Detection of Valvular Heart Diseases Using Fourier Transform and Simple CNN Model

Abstract

In recent years, automated diagnosis of the state of health of the heart, particularly cardiac valves, has gained great success using the phonocardiogram (PCG). This work provides a low-complexity, completely automated system for diagnosing and categorizing cardiac illness based on the direct application of a multiclass Convolutional Neural Network (CNN) model, either using Softmax Classifier or KNN or SVM as a classification layer to the fast Fourier transform (FFT) of PCG signals. PCG signals are supplied into the CNN and transformed from the time domain to the frequency domain. With an analysis time of fewer than 2 seconds, the suggested technology allows us to improve performance by up to 97.66%. In the second evaluation, the methodology was evaluated on PhysioNet/Computing in Cardiology Challenge 2016 dataset achieved very high accuracy.

Full text

Abstract—In recent years, automated diagnosis of the
state of health of the heart, particularly cardiac valves,
has gained great success using the phonocardiogram
(PCG). This work provides a low-complexity, completely
automated system for diagnosing and categorizing
cardiac illness based on the direct application of a
multiclass Convolutional Neural Network (CNN) model,
either using Softmax Classifier or KNN or SVM as a
classification layer to the fast Fourier transform (FFT) of
PCG signals. PCG signals are supplied into the CNN and
transformed from the time domain to the frequency
domain. With an analysis time of fewer than 2 seconds,
the suggested technology allows us to improve
performance by up to 97.66%. In the second evaluation,
the methodology was evaluated on PhysioNet/Computing
in Cardiology Challenge 2016 dataset achieved very high
accuracy.
Index Terms— Sensors, Signal Processing,
Phonocardiogram (PCG), Heart Valves Diseases, Fast
Fourier Transform (FFT), Deep Learning
I. INTRODUCTION
With increasing industrialization and development,
cardiovascular diseases (CVDs) are becoming the
most common reason for death. CVDs cause a heavy
burden on human health and finances, especially in low
economies. Heart sounds provide essential indicators for the
condition of the human heart. Hence, they have been utilized
for the early diagnosis of CVDs. This is because of their non-
invasiveness and effectiveness in reflecting the mechanical
motion of the heart and cardiovascular system. Cardiologists
perform cardiac auscultation. It is one of the most widely used
techniques to detect abnormalities in heart sounds [1].
Accurate auscultation is critical in screening patients with
heart diseases. However, identification of pathological heart
sounds by ear is challenging as it requires extensive clinical
experience and skill and an ideal environment without
ambient noise. Besides, the human ear is not sensitive to
sounds within all frequency ranges [2]. Therefore, there is a
need for automated heart sound analysis and classification
systems that can transform heart sound signals into useful
clinical informatics tools enabling the identification of
different heart conditions.
Manuscript received December 15, 2021; revised July 02, 2022.
Wafaa N. Al-Sharu is a lecturer at the Department of Electrical
Engineering, Hashemite University, P.O. Box 330127, Zarqa 13133, Jordan.
(e-mail: wafaa.al-sharo3@hu.edu.jo).
Ali Mohammad Alqudah is a researcher at the Department of Biomedical
Systems and Informatics Engineering, Yarmouk University, P.O Box 566,
Irbid 21163, Jordan. (e-mail: ali_qudah@hotmail.com).
Shoroq Qazan is a researcher at the Department of Computer Engineering,
Yarmouk University, P.O Box 566, Irbid 21163, Jordan. (e-mail:
shoroq_qazan@hotmail.com).
Amin Alqudah is a Full Professor at the Department of Computer
Engineering, Yarmouk University, P.O Box 566, Irbid 21163, Jordan. (e-
mail: amin.alqudah@yu.edu.jo).
The field of research known as computer-aided diagnosis
(CAD) is expanding quickly. Due to the potential for
seriously misleading medical treatments caused by faults in
medical diagnostic systems, research investigations have
recently concentrated on enhancing computer-aided
diagnosis applications. In CAD, machine learning (ML) is
crucial. Diabetes, liver, dengue, hepatitis, and heart
conditions are among the illnesses that ML diagnoses [3–12].
Analyzing cardiac auscultation automatically falls under the
purview of signal processing. Segmentation, feature
extraction, and classification are the three fundamental
phases in heart sound analysis [13]. Each stage is conducted
using a variety of algorithms to accurately find abnormal
events and heart sounds. The goal of segmentation is to
identify the basic components of each cardiac cycle, such as
the first heart sound (S1), which happens during the systolic
phase, and the second heart sound (S2), which happens during
the diastolic period. The outcomes of segmenting heart
sounds based on features using machine learning techniques
have been improved. The most used segmentation method is
the Hidden Markov Model (HMM) [14, 15]. Sequences of
feature vectors derived from the original phonocardiogram
are utilized as the HMM's observation end, and sufficient
samples must be pre-labeled with the precise locations of the
S1, S2, systolic, and diastolic periods at the output end to train
the HMM. Heart sound segmentation using current deep
learning techniques produces results with higher precision
than other classification techniques [16, 17]. The goal of
feature extraction is to extract distinguishing features, either
for more accurate heart sound segmentation or for the step
after illness categorization. According to time-dominate in,
frequency-domain, and time-frequency complex domain
[18], cardiac sound qualities are dependent on these three
variables. Time-domain features include intervals of (S1 and
S2, systolic intervals, diastolic intervals) and amplitude
(mean absolute amplitude of S1 and S2 intervals). The power
spectrum of each component of the heart sound across
frequency bands is referred to as a frequency-domain feature
[16].
The classification and detection of VHDs aim to categorize
heart sounds according to distinct types of cardiac diseases.
Support vector machines [19–23], neural networks [24, 25],
HMMs [26], and other common classifiers are employed in
the classification of heart sounds. Previous research
demonstrated promising potential for detecting VHDs from
heart sounds once a suitable combination of algorithms was
employed for segmentation, feature extraction, and
classification. Aortic regurgitation, mitral regurgitation, and
pulmonary stenosis could all be distinguished by Sun's
intelligent diagnostic method with accuracy rates of 98.9%,
98.4%, and 98.7%, respectively. Thompson et al. [2] used a
murmur identification technique to separate pathogenic
murmurs from no murmurs and harmless murmurs. 603
participants' 3180 cardiac auscultations at five different chest
sites were examined. The algorithm was accurate in
Detection of Valvular Heart Diseases Using
Fourier Transform and Simple CNN Model
Wafaa N. Al-Sharu, Ali Mohammad Alqudah Member, IAENG, Shoroq Qazan, Amin Alqudah
W
IAENG International Journal of Computer Science, 49:4, IJCS_49_4_02
Volume 49, Issue 4: December 2022
______________________________________________________________________________________

identifying a pathologic murmur with an accuracy of 88%,
sensitivity of 93%, and specificity of 81%. Deep learning and
machine learning methods have recently come to be
recognized as the best methods for classifying heart sounds.
According to survey studies [2, 3], deep learning-based heart
sound categorization has higher accuracy than conventional
machine learning. To distinguish between normal and
pathological heart sounds, Milani et al. [27] employed Linear
Discriminant Analysis (LDA) and Artificial Neural Network
(ANN) techniques. After using LDA to lower the
dimensionality of the retrieved features, a single-layer ANN
model was utilized to classify the normal and abnormal PCG
signals. The findings demonstrated that frequency-domain
features alone do not consistently outperform time-domain
features in categorization. When time and frequency domain
information was integrated, the LDA/ANN technique
achieved the highest classification accuracy (93.3%). To
categorize heart illness, Nahar et al. [28] employed
straightforward machine learning techniques such as Support
Vector Machine (SVM), Decision Tree (DT), Random Forest
(RF), K-Nearest Neighbors (KNN), Artificial Neural
Network (ANN), and Naive Bays (NB). Multiple cardiac
sound signal features, including MFCC, Delta MFCC,
FBANK, and a combination of MFCC and FBANK features,
were subjected to the ML models' execution. The built-in ML
model demonstrates that the combination of MFCC and
FBANK characteristics, which was not previously used in the
literature, led to the best accuracy of 99.2%. To identify
cardiac disorders, Arora et al. [29] used digital
Phonocardiogram (PCG) data categorized as heart sounds.
They combined XGBoost with meta-heuristic algorithms like
genetic algorithm and ant colony optimization for hyper-
parameter tuning. Heart sounds were correctly classified by
the XGBoost 92.8% of the time. Additionally, the
classification accuracy of XGBoost beat DT (85.5%), RF
(90.6%), and Adaboost when the authors compared it to other
approaches (82.5%). To diagnose valve heart illness from
unsegmented phonocardiogram (PCG) signals, Khan et al.
[30] utilized a variety of methods, including cartesian genetic
programming evolved artificial neural network (CGPANN),
artificial neural network (ANN), and Support Vector
Machine (SVM). The various algorithms were trained using
time- and frequency-domain characteristics that were taken
from PCG signals that were not segmented. SVM fared better
than other techniques, with a 73%accuracy rate. Cardiologists
are trained to interpret heart sounds, and Tanmay et al. [31]
employed Wavelet Transform (WT) to extract wavelet
properties from the heart sounds. They performed a two-step
classification of heart sound quality followed by a
classification of heart pathology using bagging and boosting
trees, logistic classifiers, and SVM (i.e., normal, or
abnormal). Bagging trees were shown to be the most efficient
classification algorithm for the first stage, which comprised
the signal quality classification task. Based on its greatest
validation accuracy, the boosted trees classifier utilising
Logistic Boost was chosen for classification in the second
step, which included the task of classifying cardiac
abnormalities (77%). As we shall see, numerous research
papers have discussed the use of deep learning and machine
learning (supervised and unsupervised) algorithms in the
identification and prognosis of valvular heart disorders. The
majority of these publications treat the machine learning and
deep learning algorithms like a black box without making any
attempts to enhance them. This is a significant barrier for the
healthcare industry, which calls for more adaptable, well-
understood behaviors, and comprehensible models. The
challenge of creating interpretable machine learning models
is catching up and remains unsolved. In order to identify heart
sounds to detect heart valve disorders utilizing Fourier
transform inputs, this study will present a deep learning CNN
model.
II. METHODOLOGY
This research study aims to present a novel methodology
based on transforming PCG signals into frequency
representation using FFT that can be utilized to categorize
heart sounds in both binary and multi-class classification
scenarios. Figure 1 depicts the suggested methodology's
Block diagram. The following subsections are covered in
length in this section: the dataset utilized, the suggested
approach, the classifier, and the performance evaluation.
Fig. 1. Block Diagram of the Proposed Model.
A. Materials
In this study, one dataset was used, which is from Yaseen et
al. [32]. This dataset has five classes of normal and four heart
murmurs. All files in the dataset are stored in wav format,
with a sampling rate of 8000 Hz. The samples in this dataset
have been resized to have 24,000 samples. The dataset
utilized in this study is summarized in Table 1. Figure 2
shows a sample signal from the used datasets.
TABLE I
MULTI CLASSES DATASET INFORMATION.
Dataset
Training samples
Testing samples
Total Samples
Multiclass Dataset
700
300
1000
B. Methods
1. Fast Fourier Transform (FFT)
The discrete Fourier transform (DFT) or inverse of a signal is
calculated using a fast Fourier transform (FFT) (IDFT). When
using Fourier analysis, a signal is converted from its original
domain, which is often time or space, to a representation in
the frequency domain, and vice versa. The DFT [33] is
produced by breaking down a set of numbers into components
with different frequencies.
Fig. 2. Block Diagram of the Proposed Model.
PCG
Signals
Dataset
Calculating
FFT CNN
Model
Training
and
Testing
Performance
Evaluation
IAENG International Journal of Computer Science, 49:4, IJCS_49_4_02
Volume 49, Issue 4: December 2022
______________________________________________________________________________________

Although it is useful in many fields, this process is frequently
too slow to compute directly from the definition. An FFT can
swiftly conduct such adjustments by splitting the DFT matrix
into a product of sparse (zero) elements. It succeeds in
making the DFT computation less difficult as a result [34].
Particularly for sizable data sets with N in the hundreds of
millions, the performance disparity can be significant. In the
presence of round-off error, several FFT methods are more
accurate than directly or indirectly evaluating the DFT
specification. Several published theories, including group
theory, number theory, and simple complex-number
arithmetic, constitute the foundations for several FFT
algorithms [35]. The fields of mathematics, physics,
engineering, and music frequently make use of fast Fourier
transformations. While the basic ideas gained popularity
around 1965, several algorithms were created as early as
1805. The FFT was referred to as "the most important
numerical algorithm of our lifetime" by Gilbert Strang in
1994, and it was listed as one of the Top 10 Algorithms of the
20th Century by the IEEE Journals Computing in Science &
Engineering [36]. Because the majority of the frequency
components fall within this frequency range, the Fourier
transform of signals in this study was trimmed to only include
350 Hz from the 4000 Hz spectrum [37]. Figure 3 displays all
four signal types in their entirety.
Fig. 3. Frequency Content of Different PCG Signals.
As a result, the frequency content fed to the CNN model will
be clipped only to 350 Hz (1000 Samples). Figure 4 shows
the frequency content of different PCG signals with limited
to only 350 Hz [37, 38].
Fig. 4. Frequency Content of Different PCG Signals Limited to 350 Hz.
2. CNN Model
Due to the increasing availability of massive datasets, deep
learning is one of the most recent and cutting-edge artificial
intelligence techniques [39]. In order to achieve successive
phases of input processing, deep learning develops a
distinctive architecture made up of numerous sequential
layers [1, 6, 12–14]. The deep structure of the human brain
serves as both an inspiration for and a model for deep learning
[40]. Since the human brain has a complex internal structure
with many hidden layers, we can extract and abstract deep
features at different levels and from different perspectives.
Several deep learning algorithms have recently been
introduced [40]. CNN [39, 40] (Convolutional Neural
Network). Input, convolution, RELU, totally connected,
classification, and output are just a few of the numerous
layers that make up CNN. These layers build a CNN model
that can perform the required function. In a number of
scientific disciplines, CNN has excelled, particularly in the
medical sector [39]. The main purpose of CNN layers is to
extract comprehensive, representative, and discriminative
properties. Downsampling, feature selection, and pattern
classification will all be done in the earlier layers [40]. In the
suggested methodology, we divided the input ECG beats into
six groups using a CNN model. Eight layers make up the
model. The proposed CNN model has fewer layers than
earlier CNN models used in the literature. It is more suitable
for embedded systems due to the decreased number of layers
because it requires less time and resources to run and train the
model and identify the class of newly input PCG FFT. The
proposed CNN design is shown in Figure 5.
Fig. 5. The overall framework of CNN-based beat classification model.
3. Hybrid CNN-ML Classifier
In this work, the CNN's fully connected layer—which comes
before the Softmax layer, which is used for classification—
was used as a feature extraction layer. A feature vector with
five features, each of which represents a different sort of
class, will be the output of the entirely connected layer [39].
When the used CNN is properly constructed and trained on a
sizable dataset, the features can extract representative
features for the input data [39]. MxN is the dimension of the
retrieved features used in this study, where M is the number
of photos and N is the number of classes (in our instance, five)
[39]. a. KNN Classifier
The main objective of the K-means clustering method is to
divide data with M points and N dimensions into K clusters
while minimizing the sum of squares within each cluster. The
main idea of the clustering algorithm is to define k centers,
one for each cluster. These K-centers should be strategically
placed because different settings yield varying effects. The
next step is to link each point in a given data set with the
closest center by using the least sum of squares against all
centers [37, 38].
Input Layer
[1000 1]
Convolution
Layer
[3x3, 16]
ReLu
Layer
Fully
Connected
Layer 1 [500]
Fully
Connected
Layer 2 [250]
Fully
Connected
Layer 3 [5]
Softmax
Layer Classification
Output Layer
IAENG International Journal of Computer Science, 49:4, IJCS_49_4_02
Volume 49, Issue 4: December 2022
______________________________________________________________________________________

b. SVM Classifier
A well-known and widely used supervised machine learning
method called the Support Vector Machine (SVM) is used
largely to categorize data into two groups. By determining the
optimal hyperplane between the datasets, the SVM method
uses the input training data to build a model that predicts the
new sample class. This hyperplane must maximize the
distance between the closest data point and the separation
hyperplane. In particular in Biomedical Engineering, the
SVM has been successfully used for a variety of real-world
applications, such as face identification, recognition, and
verification, image retrieval, handwritten character, and digit
recognition [38].
4. Performance Evaluation
The confusion matrix for both binary and multi-class
scenarios was generated, followed by a comparison of the
classifier outputs with the corresponding original label of the
heart sound, in order to assess the performance of the
proposed methodology in classifying heart sounds using
instantaneous frequency features [40]. The generated
confusion matrices are used to calculate accuracy, sensitivity,
and specificity, and these values are used as a metric to judge
how precisely the classifier categorizes heart sounds. The
Equations below contain the formulas for accuracy,
sensitivity, specificity, and precision.
   
 (1)
   
 (2)
   
 (3)
   
 (4)
The false positive rate (FPR) and true positive rate (TPR),
with values ranging from 0 to 1, were given as the X and Y
axes of a receiver operating characteristic (ROC) curve that
was created to illustrate the performance of the LSTM model.
The sensitivity equation was used to determine the TPR
values, whereas the FPR was determined by deducting the
specificity value from 1. When the ROC curve was more
closely positioned to the upper left corner, the model
performed better. Although the ROC's area under the curve
(AUC) was also used, the accuracy of the model's predictions
increased with increasing AUC values [38, 37, 39, 40].
III. RESULTS
In this part, the effectiveness of the suggested technique is
evaluated. We compare all of the models that have been
offered before deciding which model is the best. All models
in this experiment are run on a computer with 16GB of RAM
and an Intel(R) Core-I5 CPU clocked at 2.3GHz. Using
adaptive moment estimation, we maximize backpropagation
using a 150-batch size, a 0.001 learning rate, and several 30-
epoch iterations.
1. CNN Model Results
In this section the result of the CNN model, Figure 6 shows
the accuracy and loss results of the training during the model
training. While Figure 7 shows the confusion matrix of the
testing data using the CNN model. Finally, Figure 8 shows
the ROC curve.
Fig. 6. The Training Accuracy and Loss using Proposed CNN Model.
Fig. 7. Confusion Matrix of Testing Dataset using Proposed CNN Model.
To summarize the results of the previous figures, Table 2
shows the performance of different classes and the overall
performance using the proposed methodology.
Fig. 8. ROC Curve of Using Proposed CNN Model.
2. CNN-SVM Model Results
In this section the result of the CNN-SVM model, Figure 9
shows the confusion matrix of the training data using the
CNN-SVM model. While Figure 10 shows the confusion
matrix of the testing data using the CNN-SVM model.
Finally, Figure 11 shows the ROC curve.
IAENG International Journal of Computer Science, 49:4, IJCS_49_4_02
Volume 49, Issue 4: December 2022
______________________________________________________________________________________

Fig. 9. Confusion Matrix of Training Dataset using Proposed CNN-SVM
Model.
Fig. 10. Confusion Matrix of Testing Dataset using Proposed CNN-SVM
Model.
Fig. 11. ROC Curve of Using Proposed CNN-SVM Model.
3. CNN-KNN Model Results
In this section the result of the CNN-KNN model, Figure 12
shows the confusion matrix of the training data using the
CNN-KNN model. While Figure 13 shows the confusion
matrix of the testing data using the CNN-KNN model.
Finally, Figure 14 shows the ROC curve.
Fig. 12. Confusion Matrix of Testing Dataset using Proposed CNN-KNN
Model.
Fig. 13. Confusion Matrix of Testing Dataset using Proposed CNN-KNN
Model.
Fig. 14. ROC Curve of Using Proposed CNN-KNN Model.
4. PhysioNet/Computing in Cardiology Challenge 2016
Dataset Results
As a second assessment of the suggested technique and CNN
model robustness, the results of applying the proposed CNN
model to the PhysioNet/Computing in Cardiology Challenge
2016 Dataset are shown in this section. One of the biggest and
most popular datasets for analyzing heart sound classification
IAENG International Journal of Computer Science, 49:4, IJCS_49_4_02
Volume 49, Issue 4: December 2022
______________________________________________________________________________________

issues is the PhysioNet/Computing in Cardiology Challenge
2016 Dataset [38]. Figure 15 displays the accuracy and loss
outcomes of the model training. The confusion matrix of the
training and testing data, respectively, using the CNN model
is shown in Figures 16 and 17. Lastly, the ROC curve is
displayed in Figure 18.
Fig. 15. The Training Accuracy and Loss using Proposed CNN Model on
the Binary Dataset.
Fig. 16. Confusion Matrix of Training Dataset using Proposed CNN Model
on the Binary Dataset.
Fig. 17. Confusion Matrix of Testing Dataset using Proposed CNN Model
on the Binary Dataset.
Fig. 18. ROC Curve of Using Proposed CNN Model on the Binary Dataset.
additionally, to conduct a thorough comparison for the second
dataset, the hybrid CNN-KNN and CNN-SVM techniques
have been used. The CNN-SVM model's output is shown in
the following figures. Figure 19 displays the confusion matrix
created from the training set of data. Figure 20 displays the
confusion matrix created by the CNN-SVM model using the
testing data. Lastly, the ROC curve is displayed in Figure 21.
Fig. 19. Confusion Matrix of Training Dataset using Proposed CNN-SVM
Model on the Binary Dataset.
Fig. 20. Confusion Matrix of Testing Dataset using Proposed CNN-SVM
Model on the Binary Dataset.
IAENG International Journal of Computer Science, 49:4, IJCS_49_4_02
Volume 49, Issue 4: December 2022
______________________________________________________________________________________

Fig. 21. ROC Curve of Using Proposed CNN-SVM Model on the Binary
Dataset.
The following figures show the result of the CNN-KNN
model, Figure 22 shows the confusion matrix of the training
data using the CNN-KNN model. While Figure 23 shows the
confusion matrix of the testing data using the CNN- KNN
model. Finally, Figure 24 shows the ROC curve.
Fig. 22. Confusion Matrix of Training Dataset using Proposed CNN-KNN
Model on the Binary Dataset.
Fig. 23. Confusion Matrix of Testing Dataset using Proposed CNN- KNN
Model on the Binary Dataset.
Fig. 24. ROC Curve of Using Proposed CNN- KNN Model on the Binary
Dataset.
IV. DISCUSSION
The goal of the proposed study is to ascertain the effects of
automatically extracting features from signals' frequency
content using a deep learning model on the categorization of
heart sound signals for multiclass heart valve situations. Our
research primarily examined how the Fourier transform-
based deep learning model affected the categorization
accuracy of heart sounds. Our solution outperforms other
methods that are currently available in the literature in the
heart sound classification scheme in terms of automated
classifier results. TABLE II
THE CLASSIFICATION REPORT OF THE PROPOSED CNN MODEL USING
TESTING SET.
Class
Accuracy
%
Sensitivity
%
Specificity
%
Precision
%
AUC
AS
100
100
99.58
98.36
1.0
MR
100
100
98.33
93.75
0.99
MS
100
100
99.17
96.77
0.99
MVP
96.66
96.66
100.00
100.00
0.99
Normal
91.66
91.66
100.00
100.00
0.97
Overall
97.66
97.66
99.42
97.77
0.99
TABLE III
THE CLASSIFICATION REPORT OF THE PROPOSED CNN-SVM MODEL USING
TESTING SET.
Class
Accuracy
%
Sensitivity
%
Specificity
%
Precision
%
AUC
AS
78.33
78.33
99.58
97.92
1.0
MR
100
100
95.83
85.72
0.99
MS
93.33
93.33
93.33
77.78
0.99
MVP
96.67
96.67
98.75
95.09
1.0
Normal
81.67
81.67
100
100
0.99
Overall
90.00
90.00
97.50
91.30
0.99
TABLE IV
THE CLASSIFICATION REPORT OF THE PROPOSED CNN-KNN MODEL USING
TESTING SET.
Class
Accuracy
%
Sensitivity
%
Specificity
%
Precision
%
AUC
AS
51.67
51.67
99.58
96.88
0.99
MR
100
100
95
83.33
0.99
MS
93.33
93.33
94.17
80
0.99
MVP
95
95
91.67
74.03
0.99
Normal
81.67
81.67
100
100
0.99
Overall
84.33
84.33
96.08
86.85
0.99
IAENG International Journal of Computer Science, 49:4, IJCS_49_4_02
Volume 49, Issue 4: December 2022
______________________________________________________________________________________

Most of the research in the literature focus on using the binary
dataset (PhysioNet/Computing in Cardiology Challenge 2016
Dataset) to evaluate their proposed methodologies. We have
already used this dataset as a secondary source of evaluation.
Table V shows a summary of the testing of the proposed
methodology on the PhysioNet/Computing in Cardiology
Challenge 2016 Dataset.
TABLE V
THE CLASSIFICATION REPORT OF THE PROPOSED MODELS USING TESTING
SET OF BINARY DATASETS.
Class
Accuracy
%
Sensitivity
%
Specificity
%
Precision
%
AUC
CNN
93.69
87.42
100
100
0.9935
CNN-SVM
96.35
95.36
97.33
97.30
0.9805
CNN-KNN
100
100
100
100
1.0
Table VI displays a comparison of the suggested
methodology's findings with those of other recent techniques
in the literature. Both of the free online datasets (PhysioNet
CinC Challenge 2016 Dataset and Pascal Dataset) or their
records were used in the majority of the studies listed and
compared in Table VI. They employ a varied number of
classes, records, and characteristics, which is obvious. These
elements have a big impact on how well the various
classification techniques work. However, more than 90% of
the approaches reported in the literature have attained high
recognition rates. TABLE VI
COMPARING BETWEEN PROPOSED METHODOLOGY AND METHODS IN
LITERATURE.
Reference
Methodology
Number of Classes
Accuracy %
[2]
CAD System
2
88
[19]
ANN
2
98.9
[27]
LDA/ANN
4
93.3
[28]
ANN
2
99.2
[29]
XGBoost
2
92.8
[30]
SVM
2
73
[31]
Logistic Boost
2
77
Proposed 5
Classes
CNN-KNN
5
84.33
CNN-SVM
5
90.00
CNN Model
5
97.66
Proposed 2
Classes
CNN-KNN
2
100
CNN-SVM
2
96.35
CNN Model
2
93.69
Table VI shows that all literature studies focused on machine
learning techniques, but none did so for deep learning. While
time domain-based deep learning models contain the time
difference between the two primary components of the heart
sound signal (S1 and S2), the extracted Fourier transform
generally concentrated on the frequency domain. The
suggested system demonstrates that the newly developed way
of feeding Fourier transform data to deep learning models
rather than time-domain signals offers greater classification
rates when compared to existing methods. A desktop
computer with an Intel Core i5-6700 processor running at 2.4
GHz and 12 GB of RAM is used to test the system's time
consumption in order to determine its real-time performance.
The system demonstrates that an average of 244.71382 mS
and 9.71287 mS, respectively, are needed to calculate the
Fourier transform for each PCG signal once the signal has
been loaded. The duration of the Fourier transforms and
classification is shown in Table VII.
TABLE VII
AVERAGE CONSUMPTION TIME FOR PROPOSED METHODOLOGY.
Process
Average Time (ms)
Total Time (ms)
FFT
244.71382
Classification
Softmax
9.71287
254.42669
KNN
20.54881
265.26263
SVM
25.22153
269.93535
V. CONCLUSION
In this study, we successfully proposed a very light and
quick deep learning model based on one-dimensional CNN
with fast Fourier transform (FFT) for automated diagnosis of
heart valve dysfunction. With the CNN model, the model
achieved an overall accuracy of 97.66% on five classes from
the PCG signals dataset, and with the CNN-KNN model, it
achieved 100% accuracy for the PhysioNet/Computing in
Cardiology Challenge 2016 dataset. The study's suggested
converting strategy and model are simple and suitable for
embedded system applications. At the same time, our method
performs better than cutting-edge networks. The results
demonstrate that the suggested network architecture is
effective in obtaining deep features from PCG signal FFTs.
By employing a bigger dataset and a larger convolution layer
kernel, the test accuracy can be improved. An effective model
with a few parameters makes up our network.
REFERENCES
[1] Maganti, K., V.H. Rigolin, M.E. Sarano, And R.O. Bonow. Valvular
Heart Disease: Diagnosis and Management. In Mayo Clinic
Proceedings. 2010. Elsevier.
[2] Thompson, W.R., A.J. Reinisch, M.J. Unterberger, And A.J. Schriefl,
Artificial Intelligence-Assisted Auscultation of Heart Murmurs:
Validation by A Virtual Clinical Trial. Pediatric Cardiology, 2019.
40(3): P. 623-629.
[3] Otoom, A.F., E.E. Abdallah, Y. Kilani, A. Kefaye, And M. Ashour,
Effective Diagnosis and Monitoring of Heart Disease. International
Journal of Software Engineering and Its Applications, 2015. 9(1): P.
143-156.
[4] Vembandasamy, K., R. Sasipriya, And E. Deepa, Heart Diseases
Detection Using Naive Bayes Algorithm. International Journal of
Innovative Science, Engineering & Technology, 2015. 2(9): P. 441-
444.
[5] Parthiban, G. And S. Srivatsa, Applying Machine Learning Methods in
Diagnosing Heart Disease for Diabetic Patients. International Journal
of Applied Information Systems (IJAIS), 2012. 3(7): P. 25-30.
[6] Iyer, A., S. Jeyalatha, And R. Sumbaly, Diagnosis of Diabetes Using
Classification Mining Techniques. Arxiv Preprint Arxiv:1502.03774,
2015.
[7] Sen, S.K. And S. Dash, Application of Meta Learning Algorithms for
The Prediction of Diabetes Disease. International Journal of Advance
Research in Computer Science and Management Studies, 2014. 2(12).
[8] Sarwar, A. And V. Sharma, Intelligent Naïve Bayes Approach to
Diagnose Diabetes Type-2. International Journal of Computer
Applications and Challenges in Networking, Intelligence and
Computing Technologies, 2012. 3: P. 14-16.
[9] Vijayarani, S. And S. Dhayanand, Liver Disease Prediction Using
SVM And Naïve Bayes Algorithms. International Journal of Science,
Engineering and Technology Research (IJSETR), 2015. 4(4): P. 816-
820.
[10] Gulia, A., R. Vohra, And P. Rani, Liver Patient Classification Using
Intelligent Techniques. International Journal of Computer Science and
Information Technologies, 2014. 5(4): P. 5110-5115.
[11] Tarmizi, N.D.A., F. Jamaluddin, A. Abu Bakar, Z.A. Othman, S.
Zainudin, And A.R. Hamdan, Malaysia Dengue Outbreak Detection
Using Data Mining Models. Journal Of Next Generation Information
Technology (JNIT), 2013. 4(6): P. 96-107.
[12] Fathima, A. And D. Manimegalai, Predictive Analysis for The
Arbovirus-Dengue Using SVM Classification. International Journal of
Engineering and Technology, 2012. 2(3): P. 521-7.
IAENG International Journal of Computer Science, 49:4, IJCS_49_4_02
Volume 49, Issue 4: December 2022
______________________________________________________________________________________

[13] Dwivedi, A.K., S.A. Imtiaz, And E. Rodriguez-Villegas, Algorithms
for Automatic Analysis and Classification of Heart Sounds–A
Systematic Review. IEEE Access, 2018. 7: P. 8316-8345.
[14] Gill, D., N. Gavrieli, And N. Intrator. Detection And Identification of
Heart Sounds Using Homomorphic Envelogram and Self-Organizing
Probabilistic Model. In Computers in Cardiology, 2005. 2005. IEEE.
[15] Schmidt, S.E., C. Holst-Hansen, C. Graff, E. Toft, And J.J. Struijk,
Segmentation of Heart Sound Recordings by A Duration-Dependent
Hidden Markov Model. Physiological Measurement, 2010. 31(4): P.
513.
[16] Potes, C., S. Parvaneh, A. Rahman, And B. Conroy. Ensemble Of
Feature-Based and Deep Learning-Based Classifiers for Detection of
Abnormal Heart Sounds. In 2016 Computing in Cardiology Conference
(CinC). 2016. IEEE.
[17] Chen, T.-E., S.-I. Yang, L.-T. Ho, K.-H. Tsai, Y.-H. Chen, Y.-F.
Chang, Y.-H. Lai, S.-S. Wang, Y. Tsao, And C.-C. Wu, S1 And S2
Heart Sound Recognition Using Deep Neural Networks. IEEE
Transactions on Biomedical Engineering, 2016. 64(2): P. 372-380.
[18] Zhang, W., J. Han, And S. Deng, Heart Sound Classification Based on
Scaled Spectrogram and Tensor Decomposition. Expert Systems with
Applications, 2017. 84: P. 220-231.
[19] Sun, S., An Innovative Intelligent System Based on Automatic
Diagnostic Feature Extraction for Diagnosing Heart Diseases.
Knowledge-Based Systems, 2015. 75: P. 224-238.
[20] Zhang, W., X. Guo, Z. Yuan, And X. Zhu, Heart Sound Classification
and Recognition Based on EEMD And Correlation Dimension. Journal
Of Mechanics in Medicine and Biology, 2014. 14(04): P. 1450046.
[21] Safara, F., S. Doraisamy, A. Azman, A. Jantan, And A.R.A. Ramaiah,
Multi-Level Basis Selection of Wavelet Packet Decomposition Tree for
Heart Sound Classification. Computers In Biology and Medicine, 2013.
43(10): P. 1407-1414.
[22] Kwak, C. And O.-W. Kwon, Cardiac Disorder Classification by Heart
Sound Signals Using Murmur Likelihood and Hidden Markov Model
State Likelihood. IET Signal Processing, 2012. 6(4): P. 326-334.
[23] Kumar, D., P. Carvalho, M. Antunes, R. Paiva, And J. Henriques. Heart
Murmur Classification with Feature Selection. In 2010 Annual
International Conference of The IEEE Engineering in Medicine and
Biology. 2010. IEEE.
[24] Uğuz, H., A Biomedical System Based on Artificial Neural Network
and Principal Component Analysis for Diagnosis of The Heart Valve
Diseases. Journal Of Medical Systems, 2012. 36(1): P. 61-72.
[25] Ölmez, T. And Z. Dokur, Classification of Heart Sounds Using an
Artificial Neural Network. Pattern Recognition Letters, 2003. 24(1-3):
P. 617-629.
[26] Fahad, H., M.U. Ghani Khan, T. Saba, A. Rehman, And S. Iqbal,
Microscopic Abnormality Classification of Cardiac Murmurs Using
ANFIS And HMM. Microscopy Research and Technique, 2018. 81(5):
P. 449-457.
[27] Milani, M., P.E. Abas, L.C. De Silva, And N.D. Nanayakkara,
Abnormal Heart Sound Classification Using Phonocardiography
Signals. Smart Health, 2021. 21: P. 100194.
[28] Khalid, M.O.N., M.A.-H. Obaida, A.-E. Ashraf, And G. Nasr,
Phonocardiogram Classification Based on Machine Learning with
Multiple Sound Features. Journal Of Computer Science, 2020. 16(11).
[29] Arora, V., R. Leekha, R. Singh, And I. Chana, Heart Sound
Classification Using Machine Learning and Phonocardiogram. Modern
Physics Letters B, 2019. 33(26): P. 1950321.
[30] Khan, N.M., M.S. Khan, And G.M. Khan, Automated Heart Sound
Classification from Unsegmented Phonocardiogram Signals Using
Time Frequency Features. International Journal of Computer and
Information Engineering, 2018. 12(8): P. 598-603.
[31] Gokhale, T. Machine Learning Based Identification of Pathological
Heart Sounds. In 2016 Computing in Cardiology Conference (Cinc).
2016. IEEE.
[32] Yaseen; Son, G.-Y.; Kwon, S. Classification of Heart Sound Signal
Using Multiple Features. Appl. Sci. 2018, 8, 2344.
[33] Nussbaumer, H.J., 1981. The Fast Fourier Transform. In Fast Fourier
Transform and Convolution Algorithms (Pp. 80-111). Springer, Berlin,
Heidelberg.
[34] Heckbert, P., 1995. Fourier Transforms and The Fast Fourier
Transform (FFT) Algorithm. Computer Graphics, 2, Pp.15-463.
[35] Brigham, E.O., 1988. The Fast Fourier Transform and Its Applications.
Prentice-Hall, Inc.
[36] Brigham, E.O. And Morrow, R.E., 1967. The Fast Fourier
Transform. IEEE Spectrum, 4(12), Pp.63-70.
[37] Alqudah, A.M., Alquran, H. And Qasmieh, I.A., 2020. Classification
Of Heart Sound Short Records Using Bispectrum Analysis Approach
Images and Deep Learning. Network Modeling Analysis in Health
Informatics and Bioinformatics, 9(1), Pp.1-16.
[38] Alqudah, A.M., 2019. Towards Classifying Non-Segmented Heart
Sound Records Using Instantaneous Frequency Based
Features. Journal Of Medical Engineering & Technology, 43(7),
Pp.418-430.
[39] Alqudah, A. And Alqudah, A.M., 2021. Artificial Intelligence Hybrid
System for Enhancing Retinal Diseases Classification Using
Automated Deep Features Extracted from OCT Images. International
Journal of Intelligent Systems and Applications in Engineering, 9(3),
Pp.91-100.
[40] Alqudah, A.M., Qazan, S., Al-Ebbini, L., Alquran, H. And Qasmieh,
I.A., 2021. ECG Heartbeat Arrhythmias Classification: A Comparison
Study Between Different Types of Spectrum Representation and
Convolutional Neural Networks Architectures. Journal Of Ambient
Intelligence and Humanized Computing, Pp.1-31.
Wafaa Al-Sharu received her MSc from Jordan University of Science &
Technology in 2008 and BSc from Mutah University in 2001. Her research
areas are in the field of Signal Processing and Analysis and Artificial
Intelligence. Currently, she is working at the Department of Electrical
Engineering, Hashemite University, Zarqa, Jordan. She has extensive
experience in teaching electrical engineering courses like signals and
systems, electrical circuits, electrical machines, and electromagnetic and
communication systems in different universities in Jordan like Jordan
University of Science & Technology, Yarmouk University, and AlBalqa
Applied University. She also served on different scientific committees at the
Department of Electrical Engineering, Hashemite University, Zarqa, Jordan,
and the Department of Telecommunication Engineering, Yarmouk
University, Irbid, Jordan.
Ali Mohammad Alqudah, MSc received his B.Sc. and M.Sc. both from
Yarmouk University in 2015 and 2018 respectively. His research area is in
the field of Biomedical Signal Processing, Image Processing and Analysis,
Deep Learning, and Machine Learning. Alqudah has published high-quality
research articles in journals and conferences. In 2021, he was listed on
Stanford's list of the top 2% scientists in the world. Ali serves as a reviewer
for several peer-reviewed journals. Currently, he is working towards his
Ph.D. and working as Graduate Research Assistant and Ph.D. student at the
Biomedical Engineering Program, University of Manitoba, Winnipeg,
Canada.
Shoroq Qazan, MSc received her B.Sc. and M.Sc. both from Yarmouk
University in 2017 and 2022 respectively. Her research area is in the field of
Biomedical Signal and Image Processing, Deep Learning, Machine
Learning, and Brain Signal Processing. She published several high-quality
research articles.
Amin Alqudah received his M.Sc. and Ph.D. in Electrical Engineering from
the University of Colorado, USA in 2005, and from Colorado State
University, USA in 2009, respectively. He received his bachelor’s degree in
communications engineering from Yarmouk University, Jordan in 1999.
Since 2009, he has been working with the Department of Computer
Engineering, Hijjawi Faculty for Engineering Technology, Yarmouk
University, Jordan. His research interests include Image Processing, Neural
Networks, Machine Learning, and adaptive signal processing.
IAENG International Journal of Computer Science, 49:4, IJCS_49_4_02
Volume 49, Issue 4: December 2022
______________________________________________________________________________________