1D модель CNN для діагностики ЕКГ на кількох класифікаторах
One of the main reasons for human death is diseases caused by the heart. Detecting heart diseases in the early stage can stop heart failure or any damage related to the heart muscle. One of the main signals that can be beneficial in the diagnosis of diseases of the heart is the electrocardiogram (EC...
Gespeichert in:
| Datum: | 2022 |
|---|---|
| Hauptverfasser: | , , , , |
| Format: | Artikel |
| Sprache: | Englisch |
| Veröffentlicht: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2022
|
| Schlagworte: | |
| Online Zugang: | https://journal.iasa.kpi.ua/article/view/265099 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Назва журналу: | System research and information technologies |
| Завантажити файл: |  |
Institution
System research and information technologies| _version_ | 1866391923951927296 |
|---|---|
| author | Bassiouni, Mahmoud Hegazy, Islam Rizk, Nouhad El-Dahshan, El-Sayed Salem, Abdelbadeeh |
| author_facet | Bassiouni, Mahmoud Hegazy, Islam Rizk, Nouhad El-Dahshan, El-Sayed Salem, Abdelbadeeh |
| author_sort | Bassiouni, Mahmoud |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2023-05-21T20:04:38Z |
| description | One of the main reasons for human death is diseases caused by the heart. Detecting heart diseases in the early stage can stop heart failure or any damage related to the heart muscle. One of the main signals that can be beneficial in the diagnosis of diseases of the heart is the electrocardiogram (ECG). This paper concentrates on the diagnosis of four types of ECG records such as myocardial infarction (MYC), normal (N), variances in the ST-segment (ST), and supraventricular arrhythmia (SV). The methodology captures the data from six main datasets, and then the ECG records are filtered using a pre-processing chain. Afterward, a proposed 1D CNN model is applied to extract features from the ECG records. Then, two different classifiers are applied to test the extracted features’ performance and obtain a robust diagnosis accuracy. The two classifiers are the softmax and random forest (RF) classifiers. An experiment is applied to diagnose the four types of ECG records. Finally, the highest performance was achieved using the RF classifier, reaching an accuracy of 98.3%. The comparison with other related works showed that the proposed methodology could be applied as a medical application for the early detection of heart diseases. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2022.4.01 |
| first_indexed | 2025-07-17T10:27:57Z |
| format | Article |
| fulltext |
Mahmoud M. Bassiouni, Islam Hegazy, Nouhad Rizk, El-Sayed A. El-Dahshan, Abdelbadeeh M.
Salem, 2022
Системні дослідження та інформаційні технології, 2022, № 4 7
TIДC
ПРОБЛЕМИ ПРИЙНЯТТЯ РІШЕНЬ ТА
УПРАВЛІННЯ В ЕКОНОМІЧНИХ, ТЕХНІЧНИХ,
ЕКОЛОГІЧНИХ І СОЦІАЛЬНИХ СИСТЕМАХ
UDC 62-50
DOI: 10.20535/SRIT.2308-8893.2022.4.01
1D CNN MODEL FOR ECG DIAGNOSIS BASED
ON SEVERAL CLASSIFIERS
MAHMOUD M. BASSIOUNI, ISLAM HEGAZY, NOUHAD RIZK, EL-SAYED A.
EL-DAHSHAN, ABDELBADEEH M. SALEM
Abstract. One of the main reasons for human death is diseases caused by the heart.
Detecting heart diseases in the early stage can stop heart failure or any damage re-
lated to the heart muscle. One of the main signals that can be beneficial in the diag-
nosis of diseases of the heart is the electrocardiogram (ECG). This paper concen-
trates on the diagnosis of four types of ECG records such as myocardial infarction
(MYC), normal (N), variances in the ST-segment (ST), and supraventricular ar-
rhythmia (SV). The methodology captures the data from six main datasets, and then
the ECG records are filtered using a pre-processing chain. Afterward, a proposed 1D
CNN model is applied to extract features from the ECG records. Then, two different
classifiers are applied to test the extracted features’ performance and obtain a robust
diagnosis accuracy. The two classifiers are the softmax and random forest (RF) clas-
sifiers. An experiment is applied to diagnose the four types of ECG records. Finally,
the highest performance was achieved using the RF classifier, reaching an accuracy
of 98.3%. The comparison with other related works showed that the proposed
methodology could be applied as a medical application for the early detection of
heart diseases.
Keywords: Electrocardiogram (ECG), Continuous wavelet transform (CWT), 1D
convolutional neural network (CNN) model.
INTRODUCTION
Heart diseases are one of the main reasons for death worldwide and they are
sometimes called cardiovascular diseases (CVD). Various people suffer and die
from heart diseases annually based on recent research and survey studies. In 2022
[1], it is estimated that about 17.9 million people died from CVD, and this repre-
sents about 32% of the global death, and about 85% of these people have died
from heart attack and stroke. Moreover, CVD was responsible for 38% of all
premature deaths (under the age of 70) due to non-communicable diseases.
About 3 quarters of the deaths caused by CVD occur in the low-and middle-
income countries. Arrhythmia is one of the salient groups of CVDs. They repre-
sent the abnormal electrical conduction or impulse origin in the heart. Most of the
arrhythmias are non-life-threatening, while some of them can cause many cardio-
Mahmoud M. Bassiouni, Islam Hegazy, Nouhad Rizk, El-Sayed A. El-Dahshan, Abdelbadeeh M. Salem
ISSN 1681–6048 System Research & Information Technologies, 2022, № 4 8
vascular complications and sudden death. The early diagnosis of arrhythmia can
assist in preventing sudden death and help in treating many different cardiovascu-
lar diseases. Physicians, experts, and doctors detect arrhythmias based on electro-
cardiograms (ECG) signals. The ECG measures the variations in the electrical
potential in one cycle of the heartbeat. A single ECG signal consists of a group of
peaks defined by , , ,P Q R S and T . Moreover, various types of arrhythmia do not
appear in a short time and may require a large amount of ECG heartbeats. As a
result, a diagnosis method automated should be investigated for the identification
of different ECG records and this is the main focus of the proposed methodology.
Several methodologies based on machine learning have been built for ex-
tracting features and classifying ECG records. On one hand, extracting features
from ECG signals is essential before the classification process because it provides
a great impact on the results of the classification. P-QRS-T segment and RR in-
terval were used in almost every research [2]. In addition to this, there are other
conventional features extracted from the ECG based on morphological features,
wavelet transform features, higher-order statistics, random projection, and wave-
let packet entropy. These methodologies require providing a hand-crafted feature
before applying any conventional classifier. There are several disadvantages in
these processes of feature extraction which are depthless, large time-consuming
and they lack any implicit knowledge. On the other hand, several numbers of
classifiers were applied such as a k-nearest neighbor, artificial neural network,
support vector machine, random forest, and Gaussian mixture models. When
these conventional features are fit to these conventional classifiers they suffer
from overfitting obstacles. The main causes for overfitting are as follows: (i)
noise and unclean data used for training (ii) high variance and complexity of the
model (iii) size of the training set is not enough (iv) learning from a small dataset.
Deep learning (DL) is preserved to be part of machine learning. It is known by the
word “deep” because the network structure consists of many hidden layers [3].
The main concept in DL is that the low-level features are integrated to obtain
high-level features. In DL no hand-craft features are obtained and implicit knowledge
can be learned easily. DL has also been used in some of the ECG studies, and it
showed excellent classification results in diagnosis. Several DL structures were
used such as recurrent neural network, stacked de-noising auto-encoder, deep
neural network, convolutional neural networks, and restricted Boltzmann
machine. Finally, based on the advantages of the DL the proposed methodology
used the merits of the DL and delivered the following contributions.
The contributions stemming from this paper are two-main folds:
1. The proposed DL is used to diagnose four main ECG records based on
balanced datasets of records.
2. Development of a proposed 1D CNN model for the diagnosis of several
ECG diseases.
The manuscript is summarized based on different sections. Related works
are presented in section 2, while the proposed methodology in terms of capturing
ECG records, filtering the ECG signals, extracting features, and classifying the
ECG records is presented in section 3. Moreover, the main results and the discus-
sion are illustrated in sections 4 and 5 respectively. Finally, section 6 manifests
the conclusion and the future directions.
1D CNN model for ECG diagnosis based on several classifiers
Системні дослідження та інформаційні технології, 2022, № 4 9
RELATED WORK
Various approaches are applied for the diagnosis of ECG signals. These ap-
proaches depend on machine learning and deep learning methodologies and tech-
niques. Some methods deal with the ECG signals in the form of 1D signal and
other methods convert the 1D ECG signals to 2D images using several techniques
such as fast Fourier transform and wavelet transforms. Moreover, numerous stud-
ies applied 1D CNN models for extracting feature from the ECG records and
heartbeats. A study presented by L.A. Abdullah et al. [4] for the diagnosis of ECG
signals. The proposed model is based on a 1D CNN model for learning features,
and the results are fed to a long short term memory (LSTM). The 1D CNN model
consists of 4 (1D) convolutional and 2 fully connected layers, while the LSTM
model consists of 2 LSTM and 2 fully connected layers. Two main datasets were
used in their study which are MIT-BIH arrhythmia and PTB diagnostic datasets.
The CNN-LSTM model has achieved an accuracy of 98.1% and 98.66% in the
diagnosis of myocardial infraction (MYC) and other arrhythmia respectively.
Another study presented by E. Butun. et al. [5] for detecting various heart
diseases using ECG signals. The methodology is based on 1D version of capsule
networks (CapsNet). The 1D CapsNet model consist of several layers based on
convolutional and fully connected layers. The model starts with 1 input and 2
(1D) convolutional layers. Then, the model consists of 1D convolutional, 1 re-
shape, and 1 squash layers. Afterwards, the output of the squashing is input to an
ECG caps. The ECG caps consists of a masking layer and 3 fully connected lay-
ers. Finally, the model ends with CapsNet for the ECG diagnosis. The model clas-
sifies normal and coronary artery diseases (CAD) using 5-fold cross validation
achieving an accuracy of 99.44% and 98.62% for 2s and 5s ECG segments re-
spectively. In addition to this, a study presented by X. Hau et al. [6] for the diag-
nosis of several ECG diseases. The methodology proposed is based on pre-
processing, data augmentation, and data segmentation using R-R-R strategy. The
data were selected from the MIT-BIH arrhythmia, and the number of ECG heart-
beats selected are normal, left bundle block beat (LBBB), right bundle block beat
(RBBB), premature ventricular contraction, and the paced beat. Then, the features
are extracted using a proposed 1D CNN model. The model consists of 3 convolu-
tional, 3 pooling, 1 fully connected layer, and 1 classification layer. It was tested
on 5 classes of ECG heart beats, and the results using accuracy, area under the
curve (AUC), sensitivity, and F1-score performance measurements have achieved
0.9924, 0.9994, 0.99, and 0.99 respectively.
Moreover, a study provided by G. Petmeza et al. [7] for the diagnosis of
ECG diseases. The methodology proposed relies on Butterworth filter for ECG
signal de-noising. In addition to that, an improved version of cross-entropy loss to
solve the problem of unbalanced data. Then, the R peaks and the beats of the ECG
signals are separated before extracting features. Also, the features are extracted
from a hybrid model depending on 1D CNN layers and LSTM layers. The model
consists of 3 (1D) convolutional, 3 max pooling, 1 LSTM, and 1 dense layers.
Ten-fold cross validation is applied to denoise normal, atrial fibrillation (AFID),
atrial flutter (AFL), and AV junctional rhythm (J). The data was obtained from
the MIT-BIH atrial fibrillation database, and the model achieved a sensitivity and
specificity of 97.87% and 99.29% respectively. Finally, it can be seen from the
previous works that various types of 1D CNN models were proposed, and the re-
sults achieved or resulted from them was robust in performance. Therefore, it was
recommended to develop a 1D CNN model for diagnosis of ECG records.
Mahmoud M. Bassiouni, Islam Hegazy, Nouhad Rizk, El-Sayed A. El-Dahshan, Abdelbadeeh M. Salem
ISSN 1681–6048 System Research & Information Technologies, 2022, № 4 10
METHODOLOGY
The methodology consists of four main stages which are obtaining ECG data, fil-
tering ECG signals, extracting various ECG features, and classifying ECG re-
cords. In the data acquisition phase, six main data sets online are downloaded that
consist of four different ECG heartbeats. In the second stage, filtering or de-
noising is performed on the ECG heartbeats as the ECG signal consists of three
main common noises which are line drifting, power interference, and noise based
on a high frequency.
These distortions are removed using wavelets and a set of filters. The next
stage is to pass the filtered ECG records to a proposed 1D CNN model for feature
extraction. Finally, in the classification phase, two different classifiers are em-
ployed for ECG diagnosis which are Softmax and Random Forest (RF) as shown
in Fig. 1.
Data Acquisition
This stage is one of the most important stages in the methodology proposed.
There exist two concepts for capturing the ECG signals. The first concept is the
application of a medical device for capturing the ECG heartbeats at different
leads, whereas the second way is to download an available ECG signal online. In
this methodology, the second way is employed, and the ECG signals are captured
from six datasets which are Normal Sinus Rhythm Database (nsrdb) [8], Normal
Sinus Rhythm RR Interval Database (nsr2db) [9], MIT-BIH Supraventricular ar-
rhythmia database (svdb) [10], MIT-BIH ST change database (stdb) [11], Long
Term ST database (ltsdb) [12] and PTB diagnostic ECG [13], where the number
of ECG records in the former datasets are 18, 54, 84, 28, 86, and 549 respectively.
Four main different ECG heartbeats were chosen from these datasets.
Pre-processing
ECG signals are always nested in it some distortions and noises that are produced
from variant origins. The main three types of noises concentrated in the ECG sig-
nals are the drifting in the baseline of ECG signal, interference in the power line,
high noise frequency in the main components of the signal, and in some cases, a
combination between these types of noises can be found. As a result, a pre-
processing chain of filters and wavelets is developed to eliminate these noises by
saving the main information of the signal. The pre-processing chain should be
summarized in three main tasks which are correcting the drifting in the ECG
signal, reducing the interference, selecting low-frequency components, and en-
Feature
Extraction
• Convolution
Neural Network
based on
a proposed 1D
CNN model
Classifiers
• SoftMax
• Random
Forest (RF)
Raw ECG Signal Types
• Nwormal (N)
• Supraventricular
arrhythmia (SV)
• ST-segments
changes (ST)
• Myocardia Infraction
(MYC)
Preprocessing
• Wevelet Transform
• Adaptive band stop
filter
• low pass Batter worth
filter
• Smoothing
Evaluation
metrics
• Accuracy
• Precision
• F-measure
• Sensitivity
• Specificity
Fig. 1. Proposed Overall Methodology
1D CNN model for ECG diagnosis based on several classifiers
Системні дослідження та інформаційні технології, 2022, № 4 11
hancing the overall signal [14]. The chain contains four main stages based on
wavelet drift correction, adaptive band stop filter, low pass filter, and smoothing.
The baseline drift is removed by applying wavelet decomposition with db8 and a
decomposition level equal to 9.
Then, the powerline interference is removed using an adaptive band stop fil-
ter with a stopband frequency corner sW equals to 50Hz. In addition to this, high
frequency located in the ECG signals is removed using a low pass Butterworth
filter with a passband frequency corner and a stopband frequency corner equal to
40Hz and 60Hz respectively. The values for the passband ripple and stopband
ripple attenuation are 0.1dB and 30dB respectively [15]. Finally, a smoothing fil-
ter based on Savitzky–Golay (SG) is applied to remove the remaining noise with a
smoothing value equal to 5.
Feature Extraction
Convolutional Neural Networks based on 1D-CNN Model. The datasets faced
in this study does not have any previous information or knowledge about the fea-
tures that can be extracted from them. Also, it may be very difficult to extract ro-
bust features using traditional machine learning or feature engineering techniques.
It is recommended to learn information or features automatically using deep
learning. Therefore, the CNN model is developed to obtained robust features from
the ECG records [16].
The main core of the CNN model is the convolutional layers because these
layers work by applying a convolution operation between the local and filter
regions of the input. It is also known that CNN models are designed for two-
dimensional data that appear in most cases in the form of images. The proposed
1D CNN model depends mainly on convolutional layers at the beginning of the
model and at the middle of the model. The convolutional layers at the beginning
extracts low level features from the ECG signals that can appear in the form of
sudden variances, while the convolutional layers in the middle extracts high level
and more abstract features related to the ECG signals.
To use the CNN model that relies on the convolutional layers for 1D signals,
the convolutional layers must be redesigned to match the input. The proposed
CNN model consists of an input layer, 3 convolutional layers, 3 ReLU layers,
3 batch normalization layers, 3 max pooling layers, and ending with 3 fully
connected layers [17]. The structure of the proposed 1D CNN is shown in Fig. 2,
and the details (Filter size, Stride, Padding) of each layers in the proposed 1D
CNN are shown in table 1.
T a b l e 1 . The Whole Parameters of the proposed 1D CNN Model
Layer
No Layers Name Activations Learnables Parameters
1 Input Layer 1 65536 1 – Input Size = [1 65536 1]
Normalization = “zero center”
2 Convolutional 1 4 32 W = [1 23 1 32]
B = [1 1 32]
Filter Size = [1 32]
No. Filters = 32
Stride = [3 3]
Padding = [0 0 0 0]
3 Activation 1 4 32 – Function = “ReLU”
4 Batch
Normalization 1 4 32 Offset = 1 1 16
Scale = 1 1 16
Mahmoud M. Bassiouni, Islam Hegazy, Nouhad Rizk, El-Sayed A. El-Dahshan, Abdelbadeeh M. Salem
ISSN 1681–6048 System Research & Information Technologies, 2022, № 4 12
Continued table 1
Layer
No Layers Name Activations Learnables Parameters
5 Max Pooling 1 1 32 –
Pool Size = [1 2]
Stride = [1 1]
Padding = [ 0 0 0 0]
6 Convolutional 1 1 32
W = [1 � 23 1 32]
B = [1 1 32]
Filter Size = [1 32]
No. Filters = 32
Stride = [2 2]
Padding = [0 0 0 0]
7 Activation 1 4 32 – Function = “ReLU”
8
Batch
Normalization 1 4 32
Offset = 1 1 16
Scale = 1 1 16
9 Max Pooling 1 1 32 –
Pool Size = [1 2]
Stride = [1 1]
Padding = [ 0 0 0 0]
10 Convolutional 1 1 16
W = [1 23 1 32]
B = [1 1 32]
Filter Size = [1 16]
No. Filters = 16
Stride = [1 1]
Padding = [0 0 0 0]
11 Activation 1 4 16 – Function = “ReLU”
12
Batch
Normalization 1 4 16
Offset = 1 1 16
Scale = 1 1 16
13 Max Pooling 1 1 16 –
Pool Size = [1 2]
Stride = [1 1]
Padding = [ 0 0 0 0]
14
Fully
Connected 1 1 100
W = [100 16]
B = [100 1]
Output Size = 100
15
Fully
Connected 1 1 4
W = [100 4]
B = [100 1]
Output Size = 4
Fig. 2. The proposed 1D CNN model for ECG records Diagnosis
1D CNN model for ECG diagnosis based on several classifiers
Системні дослідження та інформаційні технології, 2022, № 4 13
Classification
This step is the final step of the proposed methodology in which the result of the
diagnosis will be determined with average accuracy. Two main classifiers are de-
termined to examine the performance of the methodology, and these classifiers
are Softmax, Random forest (RF), and XGBoost classifier.
Softmax Classifier. Softmax is known as a multinomial logistic regression
and it is well accepted in statistical mathematics as it is applied to classify a
categorical class placement. Softmax gives a more intuitive classification output
and probabilistic interpretation [18]. For instance, let us assume that 4 classes are
presented, the Softmax classifier will have 4 main nodes generated and defined by
iP , where 1,2,3,4i . These probabilities depend on a discrete target function,
and these probabilities are input to the Softmax classifier in the form of the
following equation:
i k ki
k
S Y t ,
where y is the activation produced from the nodes found in the last layers, and
the t is the weight that joins the last layers of nodes to the last layers in the DL
model [64].
The probability of the iS will be defined using the following equation:
)(exp
)(exp
2
i
i
i
S
S
P
j
.
The predicated class i will be obtained by the following equation:
)(maxarg iPi .
Random Forest (RF) Classifier. Random Forests (RF) are one of the
powerful ensemble learning methods. RF was developed to overcome the
drawbacks of decision trees (DT). The major disadvantage of the DT is the high
variance. In order words, it is not natural that a small variance in the training data
can lead to a major change in the structure of a decision tree. This makes the
decision trees as a classifier largely unstable in comparison to other decision
predictors. Also, if an error happens in a node that is near the root, it propagates to
the leaves of the tree. This leads to different and worse classification results.
Therefore, the classifier of the random forest is invented by Breiman [19]. RF is
built based on the combination of various decision trees. It integrates the output
obtained from each separate decision tree to generate the final result. In addition
to that, RF relies on uncorrelated decision trees. In other words, if similar decision
trees are used in the forest, then the overall result will not vary so much and it will
be equivalent to the result of a single decision tree. To achieve the concept of un-
correlated decision trees in RF features randomness and bootstrapping are applied.
Random forests work considering a learning set known by ),,(( 11 YXL
,( , ))i iX Y designed with i vector. Where X is a set of features and samples
and the Y is the set of labels. In the classification problems, RF maps X to Y
and new input features are recognized by each tree of the forest. Then, each tree
produces a specific classification result and the decision forest selectsthe classifi-
cation based on the most votes obtained over all the trees in the forest.
The training of the RF is achieved relying on the result obtained from each
decision tree. The training data is distributed randomly based on drawing N
examples with a special kind of replacement in which the N is considered the
Mahmoud M. Bassiouni, Islam Hegazy, Nouhad Rizk, El-Sayed A. El-Dahshan, Abdelbadeeh M. Salem
ISSN 1681–6048 System Research & Information Technologies, 2022, № 4 14
original size of the training data. The learning method produces a classifier
obtained from various trials and then the classifiers are gathered together to form
the final classifier. In the classification stage, each classifier starts to record a vote
for the class to which it belongs and the feature is drawn to the class with the
highest votes.
EXPERIMENTAL RESULTS
The experimental result was reached using the proposed model based on two
main classifiers which are SoftMax, and RF. The deep learning model was im-
plemented using MATLAB software. An experiment is applied based on the pro-
posed methodology for the diagnosis of four different ECG records. The whole
experiment is performed on a computer with Intel (R) Core i7-8565U CPU of
1.99 GHz, 12 GB memory, and NVIDIA graphical card with GM 310M. The total
number of records selected from 6 datasets for the four types of ECG heartbeats is
294 records. These records are collected as follows: 72 normal records (NSR)
from the first two datasets (18 from nsrdb) and (54 from nsr2db), 74 supraven-
tricular arrhythmias (SV) records from the third dataset (svdb), 74 records repre-
senting ST-segment changes from the fourth and the fifth datasets (28 from stdb)
and (46 from ltsdb), and finally 74 myocardial infractions (MYC) records from
the sixth dataset. The experiment was based on dividing the whole ECG records
into three different parts training, validation, and test. This division made 177 re-
cords used for training and 57 records for validation and 60 for the test. The pa-
rameters of the training are adjusted properly to achieve the highest training per-
formance and the lowest loss error.
Training Parameters Setting
The parameters of the 1D CNN model applied for the ECG diagnosis are deter-
mined in the Table 2. There exist various hyper-parameters that can be set before
the training process. The selected parameters are the optimizer, mini-batch size,
maximum epochs, and total number of iterations, regularization factor, and the
validation frequency.
T a b l e 2 . Parameters adjusted for the proposed 1D CNN Model
t
Optimizer Mini Batch
Size
Maximum
Epochs
Number of
iterations
Validation Ac-
curacy (%)
8 100 2100 94.73
16 100 1100 92.98
32 100 500 89.47
Stochastic gradient
descent
Momentum (Sgdm)
35 100 500 94.73
8 100 2100 92.98
16 100 1100 94.73
32 100 500 94.73
Adaptive Moment
estimation
(adam)
35 100 500 91.22
8 100 2100 98.24
16 100 1100 89.47
32 100 500 91.22
Root mean square
propagation (RMSprop)
35 100 500 89.47
The optimal parameters selected for the 1D CNN model for the diagnosis of
the ECG records are determined experimentally. The optimizers used for training
1D CNN model for ECG diagnosis based on several classifiers
Системні дослідження та інформаційні технології, 2022, № 4 15
the 1D CNN model are the stochastic gradient descent with momentum (SGDM),
adaptive moment estimation (adam), and root mean square propagation
(RMSprop). The mini-batch size parameter is applied with different values such
as 8, 16, 32, and 35 on the three optimizers. The maximum epochs are 100 and
the iterations vary relying on the size of the data and the mini-batch size values.
A validation data is input during the training process with a validation fre-
quency equal to 30, and an L2 regularization factor is defined with a value equal
to 41 10 . It can be seen that when the optimizer is set to rmsprop, and the mini-
batch size is 8 the performance of the validation data has the highest accuracy.
Therefore, the test data are passed to the model with the highest validation
accuracy. In the training stage, the accuracy and the loss curves are obtained for
each of the pre-trained models which is the 1D CNN model. Fig. 3 shows the
highest performance achieved based on the validation data. The blue curve
presents the training curve, whereas, the black dashed curve presents the
validation accuracy curve during the training phase.
Classification Parameters
It is important to determine the hyper-parameters required before training the 1D
CNN model. It is also essential to determine the parameters used on each classi-
fier after applying 1D CNN model for feature extraction. As mentioned before,
two main classifiers are used to determine the diagnosis performance of the 1D
CNN features.
T a b l e 3 . Classifiers applied in the methodology and its optimal parameters
Classifiers Optimal Parameters
Softmax Loss function : "Cross Entropy (CE)"
Random
Forest
Number of trees = 100
Max depth of each tree = 0 (zero indicates unlimited)
Number of features = (log2(no.of.predictors)+1)
The first classifier is the Softmax classifier and its main parameter is the loss
function which is defined by the cross-entropy. The next classifier is the random
forest and it has a set of parameters such as the number of features extracted,
number of trees, and the maximum depth of each tree. Finally, the last classifier
Fig. 3. Training and Vcalidation Accuracy curves performance of proposed 1D CNN Model
Mahmoud M. Bassiouni, Islam Hegazy, Nouhad Rizk, El-Sayed A. El-Dahshan, Abdelbadeeh M. Salem
ISSN 1681–6048 System Research & Information Technologies, 2022, № 4 16
applied is the XGBoost classifier and it also has several parameters. These
parameters are chosen depending on the kind of classification that XGBoost will
perform. In the case of multi-class classification (as in this study) the booster and
the evaluation matrix must be defined by gbtree and mlogloss respectively. The
rest of the XGBoost parameters are used based on their default values in the
library of XGBoost. Table 3 shows the main parameters’ values for the classifiers
used after applying the 1D CNN model.
Classification Results
The features obtained from the fully connected layer of the 1D CNN model are
forward for the two classifiers. The classifiers start to operate on the test data to
ensure the performance of the validation accuracy obtained during the training.
Table.4 shows various statistical performance measurements such as true positive
rate (TPR), precision, false-positive rate (FPR), recall, receiver operating
characteristic (ROC), Mathew’s correlation coefficient (MCC), and precision-
recall characteristic (PRC) value [22].
T a b l e . 4 . Classifiers performance using 1D CNN model based on different
statistical measurements
Performance Measurements (%) Classifiers
Performance TPR FPR Precision Recall F-
Measure Accuracy MCC ROC
Value
PRC
Value
Softmax 0.967 0.011 0.971 0.967 0.967 0.967 0.957 0.992 0.980
RF 0.983 0.006 0.984 0.983 0.983 98.33 0.978 0.997 0.993
These measurements are calculated for each classifier on the test data. In
addition to this, the confusion matrix is manifested to determine the overall
diagnosis performance on the classifiers. The confusion matrix is a figure or a
table that is needed to describe the diagnosis performance of the tested data. It is a
heat map in which the true value must be known. It gives the chance to visualize
the performance of the two applied classifiers on the 1D CNN deep learning mod-
el as shown in Fig. 4 (a and b). It can be manifested that the RF classifier has the
highest accuracy performance over other classifiers.
Confusion Matrix dased on 1D CNN
Model using Softmax Classifier
Confusion Matrix dased on 1D CNN
Model using Random Forest Classifier
a b
Fig. 4. Confusion matrices of Softmax (a) and RF Classifiers on the features obtained
from 1D CNN Model (b)
1D CNN model for ECG diagnosis based on several classifiers
Системні дослідження та інформаційні технології, 2022, № 4 17
DISCUSSION
The paper proposed a methodology for the diagnosis of four main types of ECG
heartbeats. The methodology consists of four main phases, and these phases are
obtaining ECG data, filtering ECG signals, extracting various ECG features, and
classifying ECG records. In the phase of gathering ECG data, the ECG records
are obtained from six various online ECG datasets. In the phase of signal filtra-
tion, the records are filtered using wavelets and a set of filters based on band stop,
low pass, and smoothing filter to eliminate the main common noises in the ECG
signals. In the phase of extracting features, the most discernment feature was ob-
tained from a proposed 1D CNN model. Finally, the classification is applied
based on two main classifiers and these classifiers are Softmax, and RF classifier.
To overcome the chances of overfitting in the proposed model, a regularization
factor is defined to shrink the learned estimates to zero. In other words, this regu-
larization can tune the loss function by providing a penalty term to the optimizer
of the 1D CNN model, and this will encourage smaller weights avoiding exces-
sive changes of the coefficient. In addition to this, the number of ECG records in
each of the four ECG classes are nearly equal leading to a balanced number of
records in each category, dropping the probability of overfitting. Moreover, the
ECG records are filtered using a pre-processing chain for reducing common nois-
es that can cause overfitting during the training. Finally, the training accuracy ob-
tained from the model with the highest accuracy validation is 99.5% and the value
of the highest test accuracy is 98.3%, the slight difference between the training
and the test accuracies shows that the model appropriately fits.
For comparison with others, several algorithms applied different methodolo-
gies for ECG diagnosis as shown in Table 5. S. Yu and M. Lee. [23], the authors
approached an accuracy of 96.38% with the bispectrum feature set and SVM as
the classifier, and when the authors added the genetic feature selector to the bis-
pectrum and the SVM was used for classification, the accuracy increased to
98.10%. The number of records used was 54 and 29 from each of the normal si-
nus rhythm (NSR) and cognitive heart failure (CHF) data sets, respectively. K.H.
Boon et al. [24] applied a diagnosis method to differentiate between normal and
abnormal based on PAF. The features were produced from 106 ECG data col-
lected from 53 ECG recordings. The SVM classifies based on 5 mins heart rate
variability (HRV) segment and its distance from the PAF event. If it is at least 45
min distant from the event, the recording is called normal, but if the HVR seg-
ment goes before the event the recording is called abnormal. The accuracy
achieved was 87.7%. Based on the improvement in the deep learning models in
the diagnosis of the ECG heartbeats. H.B. Bae et al. [25] tried to classify normal
NSR and abnormal ECG records such as AF, and ventricular fibrillation (VF) and
they also focused on balancing the number of records used. The classification was
based on Gamma distribution using probability output networks (CPON), and it
proved that the performance was higher than KNN, SVM, aiming at an accuracy
of 97.33%. R.R. Janghel et al. [26] aimed at building automated classification of
regular and irregular ECG heartbeats. They applied their system on 47 records
and the best results were achieved by using the decision tree, obtaining an accu-
racy of 88.2%.
Mahmoud M. Bassiouni, Islam Hegazy, Nouhad Rizk, El-Sayed A. El-Dahshan, Abdelbadeeh M. Salem
ISSN 1681–6048 System Research & Information Technologies, 2022, № 4 18
T a b l e 5 . Proposed DL model compared to other previous work for ECG diagnosis
Authors Records Methodology Classes Databases Performance
S.N. Yu et
al. [23]
2012
54 R from NSR
+
29 R from CHF
Features: Bisecpectrum +
genetic feature set
Classifier: SVM-
2
MIT-BIH
NSR and
CHF
Bisecptrum +
SVM = 96.38%
Bispecturm + ge-
netic feature set +
SVM = 98.10%
K.H. Boon
et al.
[24]
2018
106 data
from 53
R pairs
Features: Time domain,
spectral, Bispectrum,
nonlinear dynamics
features
Classifiers: SVM
2
Atrial
Fibrillation
prediction
(AFPDB)
Database
ACC = 87.7%
H.B. Bae et
al.
[25]
2019
NSR: 15 R
VF: 15 R
AF: 15 R
R–R interval + (CPON) 3
MIT-BIH
(NSRDB),
(VFDB),
(AFDB)
ACC = 97.33%
R.R. Jang-
hel
et al.
[26] 2020
47 R
40% of the 47 R
records are pa-
tients
Naïve Bayes
SVM
Ada-boost
RF, Decision Tree, and
KNN
2
MIT-BIH
arrhythmia
database
ACC of
the Decision
Tree = 88.2%
The
proposed
Method
294 R
177 R for train
117 R for valida-
tion and test
Proposed 1D CNN
Model 4 6 main data-
sets
Softmax
ACC =96.7%
RF
ACC = 98.3%
The proposed methodology worked on 294 recordings obtained from 4 dif-
ferent ECG heartbeats. The features are obtained from a 1D CNN model. Two
main classifiers were applied to reach 96.7% using Softmax and 98.3% using RF
classifier. The advantages of the proposed model are illustrated in three main
points. The first point is the removal of the three common noises related to the
ECG signals using a well-defined pre-processing chain. The second point is ob-
taining robust features from the 1D CNN model. The last point is the superiority
of the XB-boost in the classification because it is highly flexible, can be paral-
leled, supports generalization, and is faster than gradient boosting.
CONCLUSIONS AND FUTURE WORK
In this study, a methodology is presented for the diagnosis of the four different
types of ECG heartbeats based on a proposed 1D CNN model. The proposed
methodology produces better results makes it adaptable for the diagnosis of dif-
ferent ECG records. The data were collected from 6 public available datasets. The
ECG records were filtered to drifting in the ECG signals, powerline interference,
and the high noise frequencies. The filtering chain is based on wavelets and a set
of filters. Then, the ECG records are passed to a 1D CNN model for feature ex-
traction. Finally, the classification is based on Softmax and RF, classifiers achiev-
ing an accuracy of 96.6% and 98.3%, respectively. ECG signals have future direc-
tions that can contribute and provide assistance in the field of medical
informatics. There is a need for a real-time diagnosis application that can verify
various types of heart diseases. In addition to this, it was discovered recently that
the ECG signals can diagnose COVID patients based on the ECG image reports.
It is recommended to develop diagnosis systems that can identify COVID patients
from normal and various abnormal heartbeats. It is also suggested to use stratified
k-fold cross-validation in future experiments to provide more information about
the methodology performance. It is also advised to select the hyper-parameters
1D CNN model for ECG diagnosis based on several classifiers
Системні дослідження та інформаційні технології, 2022, № 4 19
based on various methods such as grid or random search or various metaheuristic
techniques to reach the optimal values on the parameters for the proposed model.
Finally, the XB-Boost classifier can be replaced with a sparse representation classifier
as it is considered a powerful technique for pixel-wise classification of images [27].
REFERENCES
1. World Health Organization, Cardiovascular Diseases (CVDs). 2022. [Online]. Avail-
able: http://www.who.int/mediacentre/factsheets/fs317/en/
2. S.K. Berkaya, A.K. Uysal, E.S. Gunal, S. Ergin, S. Gunal, and M.B. Gulmezoglu,
“A survey on ECG analysis,” Biomed Signal Process Control, 43, pp. 216–235, 2018.
3. O. Faust, Y. Hagiwara, T.J. Hong, O.S. Lih, and U.R. Acharya, “Deep learning for
healthcare applications based on physiological signals: A review,” Biomed. Comput.
Meth. Prog. Bio., 161, pp. 1–13, 2018.
4. L.A. Abdullah and M.S. Al-Ani, “CNN-LSTM based model for ECG arrhythmias and
myocardial infarction classification,” Adv. Sci. Technol. Eng. Syst., 5(5), pp. 601–606,
2020.
5. E. Butun, O. Yildirim, M. Talo, R.S. Tan, and U.R. Acharya, “1D-CADCapsNet: One
dimensional deep capsule networks for coronary artery disease detection using ECG sig-
nals,” Physica Medica, 70, pp. 39–48, 2020.
6. X. Hua et al., “A novel method for ECG signal classification via one-dimensional convo-
lutional neural network,” Multimedia Systems, pp. 1–13, 2020.
7. G. Petmezas et al., “Automated atrial fibrillation detection using a hybrid CNN-LSTM
network on imbalanced ECG datasets,” Biomedical Signal Processing and Control, 63,
102194, 2021.
8. MIT-BIH Normal Sinus Rhythm. [Online]. Available: https://www.physionet.org/ con-
tent/nsrdb/1.0.0/ last accessed 2-10-2021.
9. Normal Sinus Rhythm RR Interval. [Online]. Available: https://physionet.org/ con-
tent/nsr2db/1.0.0/ last accessed 2-10-2021.
10. MIT-BIH Supraventricular Arrhythmia. [Online]. Available: https://physionet.org/ con-
tent/svdb/1.0.0/ last accessed 2-10-2021.
11. MIT-BIH ST Change Database. [Online]. Available: https://physionet.org/content/
stdb/1.0.0/last accessed 2-10-2021.
12. Long Term ST Database. [Online]. Available: https://physionet.org/content/l tstdb/1.0.0/
last accessed 2-10-2021.
13. PTB Diagnostic ECG Database. [Online]. Available: https://www.physionet.org/ con-
tent/ptbdb/1.0.0/ last accessed 2-10-2021.
14. M.M. Bassiouni, E.S.A. El-Dahshan, W. Khalefa, and A.M. Salem, “Intelligent hybrid
approaches for human ECG signals identification,” Signal Image Video Process, 12(5),
pp. 941–949, 2018.
15. M. Bassiouni, W. Khaleefa, E.A. El-Dahshan, and A.B.M Salem, “A machine learning
technique for person identification using ECG signals,” Int. J. Appl. Phys., 1, pp. 37–41,
2016.
16. M.M. Bassiouni, I. Hegazy, N. Rizk, S.A. El-Dahshan, and A.M. Salem, “Combination
of ECG and PPG Signals for Healthcare Applications: A Survey”, Advances in Modelling
and Analysis, 64(1-4), pp. 63–70, 2021.
17. M.M. Bassiouni, I. Hegazy, N. Rizk, E.S.A. El-Dahshan, and A.M. Salem, “Automated
Detection of COVID-19 Using Deep Learning Approaches with Paper-Based ECG
Reports,” Circuits, Systems, and Signal Processing, 41, pp. 1–43,2022.
18. A.F. Agarap, “Deep learning using rectified linear units (relu),” arXiv preprint,
arXiv:1803.08375, 2018.
19. A.T. Azar, H.I. Elshazly, A.E. Hassanien, and A.M. Elkorany, “A random forest classi-
fier for lymph diseases,” Comput. Biol. Med., 113(2), pp. 465–473, 2014.
20. S.Y. El-Bakry, E.S. El-Dahshan, and M.Y. El-Bakry, “Total cross section prediction of
the collisions of positrons and electrons with alkali atoms using Gradient Tree Boosting,”
Indian J. Phys., 85(9), pp. 1405–1415, 2011.
Mahmoud M. Bassiouni, Islam Hegazy, Nouhad Rizk, El-Sayed A. El-Dahshan, Abdelbadeeh M. Salem
ISSN 1681–6048 System Research & Information Technologies, 2022, № 4 20
21. H.T. Weldegebriel, H. Liu, A.U. Haq, E. Bugingo, and D. Zhang, “A new hybrid convo-
lutional neural network and eXtreme gradient boosting classifier for recognizing hand-
written Ethiopian characters,” IEEE Access, 8, pp.17804–17818, 2019.
22. S. García, A. Fernández, J. Luengo, and F. Herrera, “A study of statistical techniques and
performance measures for genetics-based machine learning: accuracy and interpretabil-
ity,” Soft Comput., 13(10), 959, 2009.
23. S.N. Yu and M.Y. Lee, “Bispectral analysis and genetic algorithm for congestive heart
failure recognition based on heart rate variability,” Comput. Biol. Med., 42(8), pp. 816–825,
2012.
24. K.H. Boon, M. Khalil-Hani, and M.B. Malarvili, “Paroxysmal atrial fibrillation predic-
tion based on HRV analysis and non-dominated sorting genetic algorithm,” Comput.
Biol. Med., 153, pp. 171–184, 2018.
25. H.B. Bae, M.S. Park, R.M. Kil, and H.Y. Youn, “Classifying heart conditions based on
class probability output networks,” Neurocomputing, 360, pp. 198–208, 2019.
26. R.R. Janghel and S.K. Pandey, “A Classification of ECG Arrhythmia Analysis Based on
Performance Factors Using Machine Learning Approach,” in Computational Network
Application Tools for Performance Management. Springer, Singapore, 2020, pp. 65–74.
27. M.M. Bassiouni, I. Hegazy, N. Rizk, E.S.A. El-Dahshan, and A.M. Salem, “Deep learn-
ing approach based on transfer learning with different classifiers for ECG diagno-
sis,” International Journal of Intelligent Computing and Information Sciences, 22(2),
pp.44–62, 2022.
Received 23.09.2022
INFORMATION ON THE ARTICLE
Mahmoud M. Bassiouni, ORCID: 0000-0002-8617-8867, Egyptian E-Learning Univer-
sity (EELU), Egypt, e-mail: mbassiouni@eelu.edu.eg
Islam Hegazy, ORCID: 0000-0002-1572-463X, Ain Shams University, Egypt, e-mail:
islheg@cis.asu.edu.eg
Nouhad Rizk, ORCID: 0000-0001-9277-9741, Houston University, USA, e-mail:
njrizk@uh.ed
El-Sayed A. El-Dahshan, ORCID: 0000-0002-1221-0262, Ain Shams University, Egypt,
e-mail: seldahshan@eelu.edu.eg
Abdelbadeeh M. Salem, ORCID: 0000-0003-0268-6539, Ain Shams University, Egypt,
e-mail: absalam@cis.asu.edu.eg
1D МОДЕЛЬ CNN ДЛЯ ДІАГНОСТИКИ ЕКГ НА КІЛЬКОХ
КЛАСИФІКАТОРАХ / М.М. Басіуні, І. Хегазі, Н. Різк, E.C.A. Ел-Дашан, А.М. Салем
Анотація. Однією з основних причин смерті людини є захворювання серця.
Виявлення серцевих захворювань на ранній стадії може запобігти серцевій
недостатності або будь-якому пошкодженню серцевого м’яза. Одним з основ-
них сигналів, які можуть бути корисними в діагностиці захворювань серця, є
електрокардіограма (ЕКГ). Розглянуто діагностику чотирьох типів записів
ЕКГ, таких як інфаркт міокарда (MYC), норма (N), відхилення сегмента ST
(ST) і надшлуночкова аритмія (SV). Методологія збирає дані з шести основних
наборів даних, а потім записи ЕКГ фільтруються за допомогою ланцюжка по-
переднього оброблення. Після цього запропонована модель 1D CNN
використовується для вилучення ознак із записів ЕКГ. Потім застосовуються
два різні класифікатори, щоб перевірити ефективність виділених ознак і отри-
мати надійну точність діагностики. Два класифікатори – це softmax і
класифікатор випадкового лісу (RF). Застосовується експеримент для
діагностики чотирьох типів записів ЕКГ. Зрештою найвищої продуктивності
досягнуто за допомогою радіочастотного класифікатора з точністю 98,3%.
Порівняння з іншими суміжними роботами показало, що запропоновану мето-
дику можна застосовувати для раннього виявлення захворювань серця.
Ключові слова: електрокардіограма (ECG), безперервне вейвлет-перетворення
(CWT), одновимірна модель згорткової нейронної мережі (CNN).
|
| id | journaliasakpiua-article-265099 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:27:57Z |
| publishDate | 2022 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/c4/46b4b8491f7ba3ad1681050d8b5b9cc4.pdf |
| spelling | journaliasakpiua-article-2650992023-05-21T20:04:38Z 1D CNN model for ECG diagnosis based on several classifiers 1D модель CNN для діагностики ЕКГ на кількох класифікаторах Bassiouni, Mahmoud Hegazy, Islam Rizk, Nouhad El-Dahshan, El-Sayed Salem, Abdelbadeeh Electrocardiogram (ECG) Continuous wavelet transform (CWT) 1D convolutional neural network (CNN) model електрокардіограма (ECG) безперервне вейвлет-перетворення (CWT) одновимірна модель згорткової нейронної мережі (CNN) One of the main reasons for human death is diseases caused by the heart. Detecting heart diseases in the early stage can stop heart failure or any damage related to the heart muscle. One of the main signals that can be beneficial in the diagnosis of diseases of the heart is the electrocardiogram (ECG). This paper concentrates on the diagnosis of four types of ECG records such as myocardial infarction (MYC), normal (N), variances in the ST-segment (ST), and supraventricular arrhythmia (SV). The methodology captures the data from six main datasets, and then the ECG records are filtered using a pre-processing chain. Afterward, a proposed 1D CNN model is applied to extract features from the ECG records. Then, two different classifiers are applied to test the extracted features’ performance and obtain a robust diagnosis accuracy. The two classifiers are the softmax and random forest (RF) classifiers. An experiment is applied to diagnose the four types of ECG records. Finally, the highest performance was achieved using the RF classifier, reaching an accuracy of 98.3%. The comparison with other related works showed that the proposed methodology could be applied as a medical application for the early detection of heart diseases. Однією з основних причин смерті людини є захворювання серця. Виявлення серцевих захворювань на ранній стадії може запобігти серцевій недостатності або будь-якому пошкодженню серцевого м’яза. Одним з основних сигналів, які можуть бути корисними в діагностиці захворювань серця, є електрокардіограма (ЕКГ). Розглянуто діагностику чотирьох типів записів ЕКГ, таких як інфаркт міокарда (MYC), норма (N), відхилення сегмента ST (ST) і надшлуночкова аритмія (SV). Методологія збирає дані з шести основних наборів даних, а потім записи ЕКГ фільтруються за допомогою ланцюжка попереднього оброблення. Після цього запропонована модель 1D CNN використовується для вилучення ознак із записів ЕКГ. Потім застосовуються два різні класифікатори, щоб перевірити ефективність виділених ознак і отримати надійну точність діагностики. Два класифікатори – це softmax і класифікатор випадкового лісу (RF). Застосовується експеримент для діагностики чотирьох типів записів ЕКГ. Зрештою найвищої продуктивності досягнуто за допомогою радіочастотного класифікатора з точністю 98,3%. Порівняння з іншими суміжними роботами показало, що запропоновану методику можна застосовувати для раннього виявлення захворювань серця. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2022-12-27 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/265099 10.20535/SRIT.2308-8893.2022.4.01 System research and information technologies; No. 4 (2022); 7-20 Системные исследования и информационные технологии; № 4 (2022); 7-20 Системні дослідження та інформаційні технології; № 4 (2022); 7-20 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/265099/270160 |
| spellingShingle | електрокардіограма (ECG) безперервне вейвлет-перетворення (CWT) одновимірна модель згорткової нейронної мережі (CNN) Bassiouni, Mahmoud Hegazy, Islam Rizk, Nouhad El-Dahshan, El-Sayed Salem, Abdelbadeeh 1D модель CNN для діагностики ЕКГ на кількох класифікаторах |
| title | 1D модель CNN для діагностики ЕКГ на кількох класифікаторах |
| title_alt | 1D CNN model for ECG diagnosis based on several classifiers |
| title_full | 1D модель CNN для діагностики ЕКГ на кількох класифікаторах |
| title_fullStr | 1D модель CNN для діагностики ЕКГ на кількох класифікаторах |
| title_full_unstemmed | 1D модель CNN для діагностики ЕКГ на кількох класифікаторах |
| title_short | 1D модель CNN для діагностики ЕКГ на кількох класифікаторах |
| title_sort | 1d модель cnn для діагностики екг на кількох класифікаторах |
| topic | електрокардіограма (ECG) безперервне вейвлет-перетворення (CWT) одновимірна модель згорткової нейронної мережі (CNN) |
| topic_facet | Electrocardiogram (ECG) Continuous wavelet transform (CWT) 1D convolutional neural network (CNN) model електрокардіограма (ECG) безперервне вейвлет-перетворення (CWT) одновимірна модель згорткової нейронної мережі (CNN) |
| url | https://journal.iasa.kpi.ua/article/view/265099 |
| work_keys_str_mv | AT bassiounimahmoud 1dcnnmodelforecgdiagnosisbasedonseveralclassifiers AT hegazyislam 1dcnnmodelforecgdiagnosisbasedonseveralclassifiers AT rizknouhad 1dcnnmodelforecgdiagnosisbasedonseveralclassifiers AT eldahshanelsayed 1dcnnmodelforecgdiagnosisbasedonseveralclassifiers AT salemabdelbadeeh 1dcnnmodelforecgdiagnosisbasedonseveralclassifiers AT bassiounimahmoud 1dmodelʹcnndlâdíagnostikiekgnakílʹkohklasifíkatorah AT hegazyislam 1dmodelʹcnndlâdíagnostikiekgnakílʹkohklasifíkatorah AT rizknouhad 1dmodelʹcnndlâdíagnostikiekgnakílʹkohklasifíkatorah AT eldahshanelsayed 1dmodelʹcnndlâdíagnostikiekgnakílʹkohklasifíkatorah AT salemabdelbadeeh 1dmodelʹcnndlâdíagnostikiekgnakílʹkohklasifíkatorah |