Content uploaded by Yogesh Bhosale
Author content
All content in this area was uploaded by Yogesh Bhosale on Feb 01, 2023
Content may be subject to copyright.
Content uploaded by Yogesh Bhosale
Author content
All content in this area was uploaded by Yogesh Bhosale on Feb 01, 2023
Content may be subject to copyright.
978-1-6654-7100-8/22/$31.00 ©2022 IEEE
ECG-CCNet: Cardiovascular(Cardiac) and COVID‑19
Disease Classification Using Deep Convolutional
Neural Network Learning Pipeline Approaches From
Electrocardiography(ECG)- A Study
Yogesh H. Bhosale
Dept. of CS&E, Birla Institute of Technology Ranchi, India.
yogeshbhosale988@gmail.com
K. Sridhar Patnaik
Dept. of CS&E, Birla Institute of Technology Ranchi, India.
kspatnaik@bitmesra.ac.in
Abstract- Experimental studies demonstrate that COVID-19
illness affects the cardiovascular as well as the pulmonary / lung
tract. The limits of existing COVID-19 diagnostic procedures have
been revealed. In contrast, to present diagnoses, such as low-
sensitivity conventional RT-PCR testing and costly healthcare
scanning equipment, implementing additional approaches for
COVID-19 illness assessment would be advantageous for COVID-
19 epidemic management. Furthermore, problems generated by
COVID-19 on the cardiovascular tract must be detected rapidly and
precisely using ECG. Considering the numerous advantages of
electrocardiogram (ECG) functionalities, the proposed study offers
a novel pipeline termed ECG-CCNet for examining the feasibility
of employing ECG pulses to diagnose COVID-19. This study is a
two-phase transfer learning (TL) approach is suggested for the
prognosis of COVID-19 disorder, which includes feature mining
utilizing DCNNs models and ensemble pipelining using ECG
tracing imageries generated from ECG signals of COVID-19
diseased sufferers relying on the anomalies induced by COVID-19
pathogen on cardiovascular structures. A complete classification
performance of 93.5% accuracy, 87% recall, 87.03% F1-score,
95.66% specificity, 87.16% precision, and 95.33% AUC attained by
abnormal heartbeats, COVID-19, myocardial, and normal/healthy
classification. This experiment is considered a high possibility for
speeding up the diagnostic and treatments of COVID-19 individuals,
reducing practitioners' efforts, and improving epidemic containment
by utilizing ECG data.
Keywords- Electrocardiography (ECG) trace image, Deep
learning, cardiovascular diseases diagnosis, Convolutional neural
networks (CNN), COVID-19 detection & Classification, Biomedical
image signal processing.
I. INTRODUCTION
Coronaviruses are a vast viral subfamily that can produce
a variety of illnesses in both mammals and mankind. COVID-
19 originally surfaced at the start of December 2019 in Wuhan
region, China; it has since expanded at an extraordinary rate
around the continent [1]. Fever, coughing, and breathlessness
are the most prevalent indications. Tuberculosis, chronic lung
failure, kidney failure, pneumonia, and mortality may occur in
extreme situations. As of 20 June 2022, the 533 million cases
worldwide, with 6.3 million deaths due to COVID-19 [2].
Despite the fact that without viable treatment for COVID-19,
which the WHO proclaimed an epidemic disease on 11 March
2020 [3], there have been encouraging advancements in
antigen research. According to published resources, vaccine
efficacy is relatively good, with 1186 million dosages of
COVID-19 vaccine provided to the whole community in this
article's drafting [3].
RT-PCR screening is currently the WHO official
standardized approach for diagnosing COVID-19 [4]. Despite
the fact that those testing are the benchmark due to their
adequate sensitivity percentages, they have significant
limitations. These require a lengthy timeframe to provide
results, excellent laboratory conditions, and skilled experts to
execute the examinations [5]. It may generate a poor outcome
even if the individuals have identified COVID-19 [6]. Various
alternative tests and procedures that produce quicker and more
reliable findings are still being investigated. Biomedical
screening using X-ray and CT scan modes is one of these ways
employed for the rapid recognition of COVID-19. Lung X-
rays and CT pictures conceal important knowledge that could
aid in diagnosing lung illnesses. Many studies have
established the existence of differences in chest radiography
images recorded prior to the onset of COVID-19 signs.
Furthermore, multiple investigations have shown that X-ray
and CT scans are more reliable than PCR testing in diagnosing
COVID-19 [7], [8], and [20]. Nonetheless, due to the
closeness of COVID-19 characteristics with other kinds of
pulmonary diseases, these approaches necessitate the
involvement of a professional radiologist to diagnose in-
hospital patients[9].
Although the COVID-19 primarily affects the circulatory
tract, it impacts other essential human organs, including the
cardiovascular system [9]. Computer vision algorithms based
on deep learning(DL) are very successful and beneficial,
particularly for diagnostics and monitoring. Cardiovascular
changes [10,11] have promoted using ECG as a screening tool
for COVID-19 assessment. Considering the numerous
advantages of ECG use, such as mobility, accessibility, the
convenience of use, affordable, harmlessness, and real-time
inspection, the automated identification of COVID-19 using
ECG may be of substantially worth in addition to PCR testing
and thoracic X-ray or CT imaging. The cardio-vascular tract
is an essential area in which the infection interferes with
regular circulation. ECG physiologic impulses, on either hand,
are an excellent predictor of pathological disorders in the heart
and lungs. The usual approach to analyzing ECG imaging with
DL approaches is retrieving handcrafted characteristics and
constructing DL & ML classifiers. As an alternative to the
present diagnosis techniques, a novel, safe, affordable,
precise, rapid ensemble pipeline dubbed ECG-CCNet is
proposed to aid in the automated analysis of cardiovascular
and COVID-19.
The major contributions of the projected work are:
• The new method is based on 2-D ECG tracing
imageries to detect cardiovascular and COVID-19, which is a
unique technique for experimental analysis.
• ECG-CCNet extracts 2-levels of deep features from
four DCNNs of distinct structures. The potential of using ECG
2022 IEEE Silchar Subsection Conference (SILCON) | 978-1-6654-7100-8/22/$31.00 ©2022 IEEE | DOI: 10.1109/SILCON55242.2022.10028792
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on February 01,2023 at 07:00:51 UTC from IEEE Xplore. Restrictions apply.
data for COVID-19 diagnosis is investigated using a new
framework called ECG-CCNet.
• Exploring the integrated multi-model ensemble
features to expand the effectiveness, particularly after feature
collection and ensemble classification.
• The results of ECG-CCNet show the ECG might be
used for cardiovascular and COVID-19 disease diagnosis.
The remaining section is arranged as follows. Recent work
has been illustrated in section 2. Materials and
Implementations are discussed in section 3. Results and
concluding remarks are offered in sections 4 and 5,
respectively.
II. RELATED WORKS
Various studies have exploited DL and ML practices to
distinguish cardiac problems from ECG data. Sannino et al.
[17] and Ghosh et al. [19] utilized 1-D raw ECG gestures to
categorize the irregularities. Whereas Mishra et al. [16],
Widman et al. [18] used Fourier transformation and short-term
frequency transform to convert ECG data to 2-D pictures to
identify problems. Despite prior techniques having substantial
progress utilizing publicly available ECG signal samples, they
would be challenging to implement in a real-world medical
setting. This is due to the fact that the majority of existing
approaches rely on time series ECG readings. Srivastava et al.
[24] utilized DL variant i.e., ResNetV2 for 3-class
classification with 97.90% accuracy( i.e., covid-19, healthy,
pneumonia). Jain et al. [25] compared and reported the
performance of InceptionV3, Xception, and ResNeXt
networks using 6432 thoracic X-ray image data from their
Kaggle database. According to the researchers, the Xception
framework had the maximum accurateness of 97.97 %. The
authors proposed an innovative technique by using
LeakyReLU rather than relu as an activation function.
Furthermore, the researchers underlined that the remarkable
accuracy discovered could be cause for warning since it could
result from overfitting. The researchers advised that huge
databases be considered in the upcoming years to verify their
proposed approach.
Panwar et al. [26] presented a nCOVnet framework built
upon TL. The nCOVnet design begins with the input neurons,
which is proceeded with 18 levels of
convolution+ReLU+pooling levels from pretrained VGG16.
Thereafter included, 5-extra customized tiers as top levels in
the second method. When evaluated on Covid-19, the training
accuracy reported was 93 to 97%, while 97.62% for the test
set. The dataset used comprises 337 pictures from covid X-
rays. Researcher considered the potential of data leaking;
hence the data was carefully divided between train and test
pairs. Basu et al. [27] proposed DETL learning a CNN from
base, that takes substantial skill for a structure to operate
correctly and massive data to train. After a 5-fold Cross-
Validation, the stated precision of 82.98, 90.13, 85.98 for
Alexnet, VGGNet, ResNet, respectively. Additionally,
employed the Grad-CAM concept to determine if a classifier
earned greater focus throughout picture classification.
Luz et al. [28] employed CNNs framework to identify
Covid19 via X-ray. EfficientNet had correctness of 93.9%,
96.8% sensitivity, and 100% precision(no false negatives)
despite needing 5 to 30 percent fewer features. A database of
13569 scans of individuals classified as healthy, COVID-19
pneumonia, and non-COVID-19 pneumonia was employed to
develop the suggested techniques. A multilayer method and
cross-dataset analysis demonstrated that even the utmost
complex versions are incapable of generalization. As may be
observed, this research mainly contained a limited collection
of radiographs from Covid-19 positive individuals. The
authors evaluated the effectiveness of the proposed methods
on large and diverse data. Apostolopoulos et al. [23]
accomplished COVID-19 illness identification with an
accuracy of 96.78% utilizing additional DL-based techniques
employing X-rays. Bassiouni et al.[22] using pre-trained DL
models with SVM classifier demonstrated superior 99%
accuracy for binary classification (COVID-19 +Ve and -Ve).
Bhosale et al. [32], [33] used DCNN to diagnose COVID-19
from other 8-lung diseases(X-ray). They achieved a high
accuracy ratio by using the IoT approach deployment for
network parameter adjustment.
Despite prior techniques having substantial results
utilizing publicly available ECG signal databases, they
would've been challenging to implement in a real medical
setting. It's due to the fact that the majority of existing
approaches rely on time series ECG readings. However, it is
not often the situation in real diagnostic circumstances, as
ECG data is recorded and preserved as pictures in practice
[29]. In contrast to electronic Datasets, which comprise
numerous crisp and fine-detached lead impulses, image-based
ECG data obtained in practice is hazy. Furthermore, there is
considerable overlapping across signals from different leads
and solid adjacent supplementary axis in an ECG picture,
complicating precisely identifying valuable characteristics.
Again, the significant reduction in recording frequency
between 100 Hz in ECG electrical impulses to 10 Hz or fewer
in ECG pictures results in significant data loss, influencing the
effectiveness of AI algorithms (either using the handcrafted or
DL methods). Converting the picture into a digital file [16] is
one potential approach to this issue, although this conversion
has a significant processing burden, and the integrity of the
converted signals is confined [30].
To identify a cardiovascular ailment using ECG data, the
changes across practically all types of cardiac problems are
frequently minor; these minute variations are the primary
ingredients for abnormalities identification. Even with DL
algorithms' high learning capacity, they are unable to reliably
spot the discriminatory sections because of noise in the
converted ECG signals.
III. METHODS & IMPLEMENTATIONS
A. ECG image dataset and preprocessing
This study used recent publicly accessible datasets [15]
that included ECG tracing pictures of COVID-19 and
additional cardiovascular problems. To the utmost of my
understanding, it is the first and sole database for the new
coronavirus's publicly available ECG dataset. The collection
contains 1937 ECG pictures of various classifications. These
recordings are examined by physicians, which are classified
as healthy individuals with no heart abnormalities (859 ECG),
COVID-19 (250 ECG), irregular heartbeats (548 ECG), and
present or prior myocardial infarction-MI (300 ECG). ECG
readings using a 12-lead device were collected at a 500Hz
sampling frequency utilizing an EDANSE3-series 3-channel
ECG. Normal/healthy (250 ECG), COVID-19(250 ECG), MI
disorders(250 ECG), and atypical heartbeats(250 ECG)
pictures are used in the multi-class categorization criteria. This
value was chosen to equalize the sample and prevent the
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on February 01,2023 at 07:00:51 UTC from IEEE Xplore. Restrictions apply.
unbalanced impact, which might affect classifications'
efficiency. The additional information mentions the
participants' data being utilized. The dataset was partitioned
based on a 5-fold cross-validation ratio for train, Val, and test
sets, respectively. All ECG images were normalized using
1/255. Imagedatagenerator was used to imply the data
augmentation with Position augmentation parameters(scaling,
cropping, flipping, padding, rotation, and translation). All the
ECG images are centrally cropped to 849x748 pixel
resolution. ECG tracing images of heart abnormalities,
COVID-19, MI, and normal are shown in Fig. 1.
(a) Abnormal Heart beats ECG.
(b) COVID-19 ECG
(c) Myocardial ECG.
(d) Healthy/Normal
Fig. 1. Dataset samples of ECG trace images.
B. Proposed methodology
The proposed ECG-CCNet utilizes the 4 variants of DL-
based DCNNs. We have modified the pretrained network
structure with our own requirements. We have used Input
layer (IL) + 3-Convolutional Layers(CL) +3-Pooling and
Relu(PR) + Flatten(FL) + Dense with 5-neurons in VGG16
and ResNet152V2; IL+ 4 CL+4 PR + FL+ Dense with 5-
neurons in DenseNet201 and MobileNetV2. The remaining
layers of DCNNs are set to false(non-trainable). After training
these 4 models, we have implemented deep feature ensemble
pipeline, where multiple individual model features are
concatenated. And finally, the individual model classification
results and ensemble pipeline features classification are
obtained using benchmark performance metrics. The
flowchart of the proposed system is depicted in Fig. 2.
CNN was already programmed to acquire and analyze
many features and correlations continuously. The 3
components that comprise the system are CL, pool, and FC
layers. The CL and pool layers are used to retrieve information
in the first two layers. The last layer, i.e., FC, does
categorization, which is mapped to the last output. We
eliminate the fully connected layers of a pre-trained CNN
model, and the above said layers are trained(which acts as a
feature extractor), and the remaining layers are untrained in
the feature extraction approach. In DL, Fine-tuning is being
performed to greatly minimize the duration and cost needed to
increase the network efficiency. These are accomplished by
using the first layer of pretrained networks and associated
features on ECG dataset, assuming it contains equivalent
inputs. To execute a new ECG classification job, TL structures
are modified using the needed particular characteristics of a
previously trained model. DL approaches are widely used to
analyze and identify pictures from datasets [5]. Those
equations can be used to summarise how DL techniques
function. First, we calculate ‘z’ utilizing input variables ‘xi’ as
indicated in Eq. 1.
=
(1)
(=weights, =bias)
Next, we use ‘z’ to calculate ‘a’, that is equivalent to ‘y’ at the
outputting nodes, as given in Eq. 2.
= () (2)
(Where =activation function)
Eq. 3 is made by joining Eq. 1 and 2,
(3)
Deep TL refers to the process of acquiring deeper features
and fine-tuning pre-trained CNN classifiers [21] [32]. When
there are a minimal number of training photos, TL improves
the DL technique for the above-recommended technique's
image identification application. Rather than developing a
network from fresh, the parameters from a pre-trained model
could be used to accelerate or optimize the learning
experience. The CNNs earliest blocks may be regarded as
picture feature descriptors, whereas the latter levels are
associated with specific subcategories. As an outcome, many
layers can be used for a range of purposes. Thereafter, TL
must decide which layers of a pretrained DCNN must be
employed [33]. This method has shown to be useful in a range
of vision-based tasks, even when relocating features from
totally dissimilar fields. After adopting the framework to the
issue at hand, we create a new layer and specify the learning
experience. We accomplished modified DCNNs utilizing
DTL and adjusted the parameters for classifications by means
of proper optimization, 'Softmax' function.
The final level is the FC layer, that contains flattening.
After that, the entire pooled feature-map vectors are turned
into a separate block, and fed to CNN. We build a classifier
by integrating those features with fully linked layers. Lastly,
as described in Eq. 4, the result is categorised by utilising
SoftMax.
() = / (∑C=1 ) (4)
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on February 01,2023 at 07:00:51 UTC from IEEE Xplore. Restrictions apply.
where =softmax, =input-vector, =input/output
exponential function, C=No. of classes.
Finally, an 11-deep ensemble feature pipeline(using 4-
DCNNs) is generated using Eq. 5.
e=
(5)
where e=deep ensemble feature pipeline, m1…mn=individual
CNNs feature stream, n=4.
Lastly, 11-ensemble possible models with 4-disease
classification results for ECG-CCNet from the softmax layer
were achieved from 4 individual DCNNs for ECG-based
COVID-19 and cardiac disease classification. Ultimately, DL
model experimental performance accumulated.
Fig. 2. Flowchart of the suggested ECG-CCNet framework.
C. Experimental setup
The experimented approach is a bilinear phase, i.e.,
training TL models on the ECG image dataset and generating
a deep ensemble feature pipeline. In the first phase, DCNNs
(VGG16, DenseNet201, MobileNetV2, ResNet152V2) are
trained with ECG datasets using 5-fold cross-validation. All
the input ECG image sets 849x748 were rescaled to
acceptable model inputs, i.e., 224 x 224 x 3 set. The training
model is set to 75 epochs based on prior-halting criteria. The
actual learning rate was set to 0.00001. Training halt criteria
have automatically determined the iteration size.
Training stopped at 32 epochs for VGG16 with 90.49%
accuracy and 76.40% val_acc; 35 epochs for DenseNet201
with 91.90% accuracy and 75.55% val_acc; 25 epochs for
MobileNetV2 with 91.99% accuracy and 76.55% val_acc; 25
epochs for ResNet152V2 with 90.19% accuracy and 80.50%
val_acc. The 'adam' optimizer was utilized for training with a
learning rate of 0.00001. The batch size is set to 32.
D. Performance Metrics
A confusion matrix(CM) is plotted to check the efficacy
of a classification network ("classifier") on the testing set.
CM(Predicted Vs. Actual Labels) retains predictions in the
form of TP(they have the disease), TN(predicted no disease),
FP("Type I error"), and FN("Type II error"). With the help of
CM, we have evaluated the benchmark performance metrics
such as accuracy, precision, recall, specificity, and F1-score.
The obtained CM is shown in Fig 4.
E. Experimental environment
All the experiments were carried out on Google
Colaboratory. Python 3.10, TensorFlow 2.8, CUDA 11.2,
matplotlib, NumPy, pandas, Keras, sklearn, cm, and seaborn
APIs are the development platforms utilized for the complete
experimentations.
IV. RESULTS
We experimented with 4 prominent DCNNs on the ECG
dataset and use case. Depending upon their findings, ensemble
algorithms were tested and employed to improve efficiency in
multiple models. Table 2 shows the results of the initial
VGG16, DenseNet201, MobileNetV2, and ResNet152V2
implementations. Ensemble disclosed decent accuracy with
other benchmark metrics. With respect to Table I, Covid-19
showed good outcomes for VGG16 and MobileNetV2 with
100% results for all metrics. However, 88% accuracy was
attained by DenseNet201 for Abnormal heartbeats and
Myocardial disease and 93% accuracy for normal(healthy)
labels from individual DCNs.
Table II shows the average macro performance of all the
ensemble possible ensemble models from four DCNNs. We
tried our best to show the performance of all (2n-1), i.e., 15 (4
individual + 11 ensemble pairs) combinations from four
individual models. The ensemble pairs are depicted using
[M1…M4] series. According to Table 2 outcomes the highest
performance attained by [M2, M3] and [M1, M2, M3] pair
with 93.5% accuracy, 87% recall, 87.03% F1-score, 95.67%
specificity, 87.18% precision, 95.33% AUC-score, (0.8040
+/- 0.9359)% of 95% confidence interval. However, the
highest AUC was attained by [M1, M2, M3, M4] pair with
95.33%. Even after combining the DCN, the lowest
performance was attained by the [M1, M4] ensemble model.
Fig. 3 shows the best and optimal confusion matrix from
the combination of ensemble pairs. From the confusion matrix
shown in Fig. 3, the highest 87 TP samples are attained by
[M2, M3], [M1, M2, M3], and 86 TP samples reached by [M1,
M2, M3, M4] pairs. The most misclassified instances are
observed from Myocardial labels, whereas the best
classification is attained for COVID-19 labels.
TABLE I: INDIVIDUAL DCNNS PERFORMANCE WITH RESPECT TO ECG LABELS.
Label /
Evaluation
Metrics
VGG16 (M1)
DenseNet201 (M2)
MobileNetV2 (M3)
ResNet152V2 (M4)
Acc.
Rec.
F1-s
Spe.
Acc.
Rec.
F1-s
Spe.
Acc.
Rec.
F1-s
Spe.
Acc.
Rec.
F1-s
Spe.
Abnormal
Heart Beats
82
56
60.87
90.66
88
68
73.91
94.66
87
0.84
76.36
88
86
80
74.07
88
Covid_19
100
100
100
100
99
96
97.99
100
100
100
100
100
98
92
95.83
100
Myocardial
82
32
47.06
98.66
88
88
78.57
88
90
76
79.16
94.66
86
60
68.78
94.66
Normal
76
92
65.71
70.66
93
84
85.71
96
91
76
80.85
96
88
84
77.77
89.33
Note: Evaluation Metrics Values in %.
2-D ECG-Dataset
Pre-processing
Multi model Feature integration
Ensemble Deep Feature Pipeline
Deep DCNNs Feature Extractors
VGG16
DenseNet201
MobileNetV2
ResNet152V2
Feature Selection
Classification Results
Abnormal heartbats
COVID19
Myocardial
Healthy
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on February 01,2023 at 07:00:51 UTC from IEEE Xplore. Restrictions apply.
TABLE II: MACRO AVG. PERFORMANCE OF ENSEMBLE DCNNS.
Model
Accuracy (%)
Recall (%)
F1-score (%)
Specificity (%)
Precision (%)
AUC-ROC
-Score (%)
95% CI
[M1, M2]
92
84
83.87
94.66
83.95
89.33
(0.7681, 0.9118)
[M1, M3]
91.5
83
83.03
94.33
83.61
88.66
(0.7563, 0.9036)
[M1, M4]
90.5
81
80.47
93.66
81.73
87.33
(0.7331,0.8868)
[M2, M3]
93.5
87
87.03
95.67
87.18
95.33
(0.8040,0.9359)
[M2, M4]
92.5
85
85.02
95
85.43
90
(0.78001,0.91999)
[M3, M4]
92
84
84.10
94.66
85.26
89.33
(0.76815,0.91185)
[M1, M2, M3]
93.5
87
87.03
95.67
87.18
95.33
(0.80408,0.93592)
[M1, M2, M4]
91.5
83
83.02
94.33
83.16
88.66
(0.75638,0.90362)
[M1, M3, M4]
92.5
85
85.07
89.99
85.53
95
(0.78001,0.91999)
[M2, M3, M4]
92.5
85
85.07
89.99
85.49
95
(0.78001,0.91999)
[M1, M2, M3, M4]
93
86
85.99
90.66
86.64
95.33
(0.79199,0.92801)
TABLE III: COMPARATIVE ANALYSIS OF SOTA METHODS FOR COVID-19 CLASSIFICATION USING ECG.
Reference
Technique
Dataset
Performance Analysis
Anwar et al. [12]
EfficientNet
ECG dataset of Cardiac[15]
Accuracy:81.8%, Sensitivity:75.8%, Precision:80.8%
Rahman et al. [13]
MobileNet
ECG dataset of Cardiac [15]
Accuracy:90.79%, Sensitivity:90.8%, Precision:91.3%,
Specificity:92.8%
Irmak [14]
CNN
ECG dataset of Cardiac[15]
Accuracy:83.05%
Our Proposed
Method
ECG-CCNet
ECG dataset of Cardiac [15]
Accuracy:93.5%, Recall:87%, F1-score:87.03%,
Specificity:95.66%, Precision:87.16%, AUC:95.33%
(a) [M2, M3]
(b) [M1, M2, M3]
(c) [M1, M2, M3, M4]
Fig. 3. Best confusion matrix attained at DCNN ensembles.
Superior Area Under the ROC curve (AUC), values are
essential in healthcare diagnostics. Therefore, its
computations in healthcare analysis help the data analyst ease
the analysis with forecasting on clinical research. The
fundamental goal of ROCs is to demonstrate the classifier's
diagnosing performance while the discriminating criterion is
modified. AUC is the measurement of the full two-
dimensional region beneath the complete Graph (think
integral calculus) between (0,0) to (1,1); AUC=0.5 indicates
no discriminating (i.e., the capacity to classify individuals
having and with no illness or disease based on the test), 0.7
to 0.8 is deemed good, 0.8 to 0.9 is great, and >0.9 is rated
outstanding [31]. The defined DL modalities are estimated
using the AUC in the testing stage. At the testing phase for
individual labels, the AUC and highest True Positive Rate,
i.e., 100% attained for COVID-19, followed by 90% for
normal/healthy, 89% for abnormal heartbeats, 84% for
myocardial labels. The overall best AUC of 95.33% was
attained by [M1, M2, M3, M4] pair.
Apart from our experimented results with individual
DCNNs and all possible ensemble model pairs, we have
equated our best-attained outcomes with the state of the art
methods(the same datasets). Based on comparative analysis
with SOTA and methods shown in Table 3, our proposed
method (ECG-CCNet) proves the best macro average
performance with 93.5% accuracy, recall of 87%, and the F1-
score of the above 87.03%, specificity of 95.66%, precision
of 87.16% and 95.33% AUC.
(a) [M2, M3]
predicted labels
true labels
predicted labels
true labels
predicted labels
true labels
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on February 01,2023 at 07:00:51 UTC from IEEE Xplore. Restrictions apply.
(b) [M1, M2, M3]
(c) [M1, M2, M3, M4]
Fig. 4. Best ROC curve attained at DCNNs ensemble.
V. CONCLUSION
The study aims to create a quick, reliable diagnostic of
cardiovascular, COVID-19 illness using ECG tracing
pictures with DL-approaches. Relying on experimental
findings demonstrating that COVID-19 illness produces
alterations in the heart and lungs portion. With an average
AUC of 95.33%, a unique ECG-CCNet model effectively
separates COVID-19 from healthy, abnormal-heartbeats, and
myocardial instances. The suggested strategy surpasses well-
known CNN methods and shows encouraging results.
Though there has been much excellent research for
identifying COVID-19 illness using radiography pictures.
This totally automated experiment is further useful in the
context of an epidemic since it utilizes ECG, which is
quicker, simpler, and less expensive than radiographic
imaging. In the future, COVID19 variants classification
(Deltacron, Omicron, BA.2, Delta, Beta, Alpha) will be
studied, and the diagnostics approach described in this study
can be proactively employed over epidemic eras by being
loaded on smartphone, IoT devices in conjunction with real-
time AI and cloud technologies.
DECLARATIONS
No datasets are generated; publicly accessible datasets [15] are used
for experiments. The authors have no competing interests to declare.
REFERENCES
[1] Di Gennaro F et al. (2020) Coronavirus diseases (COVID-19) current
status and future perspectives: a narrative review. Int J Environ Res
Public Health.
[2] John Hopkins University Coronavirus Resource Center.
coronavirus.jhu.edu/map.html. Accessed 16 June 2021
[3] World Health Organization. [Online] www.who.int.
[4] S. Chauhan, Comprehensive review of coronavirus disease 2019
(COVID-19), Biomed. J. 43 (2020) 334–340.
[5] Bhosale, Y.H., Patnaik, K.S. Application of Deep Learning Techniques
in Diagnosis of Covid-19 (Coronavirus): A Systematic Review. Neural
Process Lett (2022).
[6] T. Ai, Z. Yang, H. Hou, et al., Correlation of chest CT and RT-PCR
testing in coronavirus disease 2019 (COVID-19) in China: a report of
1014 cases, Radiology 296 (2020) E32–E40.
[7] Y. Pathak, P.K. Shukla, A. Tiwari, S. Stalin, S. Singh, P.K. Shukla,
Deep Transfer Learning Based Classification Model for COVID-19
Disease, IRBM-Vol-43, Issue 2, 2022, Pages 87-92, ISSN 1959-0318.
[8] B. Abhijit, et al., A deep learning based approach for automatic
detection of COVID-19 cases using chest X-ray images, Biomed.
Signal Process Control 71 (Part B) (2022) 103182.
[9] RS Soumya, TG Unni, KG Raghu, Impact of COVID-19 on the
cardiovascular system: a review of available reports, Cardiovasc.
Drugs Ther. 35 (2021) 411–425.
[10] R.B. Azevedo, B.G. Botelho, J.V.G. de Hollanda, et al., Covid-19 and
the cardiovascular system: a comprehensive review, J. Hum.
Hypertens. 35 (2021) 4–11.
[11] B. Predabon, et al., The electrocardiogram in the differential diagnosis
of cardiologic conditions related to the covid-19 pandemic, J. Cardiac
Arrhythmias 33 (2020) 133–141.
[12] T. Anwar, S. Zakir, Effect of image augmentation on ECG image
classification using deep learning, in: 2021 International Conference
on Artificial Intelligence (ICAI), IEEE, 2021, pp. 182–186.
[13] Rahman T, Akinbi A, Chowdhury ME, et al. COV-ECGNET: COVID-
19 Detection Using ECG Trace Images with DCNN. arXiv:210600436
[14] Irmak, E. COVID-19 disease diagnosis from paper-based ECG trace
image data using a novel convolutional neural network model. Phys
Eng Sci Med 45, 167–179 (2022).
[15] Ali Haider Khan, et ECG Images dataset of Cardiac and COVID-19
Patients, Data in Brief, Volume 34, 2021, 106762, ISSN 2352-3409.
[16] S. Mishra, G. Khatwani, et al., ECG paper record digitization and
diagnosis using deep learning, J. Med. Biol. Eng. 41 (2021) 422–432.
[17] G. Sannino, G. De Pietro, A deep learning approach for ECG-based
heartbeat classification for arrhythmia detection, Future Generat.
Comput. Syst. 86 (2018) 446–455.
[18] L.E. Widman, G.L. Freeman, A-to-D conversion from paper records
with a desktop scanner and a microcomputer, Comput. Biomed. Res.
22 (1989) 393–404.
[19] Ghosh, A., Pahari, P., Basak, P. et al. Epileptic-seizure onset detection
using PARAFAC model with cross-wavelet transformation on multi-
channel EEG. Phys Eng Sci Med (2022).
[20] Sahoo, P., Roy, I., Ahlawat, R. et al. Potential diagnosis of COVID-19
from chest X-ray and CT findings using semi-supervised learning. Phys
Eng Sci Med 45, 31–42 (2022).
[21] Bhosale YH. “Digitization Of Households With Population Using
Cluster And List Sampling Frame In Aerial Images”, ISSN (Online)
2456-3293, vol 5., issue 2, pp. 22–26, Feb 2020.
[22] Bassiouni, et al. Detection of COVID-19 Using DL Approaches with
ECG Reports. Circuits Syst Signal Process (2022).
[23] Ioannis A, Mpesiana T (2020) Covid-19: automatic detection from X-
ray images utilizing transfer learning with convolutional neural
networks. Phys Eng Sci Med 43(2):635–640.
[24] Gaurav Srivastava, Aninditaa Chauhan, Mahesh Jangid, Sandeep
Chaurasia, CoviXNet: A novel and efficient deep learning model for
detection of COVID-19 using chest X-Ray images, Biomedical Signal
Processing and Control, Volume 78, 2022, 103848, ISSN 1746-8094.
[25] Jain R., et al.. Deep learning based detection and analysis of COVID-
19 on chest X-ray images Appl. Intell., 51 (3) (2021), pp. 1690-1700
[26] Panwar H,et al.Application of deep learning for fast detection of
COVID-19 in XRays using nCOVnet Chaos Solitons Fractals, 138.
[27] Basu S., Mitra S., Saha N. Deep learning for screening covid-19 using
chest x-ray images 2020 IEEE Symposium Series on Computational
Intelligence, SSCI
[28] Luz E., Silva P., et. al. Towards an effective and efficient deep learning
model for COVID-19 patterns detection in X-ray images Res. Biomed.
Eng. (2021), pp. 1-14
[29] F. Badilini, T. et al., ECGscan: a method for digitizing paper ECG
printouts, J. Electrocardiol. 36 (2003) 39.
[30] L.E. Widman, G.L. Freeman, A-to-D conversion from paper records
with a desktop scanner and a microcomputer, Comput. Biomed. Res.
22 (1989) 393–404.
[31] Jayawant N. Mandrekar, Receiver Operating Characteristic Curve in
Diagnostic Test Assessment, Journal of Thoracic Oncology, Volume 5,
Issue 9, 2010, Pages 1315-1316, ISSN 1556-0864.
[32] Y. H. Bhosale, K. Sridhar Patnaik, "IoT Deployable Lightweight Deep
Learning Application For COVID-19 Detection With Lung Diseases
Using RaspberryPi," IEEE-ICIBT-2022, pp. 1-6
[33] Y. H. Bhosale, et al., "Deep Convolutional Neural Network Based
Covid-19 Classification From Radiology X-Ray Images For IoT
Enabled Devices," IEEE-ICACCS-2022, pp. 1398-1402.
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on February 01,2023 at 07:00:51 UTC from IEEE Xplore. Restrictions apply.
