Professional Documents
Culture Documents
C28 Agus
C28 Agus
Abstract—Heart disease (cardiovascular disease) is any to determine the cluster center that will mark the average
condition that causes interference with the heart. This study location for each cluster. The output of FCM is the cluster
aims to determine the classification of heart disease based on center and the degree of membership for each data in each
phonocardiogram signals using the fuzzy system. The data cluster. The obtained information from this cluster center can
used are the heart sound recordings from patients with normal be used to build a fuzzy system.
hearts and cardiovascular abnormalities, which were recorded
using a phonocardiogram device. The signal extraction process Research on heart diseases have been carried out using
was carried out using wavelet decomposition mother Haar to various methods [4], [5] - [9]. However, electrocardiography
produce features as input variables. While the output produced (ECG) is considered as the best noninvasive arrhythmia
is a classification for heart conditions (normal or abnormal). detection method as it is a simple and noninvasive diagnostic
Furthermore, the singular value decomposition method was tool. Aside to ECG signal, a phonocardiogram (PCG) signal
utilized to determine the consequence parameters of the first- is also used for the diagnosis of heart disease. It is important
order Takagi-Sugeno-Kang (TSK) fuzzy rule. Fuzzy C-Means to note that the heart is a two-part pump, mechanical and
Clustering (FCM) was also used to optimize the number of electrical. ECG shows electrical activity and PCG shows
fuzzy rules. As for the defuzzification process, the weight mechanical activity [10]. The PCG signal is one of the non-
average method was used. The results showed that the invasive techniques for diagnosing the condition of the
accuracy and specificity of the training and testing data are human heart produced by muscle contraction and the closure
better compared to the Mamdani and the radial basis function
of the heart valve. These activities generate vibrations that
neural network (RBFNN) methods.
sound like noises and murmurs, which cardiologists can
Keywords—heart disease, phonocardiogram, wavelet, fuzzy c- analyze [11]. Some cardiac abnormality symptoms which
mean clustering, singular value decomposition cannot be analyzed by ECG, are able to be analyzed by PCG
[10].
I. INTRODUCTION
Many studies have focused on the ECG-based diagnosis
The heart is the most functional organ of the body of heart diseases. For example, the automatic diagnosis of
because of its role in pumping the blood to flow throughout congestive heart failure based on ECG signals using the
the body through blood vessels. Heart disease Convolutional Neural Network (CNN) has been carried out
(cardiovascular disease) is any condition that causes [12]. Several studies also have focused on supervised neural
interference with the heart. It is considered as the number networks, such as Multi-Layer Perceptron (MLP), Radial
one killer in the world. In 1990, there were 14.4 million Basis Function (RBF), and Learning Vector Quantization
deaths from heart attacks. This figure increased to 17.5 (LVQ), to classify ECG signals [13], [14] - [20].
million in 2005, and in 2030, it is expected to increase again
reaching 23.6 million victims [1]. Researchers have extracted various features in different
domains such as wavelet transformations to classify PCG
It is known for a fact that conventional medical diagnosis signals into various classes [21], [22] - [29]. Ari et al. have
is very difficult and requires a complex thinking process. In classified normal and abnormal heart sound using the Least
order to reduce the diagnosis time and to improve the Square Support Vector Machine (LSSVM) method based on
diagnosis accuracy, the development of a robust and reliable wavelet-based features. The results showed that the method
medical diagnosis system is, therefore, necessary. has greater accuracy than Support Vector Machine (SVM)
Classification can be carried out by using functional and classical least square [30]. Akbari et al. used the method
approach techniques known as soft computing. Soft of Digital Subtraction Phonocardiography (DSP) to detect
computing is an approach model to do computation by and classify heart murmurs [31]. Furthermore, Debbal and
imitating the reasoning and learning abilities of the human Bereksi used Continuous Wavelet Transform (CWT) to
mind in environments that have uncertainty and inaccuracy determine normal and abnormal PCG signals [32]. Prakash et
[2]. Several techniques in soft computing include fuzzy al. used four amplitude parameters of the PCG signals and
systems, Artificial Neural Network (ANN), evolutionary utilized artificial intelligence algorithms to classify heart
algorithms, and probabilistic reasoning. One method of fuzzy disease [33].
inference is the first-order Takagi-Sugeno-Kang (TSK)
method. In this method, the consequences of the fuzzy rules Classification of PCG signals using ANFIS and wavelet
are represented as linear combinations of each input. transformations has been carried out as well [34]. Lubaib and
Muneer classified heart disease based on PCG signals using
Fuzzy clustering is one of many techniques to determine various techniques, namely SVM, K-Nearest Neighbor
the optimal clusters in a vector space based on Euclidian (KNN), and Bayesian and Gaussian Mixture Model (GMM)
distance. The purpose of fuzzy clustering is to define each [35]. Ali et al. used artificial neural networks and Discrete
cluster based on its membership function [3]. The
fundamental concept of Fuzzy C-Means Clustering (FCM) is
,((( !
Wavelet Transform (DWT) for the diagnosis of heart disease The singular value decomposition of the real matrix
based on PCG signals [36]. ܣ௫ is the factorization of matrix ܣinto the multiplication
of matrices, which can be expressed as follows [40]:
In this study, a diagnosis of heart conditions was carried
out by utilizing heart sound data obtained from PCG signals.
The signal was processed through wavelet transformation to ܣൌ ܷσܸ ் (3)
be analyzed, extracted, and decomposed into several
components which were then used as input data. where U is orthogonal matrix mxm, V is orthogonal matrix
Furthermore, input data were processed using the first-order nxn, and ě is matrix mxn whose diagonal element is called
TSK fuzzy system and produced output in the form of heart singular values, namely σ ൌ diag ሺߪଵ ǡ ߪଶ ǡ ǥ ǡ ߪ ሻ and ߪଵ
disease classification. The process of constructing this fuzzy ߪଶ ڮ ߪ .
system involved FCM and singular value decomposition Furthermore, if U and V matrices are expressed as column
(SVD). vectors, ܷ ൌ ሺݑଵ ǡ ݑଶ ǡ ǥ ǡ ݑ ሻ and ܸ ൌ ሺݒଵ ǡ ݒଶ ǡ ǥ ǡ ݒ ሻ ,
II. THEORETICAL REVIEW therefore (3) can be written as (4).
!
IV. RESULTS AND DISCUSSION accuracy, sensitivity, and specifications. The following are
the results of the accuracy of several clustering.
A. Signal Preprocessing
Preprocessing is the initial step to obtain the best model. TABLE I. ACCURACY RESULTS OF SEVERAL TRIALS
This consists of 3 steps, namely signal cutting, signal Cluster Number Accuracy Sensitivity Specificity
normalization, and signal extraction. It has been previously 4 85 % 75 % 95 %
explained that the heart sound data which was recorded by 6 90 % 85 % 95 %
using the PCG device was then stored in audio format (.wav 8 100 % 100% 100%
file). However, from each recording data, only a single
heartbeat signal is needed for the analysis process, so it is
necessary to carry out the signal cutting on the obtained Based on Table I, the optimal number of clusters is 8
heartbeat sound. Cutting sound signals can be done with the clusters with its corresponding cluster center shown in Table
help of Sony Sound Forge software using the rectangle II. Hence, in this study, a fuzzy system will be constructed
cutting method. Fig. 2 displays the process of signal cutting with 8 fuzzy rules and 8 fuzzy sets in each input.
as well as its result.
TABLE II. OPTIMAL CLUSTER CENTER
No. Variable
Cluster Energy Min Max STD Target
1 31.79942 -0.21907 0.199171 0.056726 0.036468
2 26.12786 -0.16054 0.162376 0.029705 0.999714
3 59.62038 -0.20684 0.232291 0.019531 0.901047
(a)
4 37.34249 -0.34944 0.37876 0.065164 0.596377
5 66.48193 -0.46321 0.442473 0.057217 0.718326
6 40.76255 -0.44721 0.463239 0.106932 0.611135
7 43.47301 -0.35554 0.363205 0.086291 0.00203
8 49.46739 -0.26803 0.27525 0.038842 0.00012
(b)
Fig. 2. (a) PCG signal cutting process, and (b) signal cutting result
D. Classification of Heart Disease using Fuzzy System
Thereafter, the normalization process was carried out and The classification of heart disease using fuzzy system
the result, which is the normalized signal, is shown in Fig. 3. was carried out through the following steps:
1) Determining the Fuzzy Sets for the Input
a) Energy Variable (ENG). Based on the extraction
value, the minimum value of the ENG variable is 24.40, the
maximum value is 67.40, and thus, the universal set is
ܷாேீ ൌ ሾʹͶǤͶͲǡ ǤͶͲሿ. Moreover, based on Table II, the
energy variable is defined as 8 fuzzy sets, which are fuzzy
set ܩܰܧଵ ǡ ܩܰܧଶ ǡ ܩܰܧଷ ǡ ܩܰܧସ ǡ ܩܰܧହ ǡ ܩܰܧǡ ܩܰܧǡ and
଼ܩܰܧwith the membership function as follows.
ሺ௫ିଷଵǤଽଽସሻమ
ି
ߤாேீభ ሺݔሻ ൌ ݁ ଶሺଶǤହଶሻమ
Fig. 3. Normalized heartbeat signal ሺ௫ିଶǤଵଶ଼ሻమ
ି
ߤாேீమ ሺݔሻ ൌ ݁ ଶሺଶǤହଶሻమ
The normalized data were then extracted. This extraction
ڭ
process is indispensable to obtain important information ሺ௫ିସଽǤସଷሻమ
stored in the data. In this stage, the used signal ି
ߤாேீఴ ሺݔሻ ൌ ݁ ଶሺଶǤହଶሻమ
decomposition process is DWT Mother Haar at maximum
level. The DWT process is used to indicate signal b) Minimum Variable (MIN). The minimum value of
characteristics, filter data, and improve data quality. the MIN variable is -0.86, the maximum value is -0.05, and
therefore, the universal set is ܷெூே ൌ ሾെͲǤͺǡ െͲǤͲͷሿ .
B. Defining the Input and Output Variables Furthermore, based on Table II, the MIN variable is defined
In this study, 4 input variables are utilized, namely as 8 fuzzy sets, which are fuzzy set ܰܫܯଵ ǡ ܰܫܯଶ ǡ ܰܫܯଷ ǡ
energy, minimum, maximum, and standard deviation, which ܰܫܯସ ǡ ܰܫܯହ ǡ ܰܫܯǡ ܰܫܯǡ and ଼ܰܫܯwith the membership
are obtained from signal extraction using DWT Mother Haar. function as follows.
On the other hand, the output of this fuzzy system is the ሺ௫ିሺିǤଶଵଽଵሻሻమ
classification of normal heart or abnormal heart. ି
ߤெூேభ ሺݔሻ ൌ ݁ ଶሺǤସ଼ሻమ
!
൫௫ିሺିǤଶ଼ሻ൯
మ
ݕൌ σୀଵ ݓ ሺܾ ܾଵ ݔଵ ڮ ܾ ݔ ሻ (6)
ି
ߤெூேఴ ሺݔሻ ൌ ݁ ଶሺǤସ଼ሻమ
ఓభ ሺ௫భ ሻఓమ ሺ௫మ ሻǥఓሺ௫ ሻ
c) Maximum Variable (MAX). The minimum value with ݓ ൌ σಽ
సభ ఓభ ሺ௫భ ሻఓమ ሺ௫మ ሻǥఓ ሺ௫ ሻ
and the maximum value of the MAX variable are 0.05 and
0.90, respectively, therefore, the universal set is ܷெ ൌ After that, the parameter ܾ which minimizes the objective
ሾͲǤͲͷǡ ͲǤͻͲሿ . Furthermore, based on Table II, the MAX
function J is determined [41] with
variable is defined as 8 fuzzy sets, which are fuzzy set
ܺܣܯଵ ǡ ܺܣܯଶ ǡ ܺܣܯଷ ǡ ܺܣܯସ ǡ ܺܣܯହ ǡ ܺܣܯǡ ܺܣܯǡ and
ܬൌ σே ଶ ்
ୀଵሺ݀ሺ݇ሻ െ ݕሺ݇ሻሻ ൌ ሺ݀ െ ܾܺሻ ሺ݀ െ ܾܺሻ (7)
଼ܺܣܯwith the membership function as follows.
ሺ௫ିǤଵଽଽଵሻమ
ି
ଶሺǤସ଼଼ሻమ where d(k) is the actual output for the kth data pair and y(k)
ߤெభ ሺݔሻ ൌ ݁
is the first-order TSK model output for the kth data pair.
ሺ௫ିǤଵଶଷሻమ
ି
ଶሺǤସ଼଼ሻమ Furthermore, ݀ ൌ ሾ݀ሺͳሻ݀ሺʹሻ ǥ ݀ሺܰሻሿ் and X is matrix
ߤெమ ሺݔሻ ൌ ݁
sized ܰݔሾሺ݊ ͳሻܮݔሿ with N as the number of data, n as the
ڭ
ሺ௫ିǤଶହଶሻమ
number of inputs, and L as the number of rules. Moreover,
ି
ߤெఴ ሺݔሻ ൌ ݁ ଶሺǤସ଼଼ሻమ ܾ ൌ ሾܾଵ ܾଵଵ ǥ ܾଵ ǥ ܾ ܾଵ ǥ ܾ ሿ் is matrix with the size
of ሾሺ݊ ͳሻܮݔሿͳݔǤFunction (7) reaches a minimum if ݀ െ
d) Standard Deviation Variable (STD). The minimum ܾܺ ൌ Ͳ orܾܺ ൌ ݀, where X has the following form.
value of the STD variable is 0.005, while its maximum X=
value is 0.174, therefore, the universal set is ܷௌ் ൌ ሾͲǤͲͲͷǡ
ͲǤͳͶሿ. Moreover, based on Table II, the STD variable is ݓଵ ሺͳሻݓଵሺͳሻݔଵ ሺͳሻ ǥ ݓଵሺͳሻݔସ ሺͳሻ ǥ ݓସ ሺͳሻ ǥ ݓସ ሺͳሻݔସ ሺͳሻ
defined as 8 fuzzy sets, which are fuzzy set ܵܶܦଵ ǡ ܵܶܦଶ ǡ ൦
ݓଵ ሺʹሻݓଵሺʹሻݔଵ ሺʹሻ ǥ ݓଵሺʹሻݔସ ሺʹሻ ǥ ݓସ ሺʹሻ ǥ ݓସ ሺʹሻݔସ ሺʹሻ
൪
ڭ ڭ ڭ ڭ ڭ ڭ ڭ ڭ
ܵܶܦଷ ǡ ܵܶܦସ ǡ ܵܶܦହ ǡ ܵܶ ܦǡ ܵܶ ܦǡ and ܵܶ ଼ܦwith the ݓଵሺͶͲሻ ݓଵ ሺͶͲሻݔଵሺͶͲሻ ǥ ݓଵሺͶͲሻݔସ ሺͶͲሻ ǥ ݓସ ሺͺͲሻ ǥ ݓସ ሺͶͲሻݔସ ሺͶͲሻ
membership function as follows.
ሺ௫ିǤହሻమ Next, the singular value decomposition for matrix X was
ି
ߤܵܶܦభ ሺݔሻ ൌ ݁ ଶሺǤଽଽሻమ
carried out by using (3),
ሺ௫ିǤଶଽሻమ
ି
ߤܵܶܦమ ሺݔሻ ൌ ݁ ଶሺǤଽଽሻమ
ܺ ൌ ܷܸܵ ் (8)
ڭ
ሺ௫ିǤଷ଼଼ሻమ where U and V are orthogonal matrices, ܷ ൌ ሾݑଵ ǡ ǥ ǡ ݑே ሿ
ି
ߤܵܶܦఴ ሺݔሻ ൌ ݁ ଶሺǤଽଽሻమ
with the size of ܰܰݔ, ܸ ൌ ሾݒଵ ǡ ǥ ǡ ݒሺାଵሻ ሿ with the size of
2) Identifying the Fuzzy Sets for the Output ሾሺ݊ ͳሻݔܮሾሺ݊ ͳሻܮሿ , ܵ ൌ ݀݅ܽ݃ሺߪଵ ǡ ǥ ǡ ߪሺାଵሻ ሻ with the
This study uses the first-order TSK fuzzy system, size of ܰݔሾሺ݊ ͳሻܮݔሿ and singular value of ߪଵ ڮ
therefore fuzzy set is not defined in the output variable. ߪሺାଵሻ Ͳ. The optimal solution for ܾܺ ൌ ݀ is
3) Building Fuzzy Rules ௨ ௗ
Based on the results of the clustering process, the number ܾ ൌ σୀଵ ߪ ିଵ ൏ ݀ǡ ݑ ݒ ൌ σୀଵ ݒ (9)
ఙ
of optimal clusters is eight, therefore eight fuzzy rules will be
built. The fuzzy system used in this study is TSK (1). Based where r is the number of non-zero singular values.
on Table II, first-order TSK fuzzy rules can be written as Therefore, the parameter ܾ ǡ which is the entries of matrix
follows.
ܾǡ is estimated by the entries of matrix ܾ . The entries of
Rule (1) “If ݔଵ is ܩܰܧଵ and ݔଶ is ܰܫܯଵ and ݔଷ is ܺܣܯଵ and matrix ܾ are shown in Table III.
ݔସ is ܵܶܦଵ ǡ then ݕଵ ൌ ܾଵ ܾଵଵ ݔଵ ܾଵଶ ݔଶ ܾଵଷ ݔଷ
ܾଵସ ݔସ .” TABLE III. MATRIX ܾସ௫ଵ (CONSEQUENCE PARAMETERS)
ڭ Row 1-5 Row 6-10 Row 11-15 Row 16-20
Rule (8) “If ݔଵ is ଼ܩܰܧand ݔଶ is ଼ܰܫܯand ݔଷ is ଼ܺܣܯand -233.162 3.412694 -697545 239854.2
ݔସ is ܵܶ ଼ܦǡ then ଼ݕൌ ଼ܾ ଼ܾଵ ݔଵ ଼ܾଶ ݔଶ ଼ܾଷ ݔଷ 7.459094 -0.11148 48168.59 -6666.97
଼ܾସ ݔସ .” -1968.99 -8.92496 -37886.3 326387.6
-1148.93 -0.52557 -4027458 394343.9
Afterwards, the parameter ܾ is determined by using SVD -2983.76 -31.6709 -6.6E+07 -346788
method, for ݅ ൌ ͳǡ ʹǡ ǥ ǡ ͺ and ݆ ൌ Ͳǡ ͳǡ ʹǡ ͵ǡ Ͷ.
Row 21-25 Row 26-30 Row 31-35 Row 36-40
4) Defuzzification
Defuzzification is a process of converting fuzzy sets into 5.34E+09 -9.16021 429120.1 -37249.2
-1.1E+08 1.70856 -1356.35 1117.393
real numbers. In this study, the method used in the
1.85E+10 -23.0584 -156329 -5088.27
defuzzification is the weight average method [37], which is 2.36E+10 -52.7926 -50149.9 -46589.8
written as: 7.61E+09 -289.671 -4129580 -237297
σಽ
ݕൌ సభ ௬ ሺఓభ ሺ௫భ ሻఓమ ሺ௫మ ሻǥఓ ሺ௫ ሻሻ
(5) The elements of matrix ܾ are then utilized as the
σಽ
సభ ఓభ ሺ௫భ ሻఓమ ሺ௫మ ሻǥఓ ሺ௫ ሻ consequence parameters of the rules that have been built.
Hence, the generated fuzzy rules are:
Then (5) can be expressed as (6).
!
[Rule 1] “If energy is ܩܰܧଵ and minimum is ܰܫܯଵ and ܶܲ
ൌ ͲͲͳݔΨ ൌ ͳͲͲΨ
maximum is ܺܣܯଵ and standard deviation is ܵܶܦଵ ǡ then ܶܲ ܰܨ
ݕଵ ൌ െʹ͵͵Ǥͳʹ ሺǤͶͷͻ ݕ݃ݎ݁݊݁ כሻ ሺെͳͻͺǤͻͻ כ ܶܰ
݉݅݊ሻ ሺെͳͳͶͺǤͻ͵ ݔܽ݉ כሻ ሺെʹͻͺ͵Ǥ ݀ݐݏ כǤ ݀݁ݒሻ.” ൌ ͲͲͳݔΨ ൌ ͳͲͲΨ
ܶܰ ܲܨ
[Rule 2] “If energy is ܩܰܧଶ and minimum is ܰܫܯଶ and
maximum is ܺܣܯଶ and standard deviation is ܵܶܦଶ , then In addition to that, the defuzzification of testing data yielded
ݕଶ ൌ ͵ǤͶͳʹ ሺെͲǤͳͳͳͶͺ ݕ݃ݎ݁݊݁ כሻ ሺെͺǤͻʹͶͻ כ TP = 5, TN = 5, FP = 0, and FN = 0. As a result, the values
݉݅݊ሻ ሺെͲǤͷʹͷͷ ݔܽ݉ כሻ ሺെ͵ͳǤͲͻ ݀ݐݏ כǤ ݀݁ݒሻ.” of accuracy, sensitivity, and specificity are as follows:
ڭ
ൌ ͲͲͳݔΨ
[Rule 8] “If energy is ଼ܩܰܧand minimum is ଼ܰܫܯand
maximum is ଼ܺܣܯand standard deviation is ܵܶ ଼ܦǡ then ൌ ͳͲͲΨ
଼ݕൌ െ͵ʹͶͻǤʹͲͷ ሺͳͳͳǤ͵ͻ͵ ݕ݃ݎ݁݊݁ כሻ
ሺെͷͲͺͺǤʹͷ ݊݅݉ כሻ ሺെͶͷͺͻǤͷͷ ݔܽ݉ כሻ ܶܲ
ൌ ͲͲͳݔΨ ൌ ͳͲͲΨ
ሺെʹ͵ʹͻǤͳʹ ݀ݐݏ כǤ ݀݁ݒሻ.” ܶܲ ܰܨ
ܶܰ
E. Diagnosis Accuracy Test ൌ ͲͲͳݔΨ ൌ ͳͲͲΨ
ܶܰ ܲܨ
Based on the defuzzification of training data, the results
yielded True Positive (TP) = 20, True Negative (TN) = 20,
False Positive (FP) = 0, and False Negative (FN) = 0. As a Table IV presents the accuracy comparison of the
result, the values of accuracy, sensitivity, and specificity are Mamdani fuzzy system, first order TSK fuzzy system, and
as follows: RBFNN.
ൌ ͲͲͳݔΨ
ൌ ͳͲͲΨ
!
[11] A. K. Abbas, R. Bassam, and R. M. Kasim, “Mitral Regurgitan PCG- Phonocardiography Applied to The Detection and Characterization of
Signal Classification Based on Adaptive Db-Wavelet,” 4th Kuala Heart Murmurs,” BIOMED ENG. ONLINE, vol. 10, pp. 1-14, 2011.
Lumpur BMCE, Berlin Heidelberg: Springer, vol. 21, pp. 212-216, [32] S. M. Debbal, and R. F. Bereksi, “Detection of Differences of The
2018. Phornocardiogram Signals by Using The Continous Wavelet
[12] U. R. Acharya, H. Fujita, S. L. Oh, Y. Hagiwara, J. H. Tan, M. Adam, Transform Method,” Biomedical Soft Computing and Human
R. S. Tan, Deep Convolutional Neural Network for The Automated Sciences, vol. 18(2), pp. 73-81, 2012.
Diagnosis of Congestive Heart Failure Using ECG Signals, Berlin: [33] D. Prakash, M. T. Uma, K. Prabakaran, and A. Suguna, “Detection of
Springer, 2018, pp. 1-12. Heart Diseases by Mathematical Artificial Intelligence Algorithm
[13] S. Dilmac, and M. Korurek, “ECG Heart Beat Classification Method Using Phonocardiogram Signals,” International Journal of Innovation
Based on Modified ABC Alghoritm,” Appl. Soft Computing, vol. 36, and Applied Studies, vol. 3(1), pp. 145-150, 2013.
pp. 641-655, 2015. [34] A. K. Roy, and A. Misal, “Performance Evaluation of ANFIS for
[14] J. Wang, W. Chiang, Y. Hsu, Y. Yang, “ECG Arrhythmia Classification of PCG Signal Using Wavelet Transform,” IJARECE,
Classification Using a Probabilistic Neural Network with a Feature vol. 3(9), pp. 1034-1038, 2014.
Reduction Method,” Neurocomputing, vol. 116, pp. 38-45 , 2013. [35] P. Lubaib, and A. Muneer, “The Heart Defect Analysis Based on PCG
[15] M. Jayadi, AS. Abbazadeh, A. Sajedin, R. Ebrahimpour, Signals Using Pattern Recognition Techniques,” ICETEST, vol. 24,
“Classification of ECG Arrhythmia by a Modular Neural Network pp. 1024-1031, 2016.
Based on Mixture of Experts and Negatively Correlated Learning,” [36] M. N. Ali, A. Sayed, E. Dahshan, and A. Yahia, “Heart Disease
Biomed Signal Process, vol. 8, pp. 289-296, 2013. Diagnosis Using Intelligent Algorithm Based on PCG Signal
[16] Y. Sun, and A. Cheng, “Machine Learning on a Chip: A High- Analysis,” International Journal of Biology and Biomedicine, vol. 2,
Performance Low-Power Reusable Neuron Architecture for Artificial pp. 81-85, 2017.
Neural Networks in ECG Classifications,” Computer in Biology and [37] L. X. Wang, A Course in Fuzzy System and Control, New Jersey:
Medicine, vol. 42, pp. 751-757, 2012. Prentice Hall, 1997
[17] J. Kaur, and J. Raina, “An Intelligent Diagnosis System for [38] E. Cox, The Fuzzy System Handbook (A Practitioner’s Guide to
Electrocardiogram (ECG) Images Using Artificial Neural Network,” Building, Using, and Maintaining Fuzzy System), Massachussets:
International Journal of Electrical Electronics and Computer Academic Press Inc., 1994.
Engineering, pp. 47-51, 2012.
[39] D. Lay, Linear Algebra and Its Application 4th. London: Pearson
[18] M. Ansone, R. Fusco, A. Pepino, and C. Sansone, “Electrocardiogram Education, 2014.
Pattern Recognition and Analysis Based on Artificial Neural Network
and Support Vector Machines: A Review,” Journal Health Eng., vol. [40] JL. Golberg, Matrix Theory with Applications, Singapore: McGraw-
4, pp. 465-504, 2013. Hill Inc., 1991.
[41] J. Yen, L. Wang, and C. Gillespi, “Improving the Interpretability of
[19] M. Thomas, M. Das, and S. Ari, “Automatic ECG Arrhythmia
TSK Fuzzy Models by Combining Global Learning and Local
Classification Using Dual Tree Complex Wavelet Based Feature,”
Learning,” IEEE Trans. of Fuzzy Systems, vol. 6(4), pp. 530-537,
AEU International Journal of Electronics and Communication, vol.
1998.
69, pp. 715-721, 2015.
[20] X. Liu, H. Du, G. Wang, S. Zhou, and H. Zhang, “Automatic
Diagnosis of Premature Ventricular Contraction Based on Lyapunov
Exponents and LVQ Neural Network,” COMPUT. METH. PROG.
BIO, vol. 122, pp. 47-55, 2015.
[21] Z. Jiang, and S. Choi, “A Cardiac Sound Characteristic Waveform
Method for In-home Heart Disorder Monitoring with Electric
Stethoscope,” Expert Syst. Appl., vol. 31, pp. 286-298, 2006.
[22] M. Singh, and A. Cheema, “Heart Sounds Classification Using
Feature Extraction of Phonocardiography Signal,” Int. Journal
Comput. Appl. T., vol. 77, pp 13-17, 2013.
[23] Y. Chen, S. Wang, C. H. Shen, and F. K. Choy, “Intelligent
Identification of Childhood Musical Murmurs,” Journal Health Eng.,
vol. 3(1), pp. 125-139, 2012.
[24] D. Barschdor, U. Femmer, and E. Trowitzsch, “Automatic
Phonocardiogram Signal Analysis in Infants Based on Wavelet
Transforms and Artificial Neural Networks,” Comput. Cardiol.,
Vienna, Austria, pp. 753-756, 1995.
[25] T. Ohnez, Z. Dokur, “Classification of Heart Sounds Using an
Artificial Neural Network,” Pattern Recogn. Lett., vol. 24, pp. 617-
629, 2003.
[26] J. Vepa, “Classification of Heart Murmurs Using Cepstral Features
and Support Vector Machines,” IEEE 31st Annual International
Conference of the IEEE EMBS, pp. 2539-2542, September 2009.
[27] V. L. D. Avedano, L. J. I. Godino, V. M. Blanco, D. G. Castellanos,
“Feature Extraction from Parametric Time-Frequency Representations
for Heart Murmur Detection,” Annals of BIOMED Eng., vol. 38(8),
pp. 2716-2732, 2010.
[28] A. M. Gavroka, M. S. Slavkovic, M. P. Paskas, and D. M. Dujkovic,
“Joint Time-Frequency Analysis of Phonocardiograms,” 11th
Symposium on Neural Network Applications in Electrical
Engineering NEUREL, pp. 177-180, 2012.
[29] A. Cheema, and M. Singh, “Steps Involved in Heart Sound Analysis-
A Review of Existing Trends,” International Journal of Engineering
Trends and Technology, vol. 497, pp. 2921-2925, 2013.
[30] S. Ari, K. Hembram, and G. Saha, “Detection of Cardiac Abnormality
from PCG Signal Using LMS Based Least Square SVM Classifier,”
Expert Syst. Appl., vol. 37, pp. 8019-8026, 2010.
[31] M. A. Akbari, K. Hassani, J. D. Doyle, M. Navidbakhsh, M.
Sangargir, K. Bajelani, and M. Foroughi, “Digital Substraction
!