International Journal of

INTELLIGENT SYSTEMS AND APPLICATIONS IN

ENGINEERING
ISSN:2147-67992147-6799 www.ijisae.org Original Research Paper

A Machine Learning Approach for Early Identification and Prevention

of Covid-19 like Global Pandemics
Mr. Sandeep Kumar Mathariya*1, Dr. Hemang Shrivastava 2

Submitted: 25/01/2024 Revised: 03/03/2024 Accepted: 11/03/2024

Abstract: In addition to causing enormous global financial and social losses, the COVID-19 pandemic was a once-in-a-century
occurrence that claimed a great deal of human lives. Reducing the number of victims can be achieved in part by precisely and early
detection of the disease. The unanticipated surge in cases has also resulted in severe limits on the number of scans conducted and the
amount of time radiologists may spend analyzing the data to assess the severity of the diseases and potential advancements in the future.
Therefore, automated procedures that might lessen the burden on the healthcare system to offer prompt and accurate diagnoses are being
researched. Furthermore, determining future possible hotspots for these diseases might aid in the selection of micro-containment zones,
which can be a proactive measure to impede the disease's rapid spread. Numerous methods based on deep learning and machine learning
have been researched for picture categorization, finding possible hotspots, or both. This work offers a data-driven approach to categories
Covid-19 images and pinpoint possible pandemic hotspots, which will help with treatment and prevent pandemics from spreading in the
future. Moreover, a deep neural network based regression model has been developed which predicts the number of new active cases in
the days ahead. This approach is allows for early assessment of the disease spread and allows for timely management and early
prevention of widespread pandemics. To assess the effectiveness of the suggested method, a comparison with the most recent iterations
of deep learning and machine learning algorithms has been provided. Findings show that, when compared to baseline methods, the
suggested data-driven strategy with probabilistic categorization performed better. Moreover, the proposed regression model outperforms
baseline state of the art models in terms of accuracy of prediction..

Keywords: Coronavirus, COVID-19 Pandemic, Automated Detection, Probabilistic Classification, Deep Neural Networks.

1. Introduction services exposed several weaknesses. To get ready for future

pandemics, investments in healthcare infrastructure are required.
In contemporary times, human encounters with pandemics have These include having enough medical supplies, qualified medical
been sporadic, with the most recent significant outbreak being the workers, and adaptable systems that can manage spikes in cases
Spanish flu in the previous century [1]. The emergence of the [7]. Synchronization and international collaboration are essential.
Covid-19 pandemic marked a catastrophic event of Today's interconnected world means that combating pandemics
unprecedented scale, resulting in significant loss of life [2]. This need collaboration from all stakeholders. Improving international
crisis led to a substantial financial downturn and economic collaboration in areas like immunization distribution, data
slowdown, the ramifications of which continue to be felt today interchange, and resource allocation will be crucial to managing
[3]. As the pandemic gradually waned due to widespread and preventing future health catastrophes. [8].
containment efforts, vaccine development, and the establishment
of herd immunity, it became evident that a more proactive Advancements in vaccine research, antiviral therapy, and
approach is imperative to effectively tackle similar future crises. diagnostic technologies are crucial for pandemic preparedness
There are valuable lessons to be gleaned from this current [9]. Sustaining dedication to research and innovation will
catastrophe to enhance anticipation and preparedness for future enhance global readiness to address future pandemics [10]. A
pandemics [4]. The importance of early detection and swift action significant discovery from the current COVID-19 pandemic is the
is underscored as a crucial lesson. The delayed recognition of the utilization of data-driven machine learning models for both
severity of COVID-19 allowed it to spread globally before diagnosis and strategic planning of containment zones, effectively
adequate countermeasures could be implemented [5]. To expedite curbing the widespread and rapid transmission of the disease.
the identification and mitigation of future threats, it is essential Figure 1 depicts the digital technology based framework for
for pandemic preparedness efforts to prioritize the enhancement surveillance. and detection of pandemics like Covid-19.
of surveillance systems, international collaboration, and
information exchange [6]. This paper aims at developing a machine learning model which
can identify potential hotspots and predict future rate of increase
The strain that the COVID-19 epidemic placed on healthcare of Covid cases for early detection and prevention of the spread of
pandemic likecases. Also a data driven Covid classification
1 Department of Computer Science and Engineering
technique has been proposed [11].
1 Research Scholar, SAGE University, Indore, (M.P) India
ORCID ID : 0000-0002-3795-5455 COVID-19 detection utilizing Bayesian networks involves a
2 Research Supervisor, SAGE University, Indore, (M.P.), India sophisticated data-driven approach that integrates various factors
2drhemang.shrivastava@sageuniversity.in to assess the likelihood of infection.
* Corresponding Author Email: mathariya@gmaill.com

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3140–3151 | 3140
 have shown particular efficacy in automatically extracting
relevant features from medical images. By analyzing subtle
textures, patterns, and structural changes indicative of COVID-19
infection, these models can provide rapid and accurate
assessments, assisting radiologists and clinicians in making
informed decisions. Additionally, the use of ML and DL for
COVID-19 detection in imaging data offers the potential for
scalability and efficiency, enabling healthcare systems to handle
large volumes of cases more effectively.
However, continued research and validation are necessary to
ensure the reliability and generalizability of these models across
diverse populations and imaging protocols. Overall, leveraging
ML and DL techniques for COVID-19 detection in CT and X-ray
images holds promise for improving diagnostic accuracy and
aiding in the management of the pandemic.

Fig.1 Digital Technology infrastructure for pandemic detection

and surveillance.

. By incorporating information such as symptoms, exposure

history, and test results, Bayesian networks provide a
probabilistic framework for analyzing and predicting COVID-19
cases. These networks consider the conditional dependencies
between different variables, allowing for more accurate and
nuanced assessments of the disease's presence. Through Bayesian
inference, the model can update its predictions as new data
becomes available, enabling healthcare professionals to make
informed decisions regarding testing, treatment, and containment
measures. This approach not only enhances the efficiency of
COVID-19 detection but also contributes to a more
comprehensive understanding of the pandemic's dynamics and
spread.

Detecting COVID-19 hotspots through machine learning involves

leveraging large datasets and sophisticated algorithms to identify
regions with heightened transmission rates. Machine learning
models analyze various factors such as population density,
mobility patterns, case counts, and demographic information to
predict areas at higher risk of outbreaks. These models can
uncover hidden patterns and correlations that may not be
immediately apparent to human observers, enabling proactive
measures to be implemented in vulnerable areas. By continuously
updating and refining their predictions based on real-time data,
machine learning algorithms empower authorities to allocate
resources more effectively, implement targeted interventions, and
mitigate the spread of the virus. This data-driven approach to
hotspot detection not only aids in controlling the current
pandemic but also lays the groundwork for better preparedness Fig.2 A Weekly Surveillance of Covid Hotspot Clusters in USA
and response to future outbreak. (hohl et al., spatial and spatio-temporal epidemiology (2020)

2. Background Figure 2 depicts the case distribution (weekly) for global

pandemic distribution during Covid-19.
2.1 Image Based Classification
2.2 Prediction of Covid Hotspots
COVID-19 detection through the analysis of CT and X-ray
images using machine learning (ML) and deep learning (DL)
Machine learning models for detecting hotspots play a crucial
models has emerged as a promising approach to aid in diagnosis.
role in identifying regions or areas with heightened activity or
These imaging techniques offer valuable insights into the
risk related to various phenomena, including disease outbreaks,
presence and severity of lung abnormalities associated with the
natural disasters, and social unrest. These models leverage large
virus. ML and DL models are trained on large datasets of
datasets containing information such as historical records,
annotated CT and X-ray images, learning to distinguish between
demographic data, environmental factors, and real-time
normal and COVID-19-infected lungs based on visual patterns
observations to predict the likelihood of hotspot occurrences. One
and features.
commonly used approach involves clustering algorithms, which
Convolutional neural networks (CNNs), a type of DL model,
group spatial data points based on similarity, allowing for the

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3141
identification of clusters or concentrations of events. 3. Methodology:
Another approach involves spatiotemporal modeling, where
The methodology has a addresses 3 major challenges to Covid-19
machine learning algorithms analyze both spatial and temporal
like pandemics, which are:
patterns to predict the emergence and evolution of hotspots over
1. Data driven models to classify the diseases.
time. These models can incorporate diverse data sources,
2. Identifying potential hotspots in future.
including satellite imagery, social media feeds, and sensor data,
3. Predicting new active cases in future.
to provide comprehensive insights into hotspot
These three approaches are expected to result in early detection
dynamics.Additionally, ensemble methods, which combine
and prevention of Covid like pandemics.
predictions from multiple models, can enhance the robustness and
accuracy of hotspot detection. By identifying hotspots early,
The proposed approach entails image-processing followed by
decision-makers can allocate resources, implement targeted
feature extraction and classification to classify positive and
interventions, and mitigate potential risks, ultimately enhancing
negative cases [24]. While several deep learning approaches such
public safety and resilience. However, ongoing research and
as variants of CNNs are available at our disposal, yet the copious
validation are essential to ensure the reliability and effectiveness
amounts of data and processing power needed to train the
of machine learning models for hotspot detection across different
algorithms is extremely large to attain significantly high
contexts and applications.
classification accuracy [25]. Alternatively an image processing
based approach coupled with feature extraction is presented in
2.3 Prediction of future Covid Cases using Machine learning.
this paper so as to overcome the challenge of extremely large
datasets [26].
Predicting future COVID-19 cases using machine learning
involves leveraging historical data, epidemiological factors, and
3.1 Data Pre-Processing
other relevant features to forecast the spread of the virus over
The first step though is the removal of noise from raw captured
time. One common approach is time series forecasting, where
images, whose sources can be:
machine learning algorithms analyze trends and patterns in past
case counts to make predictions about future numbers of
1) Addition of electronic noise in the image due to the use
infections.
of amplifiers in the sensing device which is also termed
These models can incorporate various input variables such as
as white or Gaussian noise [27].
population density, mobility data, government interventions, and
2) The abrupt change or spikes in the analog to digital
healthcare capacity to improve the accuracy of predictions.
converters sued in the circuity of the fundus image
Additionally, ensemble methods and deep learning techniques
causing salt and pepper noise patterns [28].
such as recurrent neural networks (RNNs) can capture complex
3) The multiplicative noise effect due to the inconsistent
temporal dependencies and nonlinear relationships in the data,
gain of the adaptive gain control (AGC) circuity used
enhancing the predictive capabilities of the models.
for capturing or retrieving the fundus image29 [].
Continuous updating of the models with new data allows for
4) The lack of pixels while capturing the image resulting
adaptive forecasting, enabling authorities to adjust strategies and
in frequency mean valued interpolations in the
allocate resources in response to changing dynamics. While
reconstructed image causing Poisson image [30]-[31].
machine learning models for predicting COVID-19 cases have
shown promise, it's essential to consider uncertainties and
The removal of noise effects is fundamentally important as noisy
limitations inherent in the data and model assumptions.
images would result in erroneous feature extraction leading to
Moreover, ongoing validation and refinement of these models are
inaccurate classification of the CT/MRI images [32]. Contrary to
crucial for ensuring their reliability and effectiveness in
conventional Fourier based methods, the wavelet transform is
supporting decision-making and public health interventions.
made us of non-smooth Kernel functions such as Mayer, Haar,
Coif etc [33]-[34]. The essence of the transform lies in the fact
Machine learning (ML) has been instrumental in developing
that the wavelet transform separates the low frequency and high
various solutions to address the challenges posed by COVID-
frequency components as the approximate co-efficient ) and
19.ML techniques, particularly time-series analysis and deep
detailed co-efficient ) of the transform. Generally, detailed co-
learning, have been used to model the spread of COVID-19,
efficient ) and a low frequency resolution component termed
predict case trajectories, and forecast healthcare resource needs.
as the detailed co-efficient ) [35]..
These models take into account factors such as population
demographics, mobility patterns, interventions, and
Retaining the low frequency component ) while discarding the
environmental factors to provide insights for policymakers and
higher frequency component ) for a number of iterations helps
healthcare providers. ML models can predict individual risk
in removal of the baseline noise of the system [36]. One of the
factors for COVID-19 complications based on patient
most effective hyperspectral image restoration techniques is
demographics, comorbidities, and other clinical variables. These
based on the sub-band decomposition of images into low pass and
models enable personalized treatment plans and interventions for
high pass signal values using the wavelet transform [37]. The
high-risk individuals, leading to better outcomes.
wavelet transform, unlike the conventional Fourier methods uses
Overall, machine learning plays a critical role in advancing our
non-linear and abruptly changing kernel functions which show
understanding of COVID-19, developing effective interventions,
efficacy in analysing abruptly fluctuating signals such as images.
and guiding public health responses to mitigate the impact of the
The continuous and the discrete wavelet transforms are computed
pandemic, as well as early detection and prevention of such
as:
pandemics in future.

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3142
 (6)
(1)
Where, c) Energy which is also considered as the secondary
moment:
represent the scaling (dilation) and shifting (translation)
constants constrained to the condition .
(7)
is the Wavelet Family or Mother Wavelet
is the time variable d) Variance is the squared value of s.d. given by:
is the time domain data.
(8)
For implementing the wavelet transform on the image dataset, the
sampled version of the continuous wavelet transform yields the
discrete wavelet transform given by:
e) Contrast which is the deviation among the mean and
differential change in illuminance:
(2)
Where,
(9)
is the discrete vector.
is the discrete scaling constant. a) Entropy which is the statistical average information
content defined as:
is the discrete shifting constant.

The discrete wavelet transform yields two distinct low and high (10)
pass values based on the number of levels of decomposition and
wavelet family given by the approximate co-efficient (CA) and b) Homogeneity which is the similarity among the pixel
detailed co-efficient (CD) [38]. The approximate co-efficient value distribution:
values are typically the low pass values containing the maximum
information content of the image while the detailed co-efficient
values account for the noisy spectral part. Retaining the low pass (11)
co-efficients and recursively discarding the high pass co-efficients
allows to de-noise the image [39]. The choice of the wavelet c) Correlation which is the similarity overlap among pixel
family impacts the estimation of the noise gradient vector given values:
by [40]:

(3) (12)

The value of the second order normalizing gradient as a function d) Root Mean Square Value which is defined as the
of spatial co-ordinates is given by: squared root of the squared mean of values in the
random distribution defined as:

(4)
(13)
Here,
The normalizing factor for the gray covariance matrix (GLCM) is
denotes the original image. defined as:
denotes the fused image after normalization.
denotes the normalizing gradient. (14)

represents the gradient.

Here,
represents the Laplacian.
denotes random variable X
3.2 Feature Extraction denotes the frequency of occurrence
The next step is the statistical feature extraction of image
features expressed as [41]: denotes the image

a) Mean or average value: denotes pixels

(5) denotes avg. illuminance

A denotes the amplitude
b) Standard Deviation: denotes levels of normalized GLCM matrix
denotes the normalized GLCM matrix

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3143
 Step.9: Compute
denotes probability
Step.10: Compute
The design of the automated classifier is critically important as
the accuracy of classification critically depends on the design of Step.11:
the classifier. Generally, positive and negative CT/MRI images Step.12: Truncate training
show overlapping feature values. Hence a probabilistic approach Step.13:
is often effective. An image processing is practically carried out
Step.14: Compute MSE, MAPE,
in hospitals to medical facilities, hence deep learning algorithms
which need large computational resources may render Step.15:
infeasibility to even a novel and accurate approach [42]. This Step.16:
leads to a natural inclination towards the Bayesian Regularization
algorithm [43].
the labelled data vector is fed to The performance metrics to be computed are [47]:

Accuracy (Ac): It is mathematically defined as:

the bayesian regularized ann. the brann is chosen as it is an
effective classifier. it works on the principle of Baye’s theorem of
conditional probability. After the BRANN is trained, in the (17)
testing phase, the BRANN calculates the probability of an
element to belong to a particular category [44]. For a multi-class Recall: It is mathematically defined as:
decision, the higher probability of a particular class decides the
category of the data. In case of the BRANN tries to find out the (18)
probability of an image to be actually positive based on the
probability before passing the judgement. For this, the important
assumption which the BRANN makes is that of the accuracy of Precision: It is mathematically defined as:
the classifier [45]. This is dependent on the training accuracy
which is available to the classier (on completion of training) and (19)
the number of positive images in the dataset (already available to
the classifier as the dataset provided by the user, which the F-Measure: It is mathematically defined as:
classifier assumes to be true). The same logic applies to the
negative images [46].
(20)
The training rule for the approach is based on the Bayes theorem
of conditional probability which is effective for classifying Here,
overlapping feature vectors, based on a penalty . The TP, TN, FP and FN denote the true positive, true negative, false
positive and false negative rates respectively.
weights are updated based on the modified regularized cost
function [47]:
Bayesian networks have been explored for COVID-19 detection
] (15) as they offer a probabilistic framework to model the relationships
between various factors associated with the disease. In this
If Network error are generally low. context, Bayesian networks can integrate information from
different sources such as symptoms, exposure history,
demographic data, and test results to assess the likelihood of
else if Network errors tend to increase, in which case COVID-19 infection in individuals [48].The structure of a
Bayesian network represents the dependencies between variables,
the weight magnitude should be reduced so as to limit and the network is augmented with probabilities that quantify the
errors (Penalty). likelihood of certain events given the values of other variables.
This is done be maximizing the weight Posteriori Probability For COVID-19 detection, the network might include nodes
using the Bayes theorem of Conditional Probability as: representing symptoms (e.g., fever, cough, shortness of breath),
clinical findings (e.g., lung abnormalities on imaging), test results
(16) (e.g., PCR or antigen test outcomes), and risk factors (e.g., age,
comorbidities, recent travel or exposure).By updating the
The proposed training algorithm is presented next. probabilities in the network based on observed data (e.g., new
symptoms reported by the patient, test results), Bayesian
3.3 Proposed Algorithm: inference can be used to calculate the posterior probability of
COVID-19 infection. This approach allows for a more nuanced
Step.1: Initialize weights and learning rate randomly, assessment that takes into account the interplay between different
set maximum iterations as , factors and their uncertainties.
One advantage of Bayesian networks is their ability to handle
Step.2: , do incomplete or noisy data, making them suitable for real-world
Step.3: , medical applications where data may be sparse or uncertain.
Additionally, the transparency of Bayesian networks enables
Step.4: Retain ) while discarding )
clinicians to interpret the reasoning behind the model's
Step.5: ) and compute predictions, which can enhance trust and facilitate clinical
decision-making [49].However, developing a robust Bayesian
Step.6: Compute network for COVID-19 detection requires careful consideration
Step.7: && of the variables to include, the structure of the network, and the
Step.8: Minimize:
estimation of accurate probabilities from data. Moreover, ongoing
validation and refinement are essential to ensure the reliability

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3144
and effectiveness of the model in diverse populations and clinical 2023 has been used to predict future active cases, through
settings. regression analysis. The MAPE and values have been used for
performance evaluation.
3.4 Regression Learning
Managing pandemic like situations need estimate of the future
number of active cases so as to deviceapproaches for
management and effective prevention [50]. The following
parameters have been considered while predicting future cases:

1. Date
2. Confirmed Cases
3. Deaths
4. Recovered
5. Active New Cases

The target variable happens to be the active new cases. While

several other parameters may affect the cause of increase or
decrease in cases, such as demographics, spatial coordinates etc.,
the empirical attributes have been used for analysis. A regression
learning model has been developed for the purpose employing
deep neural networks.
1) Mean Absolute Percentage Error
2) Number of Iterations

The approximated 2nd order derivative based learning

algorithmcomputing the Hessian Matrix is employed to update
the weights of the network I n each iterations, given by [51]:

(21)

Here, Fig.3 Original Image

denote the weights of the present and subsequent
iterations.
denotes the learning rate.
denotes the error in the present iterations.

4. Experimental Results
The experiments are carried out of MATLAB 2022a with the
Deep Learning Toolbox. The experiment is performed on a
Windows Machine with i5-9300H Processor enabled with
NVIDIA GTX GPU and RAM of 8GB. The software package
used in Matrix Laboratory (Matlab), 2022a.

The classification of images into Covid positive and negative

cases has been done using an annotated dataset with 1000 images
comprising of both positive and negative cases of Covid have
been used. The images are .jpg images in this experimental setup,
but the system is compatible with all other common image data
types such as .png, .tiff etc. The images acquired are .jpg images
which three colour channels viz. R, G and B. All the images are
first converted to common dimensions of (256 x 256). The
features are then used to train a Deep Bayes Net. The
performance metrics chosen are the accuracy of classification and
prediction error in terms of the TP, TN, FP and FN values
respectively. The images are presented in the sequence of
occurrence in the experimental setup followed by a detailed Fig.4Segmentation
explanation and significance of each image obtained at each step.
Figure 3 depicts the original image while figure 4 depicts the
The hotspot identification results have been evaluated in terms of segmented image, where the affected region in localized.
the classification accuracy computed through the confusion
matrix. Geographical location and number of cases have been
used a independent variables to identify hotspots.

Finally, a six month span ranging from January 2020 to July

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3145
 After the DWT decomposition, the feature values of the images
are to be computed as defined in the feature extraction section.
The feature values computed from the images need to be fed to
the proposed machine learning model for pattern recognition for
both positive and negative cases. As an illustration, two separate
CT images have been analysed using the proposed algorithm and
their features have been tabulated in table 2. It can be observed
that the statistical feature values have identical values for both
positive and negative cases of covid, which necessitates the use
of an accurate classifier.

Table.2 Image Features

Features Normal Fundus Image Glaucoma Fundus

Image
Contrast 0.362500000000000 0.328409090909091

Correlation 0.151836572326215 0.179081293145884

Energy 0.700218879132232 0.723267045454546

Homogeneity 0.915729166666667 0.921979166666667

Mean 0.00652645881744487 0.00227212218788613

Fig.5Variations in Segmentation based of threshold Standard 0.106429183594269 0.106604988215597

Deviation
Figure 5 depicts the various segmentation results on the image Entropy 3.53356796596205 3.43675888527571
under interest.
RMS 0.106600358177805 0.106600358177805
The statistical features of the wavelet decomposition tabulated in
table 2 followed by the image features computed subsequent to Variance 0.0111929089863435 0.0111892474804911
the DWT decomposition, in table 1.
Smoothness 0.923435573787053 0.807650945366392
Table.1 Statistical Analysis
S.No. Parameter Values Class Kurtosis 6.63089615157184 6.70718347384146
1. Minimum 0
Skewness 0.512491012022217 0.483428823467213
2. Maximum 0.9295
Original Image
3. Mean 0.3415
Now, one of the most effective ways to ensure the accuracy and
4. Median 0.3703 coherence of the feature extraction process is the distribution
analysis of feature values for a multitude of images. Although
5. Standard 0.1524 there can be variations in images of a particular class of image in
Deviation
any dataset, yet the feature values should depict a certain amount
6 Mean Absolute 0.05355
Deviation
of coherence over a multitude of images. Thus the feature sets for
7. Minimum 0.002951 multiple images in the dataset (taken 10 for the sake of brevity
and ease of analysis) has been depicted in figure 6. Figure 12
8. Maximum 0.9165 Approximate Co- clearly indicates a similarity in the values of the features
efficient values extracted from the images. For instance, feature 1 (for all the 10
9. Mean 0.3415
images) clearly show a much lesser magnitude of value compared
10. Median 0.3706 to feature 11 as depicted in figure 6.

11. Standard 0.1511

Deviation
12. Mean Absolute 0.05354
Deviation
13. Minimum -0.1592
14. Maximum 0.1592 Detailed Co-efficient
values
15. Mean 0

16. Median 0

17. Standard 0.01232

Deviation
18. Mean Absolute 0.005539
Deviation

Fig.6 Distribution of feature values

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3146
A similar analysis can be done using the mesh plot depicted in
figure 7. The mesh renders a three dimensional view of the
feature value distribution. A coherent inference can be drawn
from figures 6 and 7 depicting the fact that the feature
distribution for similar class of features over a multitude of
images show similarity in values. For instance, feature 1 (for
multiple images) shows much lesser value compared to feature 11
(for multiple images). This is the exact same illustration taken in
figure 6.

Fig.8 Confusion Matrix for Image Classification

The performance of the proposed approach is can be evaluated in

terms of the true positive (TP), true negative (TN), false positive
(FP) and false negative (FN) rates of the confusion matrix. Out of
the 1000 images of the dataset, 70% i.e. 700 images have been
used for training and the rest of the 30% i.e. 300 images have
been used for testing. The TP, TN, FP and FN values are depicted
in the Confusion Matrix in figure 9. Based on the TP, TN, FP and
FN values, the accuracy, sensitivity/recall, specificity, precision
and F-Measure values have been computed and tabulated in table
5. The proposed approach attains an accuracy of 0.98, sensitivity
or recall value of 0.9866, specificity of 0.9733, precision of
0.9736 and F-Measure of 0.9800. The mean training time for the
proposed approach has been 12 minutes for the training dataset
alone.
Fig.7Mesh Plot of features
Table 3 Performance Metrics
As machine learning algorithms often suffer from imbalanced
instances, this phenomenon has been carefully considered while Accuracy% Sensitivity Specificity% Precision% F-
Or Measure
data set preparation. Imbalanced instances to imbalanced class Recall%
distributions occur when the samples or observations of one of
0.980 0.9866 .9733 .9736 .9800
the classes is either much higher or much lower compared to the
other class or classes. This may result in misleading results as
machine learning algorithms tend to statistically ignore the class
distributions. This caveat is eliminated in this experiment by
following almost an equal share of positive and negative image
classes. This can also been seen in the confusion matrix for the
testing case. As the positive and negative cases have an identical
class distribution, hence resampling (over sampling, under
sampling or SMOTE (Synthetic Minority Oversampling
Technique) has not been performed.

Three sub cases of the discriminant analysis are again considered

in this experiment which happen to be:

1) Linear Discriminants.
2) Quadratic Discriminants.
3) Optimizable Discriminants.

It is observed that for a parallel pools for the training, the 3 case
feature selection attains convergence in 30 iterations. The linear
discriminant results in an accuracy of 98%, the quadratic
discriminant results in an accuracy of 97.33% and the optimizable
discriminant results in an accuracy of 98%.
Fig.9. Confusion matrix for hotspot identification

While there can be many more parameters which may affect the
area to be a hotspot, the above parameters are the most
significant. The data parameters employed in this study are:

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3147
 1. Age Bracket
2. Gender
3. Detected Area
4. Detected State
5. Current Status
6. Travel History
7. Contacted From
8. Nationality
9. Type of Transmission
10. Identity
11. No. of Cases
12. Containment Zone (target)
The approach attains a classification accuracy of 93.75% for the
hotspot identification.To evaluate the performance of the
proposed work in comparison to the contemporary approaches in
the domain has been presented in table 4.
Fig.10 Raw Data
Table 4 Comparative Analysis w.r.t. existing work
Figure 10 depicts the raw time series data for the active number
Method Accuracy of future cases over a 188 day period from January 2020 to July
Mollaloa et al. 91% 2020. The minimum, maximum and mean have been marked.
(2021)
Kahn et al. (2020) 79%
Alsan et al. (2022) 96.29
Akbarimajd et al. 72%
(2022)
Momeny et al. 80.8%
(2021)
Das et al. (2020) 97.4%
Zebari et al. (2022) 89.87%
Gannour et al. 97%
Proposed 98%

Table 5 presents a comparative study with respect to

contemporary existing work in the domain. It can be observed
from table 5 that the proposed work attains relative higher
accuracy compared to existing methods in the domain. The
improvement in the results can be attributed to the following
reasons: Fig.11 Prediction (Active Cases)
1) Image enhancement employing illumination correction and
histogram normalization compensating inconsistencies in image Figure 11 depicts the forecasting MAPE for the proposed
capturing. approach which happens to be only 0.11%/
2) Iterative noise removal for denoising fundus images for
accurate feature extraction.
3) Computing stochastic feature and subsequent feature
optimization to enhance the training efficacy.
4) While deep learning models such as the CNN, RCNN, ResNet
etc. may have the advantage of avoiding additional effort in
handpicking features and feature combinations, they lose control
over choosing the features to be used to train a model. The
proposed approach with stochastic features to train a Bayesian
Deep Neural Network attains relatively higher accuracy of
classification compared to benchmark techniques. Moreover, the
approach is also capable of identifying potential hotspots based
on geographical and statistical survey features with an accuracy
of 93.75%

The prediction of the Covid-19 cases has been doe based on a

deep neural network architecture trained with five empirical Fig.12 Regression (Active Cases)
parameters which allow us to predict the number of new active
cases in future for Covid. The data statistics are analyzed Figure 12 depicts the regression analysis (overall) for the data
followed by pattern recognition and regression analysis. The time analyzed. The regression can be seen to be 0.9998.
range of January to July 2020 has been chosen for the purpose
which exhibited the most amount of volatility in the number of A comparative analysis with existing baseline approaches has
new cases. The results are presented next. been presented in table 5.

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3148
 Table.5 Comparative Predictive Analysis. References
[1] J. Miah, R. H. Khan, S. Ahmed and M. I. Mahmud, "A comparative
S.No. Parameter Value
study of Detecting Covid 19 by Using Chest X-ray Images– A Deep
1. Sample Span 6 months
2. ML Model Back Propagation Learning Approach," 2023 IEEE World AI IoT Congress (AIIoT),
3. Training Optimized 2nd Order Seattle, WA, USA, 2023, pp. 0311-0316
Hessian [2] A. Hussain, S. U. Amin, H. Lee, A. Khan, N. F. Khan and S. Seo,
4. Iterations to "An Automated Chest X-Ray Image Analysis for Covid-19 and
convergence Pneumonia Diagnosis Using Deep Ensemble Strategy," in IEEE
5. 0.9998 Access, 2023, vol. 11, pp. 97207-97220,.
[3] S. Mishra and A. Mantri, "Design of COVID19 Detection based on
6. MAPE (proposed) 0.11
Relative Eccentric Feature Selection using Deep Vectorized
7. MAPE, Balli. [52] 2.0762 (RF)
0.1853 (LR) Regressive Neural Network for Corona Virus," 2023 IEEE
0.8179 (MLP) International Conference on Integrated Circuits and Communication
0.1247 (SVM) Systems (ICICACS), Raichur, India, 2023, pp. 1-6.
[4] J Somasekar, PP Kumar, A Sharma, “Machine learning and image
A comparative analysis with previous work in terms of baseline analysis applications in the fight against COVID-19 pandemic:
techniques such as Random Forecast (RF), Logistic Regression Datasets, research directions, challenges and opportunities”,
(LR), Multi Layer Perceptron (MLP) and Support Vector Proceedings in Matrials Today, Elsevier 2021.
Machine (SVM) clearly indicates that the proposed deep neural [5] J. De Moura et al., "Deep Convolutional Approaches for the Analysis
network approach with 2nd Order Hessian Optimized back
of COVID-19 Using Chest X-Ray Images From Portable Devices," in
propagation beats the models in terms of MAPE of prediction.
IEEE Access, 2020, vol. 8, pp. 195594-195607
This makes the model more dependable and accurate in terms of
estimating the severity of the disease in future. [6] S Kumar, RD Raut, BE Narkhede,” A proposed collaborative
framework by using artificial intelligence-internet of things (AI-IoT)
in COVID-19 pandemic situation for healthcare workers”,
5. Conclusion International Journal of Healthcare Management, Taylor and Francis,
2020, vol.13, no.4, pp. 337-345.
This paper aims at addressing the pivotal challenges of Covid-19 [7] JB Awotunde, SO Folorunso, RG Jimoh, “Application of artificial
like unprecedented pandemics, wherein it is mandatory to arrest
intelligence for COVID-19 epidemic: an exploratory study,
the spread of the disease through early detection and prevention.
opportunities, challenges, and future prospects”, Artificial
It was observed from the Covid-19 scenario that the most
effective way to reduce the severity and human casualties was to Intelligence for COVID-19. Studies in Systems, Decision and
prevent the spread and increase in new cases in the first place. Control, Springer, 2021, vol 358.
The methodology designed in this paper tries to address the three [8] S. Hu et al., "Weakly Supervised Deep Learning for COVID-19
fundamental challenges pertaining to pandemics which are Infection Detection and Classification From CT Images," in IEEE
automated detection, identification of hotspots and predicting Access, 2020, vol. 8, pp. 118869-118883.
future trends in active new cases. The probabilistic Deep [9] J. C. Clement, V. Ponnusamy, K. C. Sriharipriya and R.
BayesNet has been proposed in this paper for both automated Nandakumar, "A Survey on Mathematical, Machine Learning and
classification of cases as well as for identification of hotspots.
Deep Learning Models for COVID-19 Transmission and Diagnosis,"
Further a regression learning based approach employing the
in IEEE Reviews in Biomedical Engineering, 2022, vol. 15, pp. 325-
Optimized back propagation deep neural network model has been
employed to predict future active cases. The results indicate that 340.
the proposed approach not only beats existing baseline techniques [10] H. S. Alghamdi, G. Amoudi, S. Elhag, K. Saeedi and J. Nasser,
in terms of classification accuracy, but also attains lower "Deep Learning Approaches for Detecting COVID-19 From Chest X-
prediction MAPE compared to state of art regression models. Ray Images: A Survey," in IEEE Access, 2021, vol. 9, pp. 20235-
Thus, the proposed methodology developed addresses the 20254.
fundamental challenges associated with pandemics and aids early [11] T Ozturk, M Talo, EA Yildirim, UB Baloglu, “Automated detection
detection and prevention in such cases. of COVID-19 cases using deep neural networks with X-ray images”,
Acknowledgements Computers and Biology and Medicine, Elsevier 2020, col.121,
103792.
We thank our colleagues from SAGE University, Indore who [12] J. C. Clement, V. Ponnusamy, K. C. Sriharipriya and R.
provided insight and expertise that greatly assisted the research. Nandakumar, "A Survey on Mathematical, Machine Learning and
We thank to Dr. Lalji Prashad HOI, IAC, Sage University for Deep Learning Models for COVID-19 Transmission and Diagnosis,"
assistance and motivation. in IEEE Reviews in Biomedical Engineering, 2022, vol. 15, pp. 325-
Author contributions 340
[13] M. M. Islam, F. Karray, R. Alhajj and J. Zeng, "A Review on Deep
Mr. Sandeep Kumar Maharaja: Data collection, Learning Techniques for the Diagnosis of Novel Coronavirus
Implementation of proposed model, compilation of results and (COVID-19)," in IEEE Access, vol. 9, pp. 30551-30572.
drafting manuscript.
[14] Y Zhang, Q Zhang, Y Zhao, Y Deng, H Zheng, “Urban spatial risk
prediction and optimization analysis of POI based on deep learning
Dr. Hemang Shrivastava: Conceptualization of Research
Problem and Methodology, Writing-Reviewing and Editing. from the perspective of an epidemic”, International Journal of
Applied Earth Observation and Geoinformation, , Elsevier, 2022, vol.
Conflicts of interest 112, 102942.
The authors declare no conflicts of interest. [15] L Xu, R Magar, AB Farimani, “Forecasting COVID-19 new cases
using deep learning methods”, Computers in biology and medicine,

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3149
 Elsevier 2022, vol.144, 105342 Engineering, Elsevier 2021, vol.41, no.3, pp. 867-879.
[16] AS Kwekha-Rashid, HN Abduljabbar, B Alhayani, “Coronavirus [32] E. -S. M. El-Kenawy et al., "Advanced Meta-Heuristics,
disease (COVID-19) cases analysis using machine-learning Convolutional Neural Networks, and Feature Selectors for Efficient
applications”, Applied Nanoscience, Springer, 2023, vol.31, pp. COVID-19 X-Ray Chest Image Classification," in IEEE Access,
2013–2025. 2021, vol. 9, pp. 36019-36037.
[17] A Mollalo, B Vahedi, S Bhattarai, LC Hopkins, “Predicting the [33] K. K. Singh and A. Singh, "Diagnosis of COVID-19 from chest X-
hotspots of age-adjusted mortality rates of lower respiratory infection ray images using wavelets-based depthwise convolution network," in
across the continental United States: Integration of GIS, spatial Big Data Mining and Analytics, 2021, vol. 4, no. 2, pp. 84-93.
statistics and machine learning algorithms”, International Journal of [34] B. King, S. Barve, A. Ford and R. Jha, "Unsupervised Clustering of
Medical Informatics, Elsevier, 2021, vol. 142, 104248. COVID-19 Chest X-Ray Images with a Self-Organizing Feature
[18] SD Khan, L Alarabi, S Basalamah -, “Toward Smart Lockdown: A Map," 2020 IEEE 63rd International Midwest Symposium on
Novel Approach for COVID-19 Hotspots Prediction Using a Deep Circuits and Systems (MWSCAS), 2020, pp. 395-398.
Hybrid Neural Network”, Journal of Computers, MDPI, 2020, vol.9, [35] ALA Dalal, MAA Al-qaness, Z Cai, EA Alawamy, “IDA: Improving
no.9, pp1-16. distribution analysis for reducing data complexity and dimensionality
[19] A Akbarimajd, N Hoertel, MA Hussain, “Learning-to-augment in hyperspectral images”, Pattern Recogniton, Elsevier 2023, vol.134,
incorporated noise-robust deep CNN for detection of COVID-19 in no. 109096.
noisy X-ray images”, Journal of Computational Science, Elsevier [36] R Mostafiz, MS Uddin, MM Reza, MM Rahman, “Covid-19
2022, vol.63, 101763. detection in chest X-ray through random forest classifier using a
[20] M Momeny, AA Neshat, MA Hussain, S Kia, “Learning-to-augment hybridization of deep CNN and DWT optimized features”, Journal of
strategy using noisy and denoised data: Improving generalizability of King Saud University - Computer and Information Sciences, Elsevier
deep CNN for the detection of COVID-19 in X-ray images”, 2022, vol.34, no.6, Part B, pp. 3226-3235.
Computers in Biology and Medicine, Elsevier, 2022, vol.136, [37] U Muhammad, MZ Hoque, M Oussalah, “SAM: Self-augmentation
104704. mechanism for COVID-19 detection using chest X-ray images”,
[21] MF Aslan, MF Unlersen, K Sabanci, A Durdu, “CNN-based transfer Lnowledge Based Systems, Elsevier 2022, vol.241, 108207.
learning–BiLSTM [38] I. Ahmad, M. Basheri, M. J. Iqbal and A. Rahim, "Performance
[22] NN Das, N Kumar, M Kaur, V Kumar, D Singh, “Automated deep Comparison of Support Vector Machine, Random Forest, and
transfer learning-based approach for detection of COVID-19 Extreme Learning Machine for Intrusion Detection," in IEEE Access
infection in chest X-rays”, IBRM, Elsevier, 2022, vol.43, no.2, 2018, vol. 6, pp. 33789-33795.
pp.114-119. [39] P. Padilla, M. Lopez, J. M. Gorriz, J. Ramirez, D. Salas-Gonzalez
[23] D. A. Zebari, A. M. Abdulazeez, D. Q. Zeebaree and M. S. Salih, "A and I. Alvarez, "NMF-SVM Based CAD Tool Applied to Functional
Fusion Scheme of Texture Features for COVID-19 Detection of CT Brain Images for the Diagnosis of Alzheimer's Disease," in IEEE
Scan Images," 2020 International Conference on Advanced Science Transactions on Medical Imaging, 2012 vol. 31, no. 2, pp. 207-216
and Engineering (ICOASE), 2020, pp. 1-6. [40] G. Rathee, S. Garg, G. Kaddoum, Y. Wu, D. N. K. Jayakody and A.
[24] O. El Gannour, S. Hamida, B. Cherradi, A. Raihani And H. Alamri, "ANN Assisted-IoT Enabled COVID-19 Patient
Moujahid, "Performance Evaluation of Transfer Learning Technique Monitoring," in IEEE Access, vol. 9, pp. 42483-42492.
for Automatic Detection of Patients with COVID-19 on X-Ray [41] C. Zhou, J. Song, S. Zhou, Z. Zhang and J. Xing, "COVID-19
Images," 2020 IEEE 2nd International Conference on Electronics, Detection Based on Image Regrouping and Resnet-SVM Using Chest
Control, Optimization and Computer Science (ICECOCS), 2020, pp. X-Ray Images," in IEEE Access, 2021, vol. 9, pp. 81902-8191.
1-6. [42] S. K. Zhou et al., "A Review of Deep Learning in Medical Imaging:
[25] SH Kassania, PH Kassanib, MJ Wesolowskic, “Automatic detection Imaging Traits, Technology Trends, Case Studies With Progress
of coronavirus disease (COVID-19) in X-ray and CT images: a Highlights, and Future Promises," in Proceedings of the IEEE, 2021,
machine learning based approach”, Biometrics and Biomedical vol. 109, no. 5, pp. 820-838.
Engineering, Elsevier, 2021, vol.41, no.3, pp. 867-879. [43] S. Tiwari, P. Chanak and S. K. Singh, "A Review of the Machine
[26] V Bolon-Canedo, B Remeseiro, “Feature selection in image analysis: Learning Algorithms for Covid-19 Case Analysis," in IEEE
a survey”, Artificial Intelligence Review, Springer 2020, vol. 53, pp. Transactions on Artificial Intelligence.
2905–2931. [44] T. Sercu, C. Puhrsch, B. Kingsbury and Y. LeCun, "Very deep
[27] R. Miikkulainen, J Liang, E Meyerson, A Rawal, “Evolving deep multilingual convolutional neural networks for LVCSR," 2016 IEEE
neural networks”, Artifical Intelligence in the Age of Neural International Conference on Acoustics, Speech and Signal Processing
Networks and Brain Computing, Academic Press, 2019, pp.293-312. (ICASSP), 2016, pp. 4955-4959.
[28] A Zargari Khuzani, M Heidari, SA Shariati, “COVID-Classifier: An [45] MA Kızrak, Z Müftüoğlu, T Yıldırım, “Limitations and challenges
automated machine learning model to assist in the diagnosis of on the diagnosis of COVID-19 using radiology images and deep
COVID-19 infection in chest x-ray images”, Scientific Repots, learning”, Data Science for COVID-19, Computational Perspectives ,
Nature, 2021, vol.11, no. 9887. Science Direct 2021, pp.91-115.
[29] M Shafiq‐ul‐Hassan, GG Zhang, K Latifi, “Intrinsic dependencies of [46] K. Sanjar, A. Rehman, A. Paul and K. JeongHong, "Weight Dropout
CT radiomic features on voxel size and number of gray levels”, for Preventing Neural Networks from Overfitting," 2020 8th
Scientific Reports, Nature, 2018, vol.8, no. 10545. International Conference on Orange Technology (ICOT), Daegu,
[30] M Toğaçar, B Ergen, Z Cömert, “COVID-19 detection using deep Korea (South), 2020, pp. 1-4.
learning models to exploit Social Mimic Optimization and structured [47] L Borawar, R Kaur, “ResNet: Solving Vanishing Gradient in Deep
chest X-ray images using fuzzy color and stacking approaches”, Networks”, Proceedings of International Conference on Recent
Computers in Biology and Medicine, Elsevier 2020, vol.121, 103805. Trends in Computing, Springer 2023, pp 235–247.
[31] SH Kassania, PH Kassanib, MJ Wesolowskic, “Automatic detection [48] M. A. Awal, M. Masud, M. S. Hossain, A. A. -M. Bulbul, S. M. H.
of coronavirus disease (COVID-19) in X-ray and CT images: a Mahmud and A. K. Bairagi, "A Novel Bayesian Optimization-Based
machine learning based approach”, Biometrics and Biomedical Machine Learning Framework for COVID-19 Detection From

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3150
 Inpatient Facility Data," in IEEE Access, vol. 9, pp. 10263-10281,
2021.
[49] M Canayaz, S Şehribanoğlu, R Özdağ, “COVID-19 diagnosis on CT
images with Bayes optimization-based deep neural networks and
machine learning algorithms”, Neural Computing, Springer, 2022,
vol.34, pp.5349–5365.
[50] F. Rustam et al., "COVID-19 Future Forecasting Using Supervised
Machine Learning Models," in IEEE Access, vol. 8, pp. 101489-
101499, 2020.
[51] W Chen, P Wang, J Cheng, “Towards mixed-precision quantization
of neural networks via constrained optimization”, Proceedings of the
IEEE/CVF International Conference on Computer Vision (ICCV),
2021, pp. 5350-5359.
[52] S Ballı, “Data analysis of Covid-19 pandemic and short-term
cumulative case forecasting using machine learning time series
methods”, Chaos, Solitons & Fractals, Elsevier, 2021, vol.142,
110512

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 3148–3151 | 3151