
From (1), it is clear that when C (m = 0 ) = 1
N f (x ), u = 0. Hence, the first DCT coefficient is the average value of
n=0
the ECG beat samples.
In this study, each ECG beat of 651 samples are compressed to of 217 DCT coefficients. A total of 217 × 92,273 DCT
coefficients are obtained for three classes.
Discrete Wavelet Transform (DWT)
The DWT technique converts the time domain signal into wavelet domain to obtain the time and frequency values in
terms of coefficients [6,61]. In DWT, the ECG signals undergo high pass filtering (HPF) and low pass filtering (LPF) to decom-
pose the signals into different scales (frequency bands) [29]. The output coefficients of the HPF are called detail coefficients
 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29 23

and LPF are called approximations. Then approximations are further subjected to decomposition using both HPF and LPF
and thus this procedure is repeated for different levels. At each level of decomposition, the samples of a signal get reduced
due to sub-sampling by 2 [66].
In this work, the DWT is performed up to six levels of decomposition using Daubechies 4 (db4) mother wavelet [57].
At sixth level of decomposition, 17 detail coefficients are obtained for each ECG beat of 651 samples/length. Thus, a total
17 × 92,273 DWT coefficients are obtained for the three ECG classes.
Empirical Mode Decomposition (EMD)
The basis of EMD is to decompose a signal x (t) into sets of frequency and amplitude modulated signal components,
Intrinsic Mode Functions (IMFs) that has its own characteristic oscillations [40].
Using an iterative method known as sifting method, IMFs are obtained for a signal [28,40,54,62]. In this sifting method,
initially for a signal, obtain local maxima and local minima. Later by joining all the maxima together and minima together
obtain the envelopes for both respectively. Finally extract the IMF from a signal by subtracting the average of envelopes
(maxima and minima) [40]. This IMF extracted is tested to confirm whether it meets the following two conditions or not.

(1) The total number of zero crossing and extrema have to be equal or at least differ by one and,
(2) The mean value of the envelope that is define by the local minima and the envelope of the local maxima have to be
zero. The EMD methodology is described in [40].

Once the conditions are satisfied, the first IMF of a signal is obtained. To obtain the subsequent (remaining) IMFs, now
consider the residual signal as a new signal and repeat the above mentioned procedure. Perform the procedure until further
no more IMFs can be extracted from the residual signal [40,54].
In this study, EMD is performed up to two levels and used only second level IMF (IMF2). In this work, 651 IMF coefficients
are extracted from each ECG beat. So, a total of 651 × 92,273 IMF coefficients are obtained using EMD for the three classes
of ECG beats in this work.

2.2.4. Features reduction - Locality Preserving Projections (LPP)

It is a linear dimensionality technique proposed by He and Niyogi [38]. This technique optimally protects local neigh-
borhood details of the data. In this method, a graph is built by including the neighborhood details of the data set. The
transformation matrix is calculated using Laplacian of the graph which maps the data to a subspace. Thus, for high dimen-
sional data, the LPPs are calculated by obtaining the optimal linear approximations to the Eigen functions of the Laplace
Betrami operator [38].

2.2.5. Features ranking – ANOVA statistical analysis

In this study, all the features extracted undergo Analysis of Variance (ANOVA) statistical test [32]. This is to determine
the difference between the three classes according to the F-value (Fisher’s discriminatory value) and p-value (probability
value). In fact, features with high F-value and p-value < 0.0 0 01 indicate good discrimination characteristics. In this study,
the features are ranked based on F-value and the highly ranked features are then fed into the classifier independently.

2.2.6. Classification – K-Nearest Neighbor (KNN)

In KNN classification, each of the unknown samples are classified based on the majority known k nearest samples [35,37].
It is a type of non-parametric lazy learning technique whereby the function is locally estimated and computed in the recall
phase. This method is non-parametric because no assumptions are made for the data distribution. In this study, variation of
k nearest neighbors (k) are varied between 5 and 10. The highest accuracy is achieved for k = 5 in this work.

3. Results

In this study, a total of 92,273 ECG beats of three classes (10,546 normal, 40,182 MIs and 41,545 CADs) are segmented
from two separate ECG signal databases. For each ECG beat, three different transformation techniques are implemented,
namely DCT, DWT and EMD methods. We have obtained 217 DCT coefficients, 17 DWT coefficients and 651 IMFs from each
ECG beat of 651 samples. A total of 217 × 92,273 DCT coefficients, 17 × 92,273 DWT coefficients and 651 × 92,273 IMF coef-
ficients are obtained for three classes. Further, the coefficients and IMFs extracted are reduced using LPP method and sub-
sequently, ranked in descending order of F value obtained from ANOVA statistical analysis. In Fig. 3 and Table 3 (Appendix)
we can see twelve highly ranked LPP coefficients, and obtained using DCT method.
From Fig. 3 and Table 3 (appendix), it is evident that, based on the mean and SD values of the features, there is a distinct
discrimination between the three classes. Likewise, in Fig. 4 and Table 4 (appendix) twelve highly ranked LPP coefficients are
presented and obtained using DWT method.
From Fig. 4 and Table 4 (appendix), it is evident that, based on the mean and SD values of the features, there is a distinct
discrimination between the three classes. Likewise, in Fig. 5 and Table 5 (appendix) twelve highly ranked LPP coefficients are
presented and obtained using DWT method.
It is evident from Fig. 5 and Table 5 that the mean and SD value discrimination between the three classes is not as strong
as DCT and DWT techniques.
24 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29

Fig. 3. Bar graph of ranked LPP feature values obtained using DCT technique.

Table 3
Results of LPP coefficients obtained using DCT method for the normal, CAD and MI ECG signals (p < 0.0 0 01).

Features Normal CAD MI F-value

Mean SD Mean SD Mean SD

LPP3 3232.956 1564.078 972.4445 1458.025 1627.738 1042.379 12826.93

LPP4 −438.311 477.9786 32.98502 910.8666 −424.117 662.378 4092.994
LPP6 −436.666 673.9099 −28.7374 543.5191 189.9879 751.1795 4016.255
LPP2 4419.172 2439.109 3717.926 3184.244 2758.037 2109.766 2204.826
LPP5 −141.719 623.6197 −646.256 830.6713 −374.927 803.1463 2193.612
LPP1 −180.797 743.7059 −1039.77 4074.509 −102.049 1015.677 1216.948
LPP10 86.48351 305.7637 −45.7363 311.5426 9.386938 384.5519 695.7197
LPP9 −99.1001 286.8649 −51.6623 671.5911 −171.672 383.5564 534.1472
LPP8 −131.602 834.2096 −0.06408 449.7575 5.777961 577.2325 268.3632
LPP11 77.96811 172.3527 38.74155 681.9676 72.07488 243.4614 57.76763
LPP12 196.945 558.2805 234.1662 329.7941 231.3159 505.0017 31.05236
LPP7 268.6169 545.2547 270.1881 387.3952 292.3687 646.5498 20.28165

Fig. 4. Bar graph of ranked LPP feature values obtained using DWT technique.
 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29 25

Table 4
Results of LPP coefficients obtained using DWT method for the normal, CAD and MI ECG signals (p < 0.0 0 01).

Features Normal CAD MI F-value

Mean SD Mean SD Mean SD

LPP7 −74.3519 12.61883 −86.7201 18.29976 −79.1606 11.81201 4062.101

LPP2 740.382 392.0473 379.4176 460.1773 515.8773 300.6312 3897.824
LPP4 −62.7729 47.3126 −101.968 43.00439 −95.1268 38.37565 3742.785
LPP8 −5.49819 40.48845 25.10654 35.40835 18.40985 31.26369 3347.735
LPP1 −1527.97 795.0 0 04 −1392.87 1333.895 −1029.79 652.9709 1705.084
LPP6 −98.1445 21.92928 −103.941 40.9712 −95.581 24.51134 678.0044
LPP12 7.351888 15.39211 6.499149 25.52394 2.169629 16.49384 530.8095
LPP9 83.11746 13.37896 87.06508 15.05469 84.65347 13.77294 463.2879
LPP3 −104.856 193.6901 −49.7272 468.1063 −89.3416 176.673 188.4666
LPP11 −77.957 13.76112 −78.2606 17.23874 −79.8429 12.70406 137.4954
LPP10 89.8902 20.37151 92.43097 23.68047 93.48275 20.16914 115.2259
LPP5 93.64179 104.2482 83.32691 66.03578 89.54842 108.7017 76.78288

Fig. 5. Bar graph of ranked LPP feature values obtained using DWT technique.

Table 5
Results of LPP coefficients obtained using EMD method for the normal, CAD and MI ECG signals (p < 0.0 0 01).

Features Normal CAD MI F-value

Mean SD Mean SD Mean SD

LPP4 3.555597 2.61875 1.455717 1.698257 1.339736 1.65405 6678.991

LPP3 0.771773 2.355374 −0.52887 3.664289 0.578863 0.997758 2137.146
LPP2 1.155968 2.28786 1.369248 5.293276 0.176529 0.785825 1125.846
LPP7 0.470724 1.551883 0.504627 2.331978 −0.01065 0.684123 1003.068
LPP11 −0.09479 1.015333 0.284995 0.901178 0.135457 0.767533 900.6026
LPP10 −0.42925 1.091849 −0.3016 0.884296 −0.11915 0.901259 665.7695
LPP1 0.391799 3.303186 2.759511 16.76502 −0.03764 0.662832 660.2858
LPP9 0.19489 1.555882 −0.23136 1.557028 −0.00854 0.777704 597.8626
LPP12 −0.08003 1.199115 −0.16441 1.045123 0.004148 0.708488 331.7623
LPP5 0.580344 0.596162 0.305453 2.308159 0.342545 0.594684 124.0459
LPP6 0.092582 0.86712 0.020902 2.479644 0.075401 0.846943 12.55041
LPP8 −0.17982 2.068406 −0.19348 2.374641 −0.13768 0.926139 9.609805

Table 6
Classification results of the DCT, DWT and EMD techniques.

Method No of features Acc (%) PPV (%) Sen (%) Spec (%)

DCT 7 98.50 99.80 99.72 98.46

DWT 12 98.16 99.77 99.69 98.25
EMD 9 81.34 93.96 97.23 51.58

Acc = accuracy; Sen = sensitivity; Spec = specificity; PPV = positive predic-

tive value [11].
26 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29

Fig. 6. Sketch of accuracies (%) for different number of features using DCT method with KNN classifier.

Table 7
Confusion matrix of the 3 classes.

Original/Predicted Normal MI CAD

Normal 10,384 96 66
MI 133 39,642 407
CAD 94 586 40,865

The highly ranked features from the respective techniques are fed independently to KNN classifier. Ten-fold cross valida-
tion is implemented to ensure that the results obtained are reliable. As shown in Table 6, the integration of DCT technique
into the proposed system yielded the highest performance results of 98.5% accuracy, 99.72% sensitivity and 98.46% speci-
ficity using only seven features. Sketch of variation of accuracies (%) for different number of features for DCT method using
KNN classifier is presented in Fig. 6. Table 7 shows the confusion matrix of our work using KNN classifier for DCT method.
From Table 7, it can be observed that majority of the beats are correctly classified into their respective classes and only
few beats are incorrectly classified and treated as outliers.

4. Discussion

In this study, performances of DCT, DWT and EMD techniques for the diagnosis of CAD and MI using ECG beats are com-
pared. In this study, 92,273 (10,456 normal, 40,182 MI and 41,545 CAD) beats are used. Overall, the proposed system yielded
a highest classification accuracy of 98.5%, 99.7% sensitivity and 98.5% specificity using only seven features by employing DCT
technique.
DCT is a class of orthogonal transforms used for ECG data compression due to its energy compaction and decorrelation
properties [8,21]. It has the capabilities to concentrate most of the ECG signal information within a small subset of the
transform coefficients with insignificant error [7,10]. Hence, allowing classification to be performed with only few significant
features. In this study, 651 samples of the individual ECG beat are compressed to 217 DCT coefficients (one-third).
DWT is also used for ECG data compression due to its non-stationary and localized properties [1]. Moreover, DWT has the
ability to analyze ECG signal at different resolution by breaking down the ECG signal into several frequency bands [18]. This
decomposition method results in fewer coefficients without any loss of signal information or energy during both wavelet
and inverse transform [1,30].
EMD is signal dependent and adaptive approach, which is highly efficient [40]. The ECG signal is decomposed into a
finite numbers of different IMFs. The characteristic oscillation of the individual IMFs are presented on separate time scale.
 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29 27

Moreover, the analysis of IMFs together will provide instantaneous frequency information with respect to time [40]. This
helps to capture the subtle changes in the ECG signal, which could be related to a particular type of disease.
Based on the results, the proposed system performed well using DCT and DWT techniques and achieved significantly
better accuracy than EMD method.
In the case of DWT, the accuracy achieved depends on the assumption that the shape of the basis function is similar to
that of the ECG signal [60]. However, predicting the shape of the basis function is challenging especially for nonstationary
and nonlinear signal. This is because the exact time and time-scale that the signal changes are not known [56]. However, in
this work, we have used Daubechies 4 (db4) mother wavelet [57]. In the previous work by Martis et al., have classified five
arrhythmias using DCT-PCA method with highest performance [56]. Hence, in our classification also, using each ECG beat,
DCT method yielded highest classification accuracy due to its energy compaction property.
Our proposed system has the following key advantages:

(a) As summarized in Table 1, this is the first study involving the classification of three classes achieving high classifica-
tion performance using 92,273 ECG beats with just seven features.
(b) The DCT method aids in noise suppression or removal, the proposed diagnosis system is insensitive to the ECG signal
noise.
(c) The proposed system is completely automated, non-invasive, repetitive, fast and simple to use. Thus, it can be used
in hospitals to aid the clinicians in their diagnosis.
(d) The main application of our proposed method is to spot abnormal beat in a long ECG recording, so that the clinician
can further investigate the subject when a beat is classified as either CAD or MI.
(e) Proposed technique is robust and reliable due to the 10-fold cross validation technique.
(f) Another important finding of our work is that, only 133 MI and 94 CAD ECG beats are wrongly classified as normal.
This clearly shows that, our method is robust and number of false positive is very less (< 0.35%).

5. Conclusion

Cardiovascular diseases are the leading cause of death in the world. In this study, performance of diagnostic support
system developed using DCT, DWT, and EMD methods are compared in the classification of normal, CAD and MI classes.
The proposed system yielded highest classification performance of 98.5% accuracy, 99.7% sensitivity and 98.5% specificity by
using seven features using DCT technique. The proposed system provides a reliable and accurate detection of CAD and MI
(high classification performance). In addition, our technique is cost effective and fast as compared to other cardiac diagnostic
imaging tools Magnetic Resonance Imaging (MRI), Intravascular Ultrasound (IVUS), and coronary angiogram etc. The most
important finding of this work is that, we are able to detect the CAD with an accuracy of 98.75%. Only 0.23% of CAD ECG
signals and 0.33% of MI ECG beats are wrongly classified as normal ECG signals (Table 7). This performance can be further
improved by using more subjects in each class and better nonlinear features as parameters for the classifiers. In future study,
authors aim to extend their work for different stages of CAD and MI. This contributes in preventing patients from having MI
and save life significantly.

Appendix

Tables 3–5.

References

[1] M. Abo-Zahhad, ECG signal compression using discrete wavelet transform, in: Dr.Juuso T. Olkkonen (Ed.), Discrete Wavelet Transforms – Applications
and Theory, InTech, 2011, pp. 143–169.
[2] U.R. Acharya, M. Sankaranarayanan, J. Nayak, C. Xiang, T. Tamura, Automatic identification of cardiac health using modeling techniques: a comparative
study, Inf. Sci. 178 (2008) 4571–4582.
[3] U.R. Acharya, O. Faust, S. Vinitha, G. Swapna, R.J. Martis, N.A. Kadri, J.S. Suri, Linear and nonlinear analysis of normal and CAD-affected heart rate
signals, Comput. Methods Programs Biomed. 113 (2014) 55–68.
[4] U.R. Acharya, H. Fujita, V.K. Sudarshan, S.L. Oh, M. Adam, J.E.W. Koh, J.H. Tan, D.N. Ghista, R.J. Martis, C.K. Chua, C.K. Poo, R.S. Tan, Automated detection
and localization of myocardial infarction using electrocardiogram: a comparative study of different leads, Knowl.-Based Syst. 99 (2016) 146–156.
[5] U.R. Acharya, V.K. Sudarshan, J.E.W. Koh, R.J. Martis, J.H. Tan, S.L. Oh, A. Muhammad, Y. Hagiwara, M.R.K. Mookiah, K.P. Chua, K.C. Chua, R.S. Tan,
Application of higher-order spectra for the characterization of coronary artery disease using electrocardiogram signals, Biomed. Signal Process. Control
31 (2017) 31–43.
[6] P.S. Addison, Wavelet transforms and the ECG: a review, Physiol. Measure. 26 (2005) 155–199.
[7] N. Ahmed, T. Natarajan, K.R. Rao, Discrete cosine transform, IEEE Trans. Comput. 23 (1974) 90–93.
[8] M.S. Alam, N.M.S. Rahim, Compression of ECG signal based on its deviation from a reference signal using discrete cosine transform, in: IEEE, 5th
International Conference on Electrical and Computer Engineering (ICECE), 2008, pp. 53–58.
[9] S.G. Al-Kindi, F. Ali, A. Farghaly, Towards real-time detection of myocardial infarction by digital analysis of electrocardiograms, IEEE, 1st Middle East
Conference on Biomedical Engineering, 2011.
[10] V.A. Allen, J. Belina, ECG data compression using the discrete cosine transform (DCT), in: IEEE, Proceedings of Computers in Cardiology, 1992,
pp. 687–690.
[11] A.K. Akobeng, Understanding diagnostic tests 1: sensitivity, specificity and predictive values, Acta. Paediatr. 96 (2006) 338–341.
[12] American Heart Association, AHA, Heart disease and stroke statistics – 2016 update, a report from the American Heart Association (AHA), Circulation.
(2016) e2–e11.
[13] S. Arafat, M. Dohrmann, M. Skubic, Classification of Coronary Artery Disease Stress ECGs Using Uncertainty Modeling, IEEE, 2005.
28 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29

[14] M. Arif, I.A. Malagore, F.A. Afsar, Automatic detection and localization of myocardial infarction using back propagation neural networks, IEEE, 4th
International Conference on Bioinformatics and Biomedical Engineering (iCBBE), 2010.
[15] M. Arif, I.A. Malagore, F.A. Afsar, Detection and localization of myocardial infarction using k-nearest neighbor classifier, J. Med. Syst. 36 (2012) 279–289.
[16] I. Babaoglu, O. Findik, M. Bayrak, Effects of principle component analysis on assessment of coronary artery diseases using support vector machine,
Expert Syst. Appl. 37 (2010) 2182–2185.
[17] I. Babaoglu, O. Findik, E. Ulker, A comparison of feature selection models utilizing binary particle swarm optimization and genetic algorithm in deter-
mining coronary artery disease using support vector machine, Expert Syst. Appl. 37 (2010) 3177–3183.
[18] S. Banarjee, M. Mitra, ECG feature extraction and classification of anteroseptal myocardial infarction and normal subjects using discrete wavelet trans-
form, IEEE, International Conference on Systems in Medicine and Biology, 2010.
[19] S. Banerjee, M. Mitra, A classification approach for myocardial infarction using voltage features extracted from four standard ECG leads, IEEE, Interna-
tional Conference on Recent Trends in Information Systems, 2011.
[20] S. Banerjee, M. Mitra, Cross wavelet transform based analysis of electrocardiogram signals, Int. J. Electr. Electron. Comput. Eng. 1 (2012) 88–92.
[21] A. Bendifallah, R. Benzid, M. Boulemden, Improved ECG compression method using discrete cosine transform, Electronic Lett. (2011) 47.
[22] L.M. Buja, J.T. Willerson, The role of coronary artery lesions in ischemic heart disease: insights from recent clinicopathologic, coronary arteriographic,
and experimental studies, Hum. Pathol. 18 (1987) 451–461.
[23] L.M. Buja, F.J. Clubb Jr, D.W. Bilheimer, J.T. Willerson, Pathobiology of human familial hypercholesterolemia and a related animal model, the Watanabe
heritable hyperlipidaemic rabbit, Eur. Heart. J. 11 (1990) 41–52.
[24] L.M. Buja, H.A. McAllister Jr, Coronary artery disease: pathological anatomy and pathogenesis, in: J.T. Willerson, J.N. Cohn, H.J.J. Wellens, D.R. Holmes
(Eds.), Cardiovascular Medicine, third edition, Springer, London, 2007, pp. 593–610.
[25] L.M. Buja, H.A. McAllister Jr., Atherosclerosis: pathologic anatomy and pathogenesis, in: J.T. Willerson, J.N. Cohn, H.J.J. Wellens, D.R. Holmes (Eds.),
Cardiovascular Medicine, third ed., Springer, London, 2007, pp. 1581–1591.
[26] P.C. Chang, J.C. Hsieh, J.J. Lin, Y.H. Chou, C.H. Liu, A hybrid system with hidden markov models and Gaussian mixture models for myocardial infarction
classification with 12-lead ECGs, 11th IEEE Conference on Hugh Performance Computing and Communications, 2009.
[27] J. Chee, S.C. Seow, The Electrocardiogram, in: U.R. Acharya, J.S. Suri, J.A.E. Spaan, S.M. Krishnan (Eds.), Advances in Cardiac Signal Processing,
Springer-Verlag, Berlin, Heidelberg, 2007, pp. 1–53.
[28] J.S. Cheng, D.J. Yu, Y. Yang, Research on the intrinsic mode function (IMF) criterion in EMD method, Mech. Syst. Sig. Process. (2006) 817–824.
[29] I. Daubechies, The wavelet transform, time-frequency localization and signal analysis, IEEE Trans. Inf. Theory (1990) 961–1005.
[30] K.D. Desai, M.S. Sankhe, A real-time fetal ECG feature extraction using multiscale discrete wavelet transform, IEEE, 5th International Conference on
Biomedical Engineering and Informatics (BMEI), 2012.
[31] U. Desai, R.J. Martis, C.G. Nayak, G. Sheshikala, K. Sarika, R.K. Shetty, Decision support system for arrhythmia beats using ECG signals with DCT, DWT
and EMD methods: a comparative study, J. Mech. Med. Biol. 16 (2016) 1640012 (19 pages).
[32] R.O. Duda, P.E. Hart, D.G. Stork, Pattern Classification, second ed., A Wiley-Inter science Publication, 2001.
[33] D. Giri, U.R. Acharya, R.J. Martis, S.V. Sree, T.C. Lim, T.V.I. Ahamed, J.S. Suri, Automated diagnosis of coronary artery disease affected patients using
LDA, PCA, ICA and discrete wavelet transform, Knowl.-Based Syst. 37 (2013) 274–282.
[34] A.L. Goldberger, L.A.N. Amaral, L. Glass, J.M. Hausdorff, P.C. Ivanov, R.G. Mark, J.E. Mietus, G.B. Moody, C.-K. Peng, H.E. Stanley, PhysioBank, Phys-
ioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals, Circulation. 101 (20 0 0) 215–220.
[35] M. Gonzalez, C. Bergmeir, I. Triguero, Y. Rodriguez, J.M. Benitez, On the stopping criteria for k-nearest neighbor in positive unlabeled time series
classification problems, Inf. Sci. 328 (2016) 42–59.
[36] A.C. Guyton, J.E. Hall, Text Book of Medical Physiology, 11th Edition, Elsevier, New York, NY, USA, 2006.
[37] J. Han, M. Kamber, Data Mining: Concepts and Techniques Morgan Kaufmann, 2nd Edition, 2006.
[38] X. He, P. Niyogi, Locality Preserving Projection, University of Chicago, Chicago, IL, USA, 2005.
[39] N. Herring, D.J. Paterson, ECG diagnosis of acute ischemia and infarction: past, present and future, QJM 99 (2006) 219–230.
[40] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shih, Q. Zheng, N.C. Yen, C.C. Tung, H.H. Liu, The empirical mode decomposition and the Hilbert spectrum
for nonlinear and nom-stationary time series analysis, Proc. R. Soc. 454 (1998) 903–995.
[41] N. Iyengar, C.K. Peng, R. Morin, A.L. Godlberger, L.A. Lipsitz, Age-related alterations in the fractal scaling of cardiac interbeat interval dynamics, Am.
J. Physiol. 271 (1996) 1078–1084.
[42] E.S. Jayachandran, K.P. Joseph, U.R. Acharya, Analysis of myocardial infarction using discrete wavelet transform, J. Med. Syst. 34 (2010) 985–992.
[43] M.J. Katz, S.M. Ness, Coronary Artery Disease (CAD), Wild Iris Medical Education, 2015.
[44] A. Kaveh, W. Chung, Automated classification of coronary atherosclerosis using single lead ECG, IEEE Conference on Wireless Sensors, 2013.
[45] W.S. Kim, S.H. Jin, Y.K. Park, H.M. Choi, A study on development pf multi-parameter measure of heart rate variability diagnosing cardiovascular disease,
IFMBE Proc. 14 (2007) 3480–3483.
[46] M. Kumar, R.B. Pachori, U.R. Acharya, An efficient automated technique for CAD diagnosis using flexible analytic wavelet transform and entropy features
extracted from HRV signals, Expert Syst. Appl. 63 (2016) 165–172.
[47] T. Lahiri, U. Kumar, H. Mishra, S. Sarkar, A.D. Roy, Analysis of ECG signal by chaos principle to help automatic diagnosis of myocardial infarction, J. Sci.
Ind. Res. 68 (2009) 866–870.
[48] H.G. Lee, K.Y. Noh, K.H. Ryu, in: Mining Biosignal Data: Coronary Artery Disease Diagnosis Using Linear and Nonlinear Features of HRV, Springer-Verlag,
Berlin Heidelberg, 2007, pp. 218–228.
[49] H.G. Lee, K.Y. Noh, K.H. Ryu, A data mining approach for coronary heart disease prediction using HRV features and carotid arterial wall thickness, IEEE,
International Conference on Biomedical Engineering and Informatics, 2008.
[50] R. Lehtinen, H. Holst, V. Turjanmaa, L. Edenbrandt, O. Pahlm, J. Malmivuo, Artificial neural network for exercise electrocardiographic detection of
coronary artery disease, 2nd International Conference on Bioelectromagnetism, 1998 February.
[51] K. Lewenstein, Radial basis function neural network approach for the diagnosis of coronary artery disease based on the standard electrocardiogram
exercise test, Med. Biol. Eng. Comput. 39 (2001) 1–6.
[52] B. Liu, J. Liu, G. Wang, K. Huang, F. Li, Y. Zheng, Y. Luo, F. Zhou, A novel electrocardiogram parameterization algorithm and its application in myocardial
infarction detection, Comput. Biol. Med. 61 (2015) 178–184.
[53] H.L. Lu, K. Ong, P. Chia, An automated ECG classification system based on a neuro-fuzzy system, IEEE Comput. Cardiol. 27 (20 0 0) 387–390.
[54] R.J. Martis, U.R. Acharya, J.H. Tan, A. Petznick, R. Yanti, C.K. Chua, E.Y.K. Ng, L. Tong, Application of empirical mode decomposition (EMD) for automated
detection of epilepsy using EEG signals, Int. J. Neural Syst. 22 (2012) 1250027.
[55] R.J. Martis, U.R. Acharya, C.M. Lim, ECG beat classification using PCA, LDA, ICA and discrete wavelet transform, Biomed. Signal Process. Control 8 (2013)
437–448.
[56] R.J. Martis, U.R. Acharya, C.M. Lim, J.S. Suri, Characterization of ECG beats from cardiac arrhythmia using discrete cosine transform in PCA framework,
Knowl.-Based Syst. 45 (2013) 76–82.
[57] R.J. Martis, C. Chakraboty, A.K. Ray, Wavelet-based machine learning techniques for ECG signal analysis, Mach. Learn. Healthcare Inf. 56 (2014) 25–45.
[58] R.J. Martis, U.R. Acharya, H. Adeli, Current methods in electrocardiogram characterization, Comput. Biol. Med. 48 (2014) 133–149.
[59] S. Mendis, et al., Global Status Report on Non-Communicable Diseases 2014, World Health Organization, 2014.
[60] Muhidin A. Mohamed, Mohamed A. Deriche, An approach for ECG feature extraction using daubechies 4 wavelet, Int. J. Comput. Appl. 96 (2014) 36–41.
[61] J. Olkkonen, Discrete Wavelet Transforms – Theory and applications, published by InTech, Croatia, 2011.
[62] R.B. Pachori, P. Avinash, K. Shashank, R. Sharma, U.R. Acharya, Application of empirical mode decomposition for analysis of normal and diabetic
RR-interval signals, Expert Syst. Appl. 42 (2015) 4567–4581.
 U.R. Acharya et al. / Information Sciences 377 (2017) 17–29 29

[63] J. Pan, W.J. Tompkins, A real time QRS detection algorithm, 11th Edition, WB Saunders Co, Philadelphia, 2006.
[64] S. Patidar, R.B. Pachori, U.R. Acharya, Automated diagnosis of coronary artery disease using Tunable-Q wavelet transform applied on heart rate signals,
Knowl.-Based Syst. 82 (2015) 1–10.
[65] M.M. Rahhal, Y. Bazi, H. AlHichri, N. Alajlan, F. Melgani, R.R. Yager, Deep learning approach for active classification of electrocardiogram signals, Inf.
Sci. 345 (2016) 340–354.
[66] M. Ratnakar, K.S. Sunil, J. Nitisha, Signal filtering using discrete wavelet transform, Int. J. Recent Trends Eng. (2009) 2.
[67] N. Safdarian, N.J. Dabanloo, G. Attarodi, A new pattern recognition method for detection and localization of myocardial infarction using T-wave integral
and total integral as extracted features from one cycle of ECG signal, J. Biomed. Sci. Eng. 7 (2014) 818–824.
[68] L.N. Sharma, R.K. Tripathy, S. Dandapat, Multiscale energy and eigenspace approach to detection and localization of myocardial infarction, IEEE Trans.
Biomed. Eng. 62 (2015) 1827–1837.
[69] D.M. Schreck, L. Ng, B.S. Schreck, S.F. Bosco, J.R. Allegra, D. Zacharias, J.V. Wortzel, Detection of coronary artery disease from the normal resting ECG
using nonlinear mathematical transformation, Ann. Emerg. Med. 17 (1988) 132–134.
[70] E.N. Silber, N. Katz, in: Heart Disease, Macmillan publishing Co., New York, 1975, p. 498.
[71] S. Sood, M. Kumar, R.B. Pachori, U.R. Acharya, Application of empirical mode decomposition-based features for analysis of normal and CAD heart rate
signals, J. Mech. Med. Biol. 16 (2016) 1640 0 02 (20 pages).
[72] L. Sun, Y. Lu, K. Yang, S. Li, ECG analysis using multiple instance learning for myocardial infarction detection, IEEE Trans. Biomed. Eng. 59 (2012)
3348–3356.
[73] N. Townsend, K. Wickramasinghe, P. Bhatnagar, K. Smolina, M. Nichols, J. Leal, R. Luengo-Fernandez, M. Rayner, Coronary Heart Disease Statistics, a
Compendium of Health Statistics 2012 edition, British Heart Foundation, London, 2012.
[74] N. Townsend, J. Williams, P. Bhatnagar, K. Wickramasinghe, M. Rayner, Cardiovascular Disease Statistics, British Heart Foundation, London, 2014.
[75] J.T. Willerson, L.D. Hillis, L.M. Buja, Ischemic Heart Disease Clinical and Pathophysiological Aspects, Raven, New York, 1982.
[76] L. Yin, Y. Chen, W. Ji, A novel method of diagnosing coronary heart disease by analyzing ECG signals combined with motion activity, IEEE International
Workshop on Machine Learning for Signal Processing, 2011 September 18-21.