See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/341980921

COVID-19 Prediction and Detection Using Deep Learning

Article · May 2020

CITATIONS READS

4 12,163

6 authors, including:

Moutaz Alazab Abdelwadood Mesleh

Al-Balqa Applied University Al-Balqa Applied University
17 PUBLICATIONS   186 CITATIONS    38 PUBLICATIONS   578 CITATIONS   

SEE PROFILE SEE PROFILE

Vansh Jatana Salah Alhyari

SRM Institute of Science and Technology JEPCO
14 PUBLICATIONS   10 CITATIONS    18 PUBLICATIONS   71 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

computer applications View project

Analysis of malicious and benign android applications View project

All content following this page was uploaded by Moutaz Alazab on 12 June 2020.

The user has requested enhancement of the downloaded file.

International Journal of Computer Information Systems and Industrial Management Applications.
ISSN 2150-7988 Volume 12 (2020) pp. 168-181
© MIR Labs, www.mirlabs.net/ijcisim/index.html

Received: 27 April 2020; Accepted: 28 May, 2020; Published: 29 May, 2020

COVID-19 Prediction and Detection Using Deep

Learning
Moutaz Alazab1, Albara Awajan1, Abdelwadood Mesleh1, Ajith Abraham2, Vansh Jatana3, Salah Alhyari4,
1 Faculty of Artificial Intelligence, Al-Balqa Applied University, Al-Salt, Jordan,
Amman, Jordan
(m.alazab, a.awajan, wadood)@bau.edu.jo
2 Machine Intelligence Research (MIR) Labs
Auburn, WA, US
ajith.abraham@ieee.org
3 School of IT, SRM University
Kattankulathur, India,
vs7182@srmist.edu.in
4 IT Department, JEPCO
Amman, Jordan
salah.alhyari@jepco.com.jo

Abstract: Currently, the detection of coronavirus disease 2019 vaccine for combating it. COVID-19 propagation is faster
(COVID-19) is one of the main challenges in the world, given the when people are in close proximity. Thus, travel restrictions
rapid spread of the disease. Recent statistics indicate that the control the spread of the disease, and frequent hand washing
number of people diagnosed with COVID-19 is increasing is always recommended to prevent potential viral infections.
exponentially, with more than 1.6 million confirmed cases; the Meanwhile, fever and cough are the most common infection
disease is spreading to many countries across the world. In this
study, we analyse the incidence of COVID-19 distribution across
symptoms. Other symptoms may occur, including chest
the world. We present an artificial-intelligence technique based discomfort, sputum development, and a sore throat. COVID-
on a deep convolutional neural network (CNN) to detect COVID- 19 may progress to viral pneumonia which has a 5.8%
19 patients using real-world datasets. Our system examines chest mortality risk. The death rate of COVID-19 is equivalent to 5%
X-ray images to identify such patients. Our findings indicate that of the death rate of the 1918 Spanish flu pandemic.
such an analysis is valuable in COVID-19 diagnosis as X-rays are The total number of people infected with COVID-19
conveniently available quickly and at low costs. Empirical worldwide is 5,790,103 as of May 27, 2020 whereas the
findings obtained from 1000 X-ray images of real patients numbers of reported deaths and recoveries are 357,432 and
confirmed that our proposed system is useful in detecting 2,497,618 respectively. Most of the cases were recorded in the
COVID-19 and achieves an F-measure range of 95–99%.
Additionally, three forecasting methods—the prophet algorithm
USA, Spain, Italy, France, Germany, mainland China, UK,
(PA), autoregressive integrated moving average (ARIMA) model, and Iran [2]. Saudi Arabia, with 78,541 cases, has the highest
and long short-term memory neural network (LSTM)—were number of reported cases among all the Arab countries.
adopted to predict the numbers of COVID-19 confirmations, Meanwhile, the number of reported cases in Jordan is 720,
recoveries, and deaths over the next 7 days. The prediction whereas the numbers of deaths and recoveries are 9 and 586
results exhibit promising performance and offer an average respectively. The number of reported cases in Australia is
accuracy of 94.80% and 88.43% in Australia and Jordan, 7150, whereas the numbers of deaths and recoveries are 103
respectively. Our proposed system can significantly help identify and 6579, respectively. Since February 2020, information
the most infected cities, and it has revealed that coastal areas are technology services, such as mobile apps, have been used to
heavily impacted by the COVID-19 spread as the number of
cases is significantly higher in those areas than in non-coastal
curb the potential risk of infection in mainland China. The
areas. mobile apps suggest users to self-quarantine and alert the
Keywords: Artificial Intelligence, X-ray, Convolutional Neural concerned health authorities when someone infected by the
Network, Machine Learning, COVID-19. virus. They also monitor infected people, and the last persons
that they had contact with [3].
I. Introduction Since it was first reported, the disease has spread
exponentially across the world and has become an
The coronavirus disease (COVID-19) is a global pandemic international concern. A research conducted by Jiang et al. [4]
that was discovered by a Chinese physician in Wuhan, the revealed that the death rate of COVID-19 is 4.5% across the
capital city of Hubei province in mainland China, in world. The death rate of patients in the age range of 70–79
December 2019 [1]. Currently, there is no approved human years is 8.0%, whereas that of patients above 80 years is
MIR Labs, USA
COVID-19 Prediction and Detection Using Deep Learning 169

14.8%. The authors also confirmed that patients above the age  Empirical findings obtained from 1000 chest X-ray images
of 50 years with chronic illnesses are at the highest risk and of patients confirmed that our proposed system can detect
should therefore take special precautions. One of the main COVID-19 patients with an accuracy of 95–99%.
threats of COVID-19 is its rapid propagation, with an  We provide an intelligent prediction system for predicting
estimated 1.5–3.5 people getting infected by the disease upon the number of patients confirmed to have contracted the
contact with an infected person [5]. This implies that if 10 disease, recovered from the disease, and died from the disease
people are COVID-19 positive, they are more likely to infect over the next 7 days using three forecasting methods. Our
15–35 other people. Therefore, COVID-19 can infect a very proposed system has been trained and tested on datasets
large number of people in a few days unless intervention generated from real-world cases and has predicted the
measures are implemented.
numbers of COVID-19 confirmations, recoveries, and deaths
The standard diagnostic technique is the reverse
in Australia and Jordan with an average accuracy of 94.80%
transcription-polymerase chain reaction (RT-PCR) method
and 88.43%, respectively.
[1], a laboratory procedure that interacts with other
ribonucleic (RNA) and deoxyribonucleic acids (DNA) to  We highlight the most affected areas and show that coastal
determine the volume of specific ribonucleic acids using areas are heavily impacted by COVID-19 infection and spread
fluorescence. RT-PCR tests are performed on clinical research as the number of cases in those areas is significantly higher
samples of nasal secretions. The samples are collected by than that in other non-coastal areas.
inserting a swab into the nostril and gently moving it into the The rest of this paper is organised as follows. Section 2
nasopharynx to collect secretions. Although RT-PCR can presents the related works on recent COVID-19 detection and
identify the severe acute respiratory syndrome coronavirus 2 prediction methods for chest X-ray images. Section 3 presents
(SARS-CoV-2) strain that causes COVID-19, in some cases, the detailed system design, dataset description, and
it produced negative test results even though the patients performance-evaluation metrics. Sections 4 and 5 present the
showed progression on follow-up chest computed tomography results and discussions, respectively. Section 6 concludes the
(CT) scans [6]. In fact, several studies [6-9] have recommend paper and provides an outlook to future research.
the use of CT scans and X-rays rather than RT-PCR owing to
its limited availability in some countries. II. Related Works
The detection of COVID-19 symptoms in the lower parts of
The analysis and detection of COVID-19 have been
the lungs has a higher accuracy when using CT scans or X-
extensively investigated in the last few months. The first part
rays than that when using RT-PCR [7]. In certain cases, CT
of this section addresses issues related to COVID-19 detection
scans and X-ray tests can be substituted with RT-PCR tests.
based on deep-learning approaches using CT scans and chest
However, they cannot exclusively address the problem owing
X-ray images. The second part reviews the related literatures
to the relatively limited number of radiologists, compared to
to assess future estimates of the number of COVID-19
new residents, and the high volume of re-examinations of
confirmations, recoveries, and deaths.
infected people who wish to know the progression of their
COVID-19 has now become a global pandemic owing to its
illness. To overcome the challenges of CT scans and X-rays
rapid spread. It is very challenging to detect exposed persons
and to assist radiologists, we need to improve the speed of the
because they do not show disease symptoms immediately.
procedure. This can be achieved by designing advanced
Thus, it is necessary to find a method of estimating the number
diagnostic systems that utilise artificial intelligence (AI) tools.
of potentially infected persons on a regular basis to adopt the
The aim is to reduce the time and effort required to perform
appropriate measures. AI can be used to examine a person for
CT scans and X-rays of COVID-19-positive patients and
evaluate the rate of disease development [7-9]. COVID-19 as an alternative to traditional time-consuming
and expensive methods. Although there are several studies on
Radiological imaging is considered an important screening
COVID-19, this study focused on the use of AI in forecasting
method for COVID-19 diagnosis [10]. Ai et al. [6]
COVID-19 cases and diagnosing patients for COVID-19
demonstrated the consistency of the radiological history of
infection through chest X-ray images.
COVID-19-related pneumonia with the clinical nature of the
Several research areas have implemented AI (e.g. disease
disease. When examined by CT scans, almost all COVID-19
diagnoses in healthcare) [11-13]. One of the main advantages
patients have exhibited similar features including ground-
of AI is that it can be implemented in a trained model to
glass opacities in the early stages and pulmonary
consolidation in the latter stages. In fact, the morphology and classify unseen images. In this study, AI was implemented to
detect whether a patient is positive for COVID-19 using their
peripheral lung distribution can be rounded [6]. AI can be used
chest X-ray image.
to initially evaluate a COVID-19 patient as an alternative
AI can also be used for forecasting (e.g., how the population
solution to traditional approaches that are time-consuming and
will increase over the next 5 years) through existing evidence.
labour-intensive. In this paper, we advocate the use of AI to
forecast COVID-19 cases and diagnose COVID-19 patients Thus, predicting possibilities in the immediate future can help
authorities to adopt the necessary measures [14]. Wynants et
via chest X-ray images.
al. [15] focused on two main concepts. The first concept
A. Contributions of This Study involved studies related to the diagnosis of COVID-19, and
The following are the core contributions of this study: the second involved studies related to the prediction of the
 We propose an automated intelligent system for number of people who will be infected in the coming days.
distinguishing COVID-19 patients from non-patients on the The study analysis maintained that most of the existing
basis of chest X-ray images. Our system instantly reads the models are poor and biased. The authors suggested that
structure of a chest X-ray image, leverages hidden patterns to research-based COVID-19 data should be publicly available
identify COVID-19 patients, and reduces the need for manual to encourage the adoption of more specifically designed
detection and prediction models.
pre-processing steps.
170 Alazab et al.

512×512 pixels. Their classification system achieved a

COVID-19 Diagnosis Using Deep Learning sensitivity of 95.76%, a specificity of 99.7% and an accuracy
The use of machine learning (ML) has been rapidly increasing of 97.48%.
in various fields including malware detection [16-19], mobile
1) CT Scan Diagnosis Using Deep Learning
malware detection [20-24], medicine [25-27] and information
retrieval [27-31]. In 2012, a modern ML system called deep The CT scan was developed by Godfrey Hounsfield and Allan
learning was introduced, which is based on a convolutional Cormack in 1972. It utilises an advanced X-ray technology to
neural network (CNN). It won the ImageNet classification carefully diagnose delicate internal organs [34]. CT scanning
competition, the world’s best-known computer-vision is quick, painless, non-invasive, and precise and can produce
competition [32]. Deep-learning algorithms enable three-dimensional (3D) images [34]. CT scans of internal
computational models composed of multiple processing layers organs, muscles, soft tissues, and blood vessels offer greater
to learn data representation through several abstraction layers. clarity than standard X-rays, especially for soft tissues and
They train a computer model to perform classification tasks blood vessels. The main disadvantage of the CT scan is that it
directly from pictures, texts, or sounds. According to LeCun is expensive, compared to X-rays [34].
et al. [33], deep-learning models feature high accuracies and The sensitivity and specificity of RT-PCR for COVID-19
can improve human output in certain instances. detection have been criticised in several studies [4, 38, 39].
X-Ray Diagnosis Using Deep Learning Although RT-PCR is the standard method for this purpose, it
X-ray machines use light or radio waves as radiation to generates a relatively large number of false negatives owing
examine the affected parts of the body because of cancers, to several reasons, including methodological drawbacks,
lung diseases, bone dislocations, and injuries. Meanwhile, CT disease stages, and methods of obtaining the specimens, which
scans are used as sophisticated X-ray machines to examine the delay disease diagnosis and control. Therefore, RT-PCR tests
soft structures of active body parts for better views of the are not sufficient for assessing the disease status. Recent
actual soft tissues and organs [34]. The advantages of using results have revealed that nucleic acid testing is not reliable
X-rays over CT scans are that X-rays are quicker, safer, and can only achieve an accuracy of 30–50% [38].
simpler, and less harmful than CT scans [7, 34]. Jiang et al. [4] compared RT-PCR to CT scans and
Narin et al. [7] proposed a CNN-based model to detect examined 51 patients (29 men and 22 women) with a history
COVID-19 patients using 100 chest X-ray images, half of of travel to or residency in endemic areas and with severe
which belong to COVID-19 patients and the other half belong respiratory and fever symptoms due to unknown causes. The
to healthy people. They evaluated three CNN models— authors obtained a sensitivity of 98% in a non-contrast chest
ResNet-50, Inception-v3, and Inception-ResNet-v2—using CT scan for the detection of COVID-19, compared to the
five-fold cross-validation and reported that ResNet-50 had the initial RT-PCR sensitivity of 71%. Owing to the shortage of
best detection accuracy (98%). RT-PCR kits and the growing number of COVID-19 cases, it
In a similar study conducted by Sethy and Behera [35], the is important to introduce an automated detection system as an
authors extracted features from chest X-ray images using a alternative diagnostic method to prevent the spread of
deep-learning algorithm and classified the images as either COVID-19 [7].
infected or healthy using a support vector machine (SVM). Meanwhile, Gozes et al. [40] employed a deep-learning
The authors employed 11 deep-learning models: AlexNet, approach to automatically identify COVID-19 patients and
VGG16, VGG19, GoogLeNet, ResNet-18, ResNet-50, examine the disease burden quantification on CT scans using
ResNet-101, Inception-v3, Inception-ResNet-v2, DenseNet- a dataset of CT scans from 157 foreign patients from China
201, and XceptionNet. They collected two datasets—the first and the USA. Their proposed system analyses the CT scan at
containing chest X-ray images of 25 infected patients and 25 two distinct levels: subsystems A and B. Subsystem A
non-infected patients and the other containing chest X-ray performs a 3D analysis, and subsystem B performs a 2D
images of 133 infected patients (e.g. MERS, SARS, and analysis of each segment of the scan to identify and locate
ARDS patients) and 133 non-infected patients. They broader diffuse opacities, including ground-glass infiltrates
performed separate feature extractions on each dataset using (which have been clinically identified as representative of
various models and achieved a 95.38% accuracy with ResNet- COVID-19). To evaluate their system, the authors applied
50 and SVM. Resnet-50-2 to subsystem B and obtained an area under the
Furthermore, Hemdan et al. [36] proposed a framework, curve of 99.6%. The sensitivity and specificity were 98.2%
called COVIDX-Net, that can assist radiologists in diagnosing and 92.2%, respectively.
COVID-19 patients using X-ray. They evaluated their Moreover, Wang et al. [38] developed a deep-learning
framework using a dataset of 50 X-ray images divided into approach for extracting information from CT scans. Their
two classes: 25 COVID-19-positive images and 25 COVID- study included a collection of 453 CT scans from 99 patients.
19-negative images. The images used were resized to They extracted 195 regions of interest (ROIs) of sizes ranging
224×224 pixels. The COVIDX-Net framework employs seven from 395×223 to 636×533 pixels from the CT scans of 44
deep learning models: MobileNet, ResNet-v2, Inception- COVID-19-positive pneumonia patients and 258 ROIs from
ResNet-v2, Xception, Inception-v3, DenseNet, and modified those of 50 COVID-19-negative patients. They applied a
VGG19. Their evaluation results indicate that the VGG19 and modified network inception model and obtained an accuracy
DenseNet models delivered comparable performances with an of 82.9% for the internal validation with a specificity of 80.5%
F-score of 91% for COVID-19 cases. and a sensitivity of 84%. The external testing dataset exhibited
In addition, Hassanien et al. [37] proposed a classification a total accuracy of 73.1% with a specificity of 67% and a
system that uses multi-level thresholding and an SVM to sensitivity of 74%.
detect COVID-19 in lung X-ray images. Their system was Fu et al. [41] proposed a classification system based on
tested on 40 contrast-enhanced lung X-ray images (15 healthy ResNet-50 to detect COVID-19 and some other infectious
and 25 COVID-19-infected regions) with a resolution of lung diseases (bacterial pneumonia and pulmonary
COVID-19 Prediction and Detection Using Deep Learning 171

tuberculosis). The authors collected a dataset of 60,427 CT would be urgently required to stop the disease from spreading.
scans from 918 patients; 14,944 of these CT scans were from Although the prediction of COVID-19 cases for the USA was
150 COVID-19 patients and 15,133 from 154 non-COVID-19 1 million between 8 April and 30 April 2020, it reached
viral pneumonia patients. They performed several tests for 677,570 on 17 April 2020. Furthermore, Italy had 168,941
several lung diseases. The achieved accuracy, sensitivity, and cases, although it was predicted to have 300,000 cases [2].
specificity were 98.8%, 98.2%, and 98.9%, respectively. Huang et al. [45] applied a CNN to a limited dataset, which
Xu et al. [42] reported that real-time RT-PCR has a low was not specifically defined in their study, to evaluate and
positive rate at the early stage of COVID-19. They developed estimate the number of reported cases in China. The authors
an early screening model that uses deep-learning techniques used the mean absolute and root mean square errors to
for distinguishing COVID-19 pneumonia from influenza (a compare their model with other deep-learning models,
viral pneumonia) and stable cases using pulmonary CT images. including multilayer perceptron, long short-term memory
A dataset of 618 CT samples was obtained for the analysis, (LSTM), and gated recurrent units. The authors concluded that
and the images were classified as COVID-19, influenza (a the obtained results promise a high predictive efficiency.
viral pneumonia), and other cases using ResNet-18 and Pandey et al. [46] utilised two statistical algorithms—the
ResNet-based methods. The authors employed a noisy or susceptible-exposed-infectious-recovered (SEIR) and
Bayesian function to differentiate the infected images and regression models—to evaluate and forecast the distribution
obtained a detection accuracy of 86.7%. of COVID-19 in India. They used a dataset retrieved from the
John Hopkins University repository. The prediction results
2) COVID-19 Infection Prediction Using Machine
from the SEIR and regression models showed that the number
Learning Techniques
of confirmed COVID-19 cases would reach 5300 and 6153
ML is the science of training machines using mathematical cases, respectively, by 13 April 2020. However, the confirmed
models to learn and analyse data. Once ML is implemented in cases in India on that date were 10,453 and 6153 for the SEIR
a system, the data are analysed, and interesting patterns are and regression models, respectively [2].
detected. The validation data are then categorised according
to the patterns learned during the learning process. As III. System Design
COVID-19 infection has rapidly spread worldwide and
international action is required, it is important to develop a Our proposed deep learning-based COVID-19 detection
strategy to estimate the number of potentially infected people comprises several phases, as illustrated in Figure 1. The
on a regular basis to adopt the appropriate measures. Currently, phases are summarised in the following five steps:
decision-makers rely on certain decision-making statistics Step 1: Collect the chest X-ray images for the dataset from
such as imposing lockdowns on infected cities or countries. COVID-19 patients and healthy persons.
Therefore, ML can be used to predict the behaviours of new Step 2: Generate 1000 chest X-ray images using data
cases to stop the disease from spreading. augmentation.
Li et al. [43] developed a prediction model using ML Step 3: Represent the images in a feature space and apply deep
algorithms to combat COVID-19 in mainland China and in learning.
other infected countries in the world. The authors developed a Step 4: Split the dataset into two sets: a training set and a
model to estimate the number of reported cases and deaths in validation set.
mainland China and in the world. The data used to build the Step 5: Evaluate the performance of the detector on the
models were collected between 20 January 2020 and 1 March validation dataset.
2020. They estimated that the peak of the COVID-19 outbreak A. Dataset
in mainland China occurred on 22 February 2020 and on 10
Two types of datasets were used in the evaluation, the original
April 2020 worldwide. The authors also stated that COVID-
19 would be controlled at the beginning of April 2020 in dataset (without augmentation) and the augmented dataset,
mainland China and in mid-June 2020 across the world. They which are summarised in Tables 1 and 2, respectively. The
concluded that the estimated number of COVID-19 cases dataset contained the following: a) a healthy dataset
would be approximately 89,000 in China and 403,000 containing chest X-ray images of healthy persons and b) a
worldwide during the outbreak. As of 17 April 2020, the COVID-19 dataset containing chest X-ray images of COVID-
estimated number of deaths was 4000 in mainland China and 19 patients. The original dataset was obtained from the Kaggle
18,300 worldwide. It is clear that their forecast was similar to database, and its total number of images is 128, as presented
the actual situation in China as the total numbers of infected in Table 1 [47].
cases and deaths had exceeded 82,367 and 3342, respectively.
However, the number of confirmed cases worldwide exceeded
their estimations as the numbers of infected cases and deaths
exceeded 2 million and reached 145,416 as of 17 April 2020,
respectively [2].
Kumar et al. [44] predicted the COVID-19 spread in the 15
most-infected countries in the world using the autoregressive
integrated moving average (ARIMA) model. The outcome of
their prediction indicates that circumstances would worsen in
Iran and Europe, especially in Italy, Spain, and France.
Moreover, their prediction indicated that the number of cases
in South Korea and mainland China would become more
stable. The study also indicated that COVID-19 would spread
exponentially in the USA and that strict official measures Figure 1. Architecture of the proposed system
172 Alazab et al.

Owing to the limited availability of chest X-ray images, we true-negative (TN), false-positive (FP), and false-negative
generated our dataset using data augmentation [48]. Data (FN) scores:
augmentation is an AI method for increasing the size and the - TP is the proportion of positive COVID-19 chest X-
diversity of labelled training sets by generating different ray images that were correctly labelled as positive.
iterations of the samples in a dataset. Data augmentation - FP is the proportion of negative (healthy) COVID-19
methods are commonly used in ML to address class imbalance chest X-ray images that were mislabelled as positive.
problems, reduce overfitting in deep learning, and improve - TN is the proportion of negative (healthy) chest X-
convergence, which ultimately contributes to better results. ray images that were correctly labelled as healthy.
The total number of images in the dataset became 1000 after - FN is the proportion of positive COVID-19 chest X-
applying augmentation, as presented in Table 2 [47]. ray images that were mislabelled as negative
(healthy).
Table 1 Original dataset (without augmentation)  Accuracy: This metric measures the percentage of
X-ray images Number correctly identified cases relative to the entire dataset. The
Healthy 28 ML algorithm performs better if the accuracy is higher.
COVID-19 70 Accuracy is a significant measure for a test dataset that
Total 128 includes a balanced class. It is computed as follows:
𝑨𝒄𝒄𝒖𝒓𝒂𝒄𝒚 = (𝑇𝑃
Table 2 Augmented dataset + 𝑇𝑁)/ (𝑇𝑃
(1)
X-ray images Number + 𝑇𝑁 + 𝐹𝑃
Healthy 500 + 𝐹𝑁)
COVID-19 500  Precision: This metric is a measure of exactness, which is
Total 1000 calculated as the percentage of positive predictions of
B. Environment COVID-19 that were true positives divided by the number
of predicted positives. It is computed as follows:
A computer with Microsoft Windows 10 was used for the
experiment. It has the following specifications: Intel Core i7- 𝑷𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏 = 𝑇𝑃/(𝑇𝑃 + 𝐹𝑃) (2)
8565U 1.80-GHz processor, 16 GB of DDR4 RAM, and 1 TB  Recall: This metric is a measure of completeness, which is
of hard disk. We installed the virtual machine tool VMware calculated as the percentage of positives that were
Workstation Pro version 14.1.8 build-14921873 on it. Then, correctly identified as true positives divided by the number
we installed Ubuntu 18.04.4 (64 bit) on the virtual machine of actual positives. It is computed as follows:
and the following libraries and software:
 ARIMA: 𝑹𝒆𝒄𝒂𝒍𝒍 = 𝑇𝑃/(𝑇𝑃 + 𝐹𝑁) (3)
https://www.statsmodels.org/stable/generated/statsmodels.ts  F-measure: This is a combination of precision and recall
a.arima_model.ARIMA.html that provides a significant measure for a test dataset that
 Fbprophet: https://pypi.org/project/fbprophet/ includes an imbalanced class. It is computed as follows:
 ImageDataGenerator: https://keras.io/preprocessing/image/ 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑥 𝑅𝑒𝑐𝑎𝑙𝑙
𝑭 − 𝑴𝒆𝒂𝒔𝒖𝒓𝒆 = 𝟐 ( )
 Keras: https://keras.io/ (𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙) (4)
LSTM:  Root Mean Square Error (RMSE): This metric
https://www.tensorflow.org/api_docs/python/tf/keras/layers/LS measures the differences between the actual (𝑥𝑖 ) and the
TM
predicted ( 𝑥̂𝑖 ) numbers of COVID-19 confirmations,
 Matplotlib: https://matplotlib.org/
 NumPy: https://numpy.org/ recoveries, and deaths (𝑁). The main advantage of RMSE
 Pandas: https://pandas.pydata.org/ is that it penalises large prediction errors. RMSE was used
to compare the prediction errors of the three prediction
 Python: https://www.python.org/
algorithms. It is defined as follows:
 Scikit: https://scikit-learn.org/ 𝑁
 SciPy: https://www.scipy.org/ 1
𝑹𝑴𝑺𝑬 = √ ∑(𝑥𝑖 − 𝑥̂𝑖 )2 (5)
 TensorFlow: https://www.tensorflow.org/ 𝑁
𝑖=1
All the results and predictions made in this study have been  Correlation coefficient: This metric is often used to
uploaded to the Kaggle database [49, 50]. We believe that by evaluate the performance of a prediction algorithm. It is
making the system and solution publicly available, we draw defined as follows:
attention to the most affected areas, thereby preventing the
𝑁
spread of the COVID-19 outbreak and fostering the use of 1
deep-learning techniques in COVID-19 research. 𝑪𝑪 = (1 − ∑|𝑥𝑖 − 𝑥̂𝑖 |) ∗ 100% (6)
𝑁
𝑖=1
C. Evaluation Metrics
To assess the reliability of the proposed deep learning-based
COVID-19 detector, we adopted the same metrics as those
IV. Experimental Results
used by Alazab et al. [51-54] and considered the following Firstly, we examined the most infected areas across the world.
standard metrics: precision, recall, and F-measure. These In Section 4.1, we show that coastal areas are heavily affected
metrics are calculated on the basis of the true-positive (TP), by the COVID-19 outbreak as the number of cases in those
COVID-19 Prediction and Detection Using Deep Learning 173

areas is significantly higher than those in other non-coastal

areas (Figures. 3–6).
Secondly, we predicted the number of COVID-19
confirmations, recoveries, and deaths in Jordan and Australia
over the next 7 days using three well-known time-series
forecasting algorithms: PA (prophet algorithm), ARIMA, and
LSTM. In Section 4.2, we show how the application of these
algorithms allowed us to estimate the forecasting outcomes for
certain countries with a detection rate of 99% (Tables 3 and 4)
and (Figures. 7 -13).
Thirdly, we examined whether chest X-ray images can be
used to develop sophisticated classification models for
COVID-19 prediction. In Section 4.3, we present the
application of a deep-learning algorithm on two datasets.
Empirical findings indicated that our proposed system is Figure 3. Areas with the highest number of confirmed cases
reliable in detecting COVID-19 and has an F-measure range in Australia
of 95–99%, as revealed by Figures. 15 -16.
 In South Korea, the disease peaked on 20 January 2020, and
A. Coronavirus Statistics the number of confirmed cases exceeded 11,000. Figure 4
SARS-CoV-2 is a new family of viruses that has never been shows that the regions that were most severely affected by
encountered before. The virus was first discovered in COVID-19 were coastal or near-coastal areas.
pangolins before its spread to humans [55]. The typical
symptoms of COVID-19 include fever, dry cough, fatigue,
sputum production, shortness of breath, sore throat, headache,
chills, nausea or vomiting, nasal congestion, diarrhoea,
haemoptysis, and conjunctival discomfort, although some
patients also suffer from general tiredness, runny nose, and
loss of taste and/or scent. Figure 2 shows a bar graph of the
common COVID-19 symptoms sorted by their percentage of
occurrences [56].

Figure 4. Areas with the highest number of confirmed cases

in South Korea

 In India, the number of reported confirmed cases was more

than 158,000, and the most-affected places were Maharashtra;
Tamil Nadu, Delhi, Gujrat, Rajasthan, Madhya Pradesh, Uttar
Pradesh etc. as shown in Figure 5 [58].

Figure 2. Common COVID-19 symptoms

In this section, the connection between coastal and non-

coastal areas is further explored as a large proportion of
infected cases were recorded in coastal areas. On the basis of
the collected statistics, COVID-19 has a rapid spread in
coastal areas. The following examples support this conclusion:
 In Australia, there were more than 7000 confirmed cases.
The highest number of confirmed cases were in New South
Wales, Victoria, Queensland, Western Australia, South
Australia, Tasmania, and Australian Capital Territory and
Northern Territory [57]. Figure 3 reveals the most affected
areas in Australia.
Figure 5. Areas with the highest number of confirmed cases
in India
174 Alazab et al.

 In the USA, there were more than 1,745,843 confirmed observations and the residual error values by using the moving
cases on 27 May 2020. The first case was found in Oregon, average for the lagged observations. ARIMA uses the order
which is located in the Pacific Coast. Coastal states, including factors p, d, and q, where p is the order of the autoregressive
Washington, Oregon, California, Arizona, and Texas, model, d is the order of the differencing, and q is the order of
reported high numbers of confirmed cases. Furthermore, the moving average. The algorithm is computed as follows:
states including Wisconsin and Illinois with long lake ̂ 𝑝
𝑦 = 𝑐 + 𝜖𝑡 + ∑ ∅𝑖 𝑦𝑡−𝑖
coastlines also reported confirmed cases at the initial stage of 𝑖=1
(8)
the COVID-19 spread. Other eastern coastal states including ̂ 𝑞
+ 𝑦 = 𝑐 + 𝜖𝑡 + ∑ 𝜃𝑖 𝜖𝑡−1
New York, Maine, New Hampshire, Massachusetts, Rhode 𝑖=1
Island, Connecticut, New Jersey, Delaware, Maryland, where 𝜖𝑡 is an independent and homogeneously distributed
Virginia, North Carolina, South Carolina, Georgia, Florida, error term, 𝑐 is a constant term, 𝑦 is an actual value at time 𝑡,
and Indiana also reported high numbers of confirmed cases, as and ϕ and θ are the tuning parameters of the autoregressive
well as other coastal areas such as Colorado and Nebraska. and moving-average models, respectively.
Thus, most of the states that reported the highest numbers of LSTM is a form of a recurrent neural network (RNN) that
cases are located in the coastal regions. Figure 6 highlights the memorises earlier patterns in data sequences. It was originally
coastal regions with the highest number of cases in maroon. proposed by Hochreiter and Schmidhuber [62]. It replaces the
hidden layer neurons of the RNN with a series of memory cells.
The key is the state of the memory cell that filters data using
a gate structure that updates the state of the memory cell. It
includes the input, forgotten, and output gates for its gate
structure. Each cell has three sigmoid layers and one tanh layer,
as shown in Figure 7 [61].

Figure 6. Areas with the highest number of confirmed cases

in USA
B. Prediction
In our system, we predicted the number of COVID-19 Figure 7. Structure of a memory cell in an LSTM
confirmations, recoveries, and deaths in Jordan and Australia
using the following three well-known forecasting algorithms: (i) The forgotten gate determines the cell state data that must
i) PA [59], ii) ARIMA [60], and iii) LSTM [61]. These be discarded. Each memory cell of the LSTM combines the
algorithms were trained to make predictions for the next 7 previous output ℎ𝑡−1 at time 𝑡 − 1 and the external current
days from datasets that were collected from a statistics website data 𝑥𝑡 in a vector [ℎ𝑡−1 , 𝑥𝑡 ] through a special sigmoid
[2]. function 𝜎, shown as follows:
Originally proposed by Taylor and Letham, PA is one of 𝑓𝑡 = 𝜎(𝑊𝑓 . [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑓 ) (9)
the well-known algorithms for solving multi-seasonal time- where 𝑊𝑓 and 𝑏𝑓 are the weight and the bias, respectively, of
series forecasting problems on Facebook [59]. It additively the forgotten gate. On the basis of its inputs, if the forgotten
decomposes a time series-based forecasting problem into gate outputs ‘1’, it indicates ‘reserve’, and if it outputs ‘0’, it
trends, seasonal, remainder, and holiday components. It indicates ‘discard’ data in a cell.
achieves time-series forecasting using the following relation: (ii) The input gate determines the period necessary to preserve
𝑥𝑡 = 𝑆̂𝑡1 + 𝑆̂𝑡2 + ⋯ + 𝑆̂𝑡𝑛 + 𝑇̂𝑡 + 𝑅̂𝑡 + 𝐻
̂𝑡 (7) the current data 𝑥𝑡 in a cell 𝐶𝑡 . It finds the state of the cell 𝐶𝑡 ,
where 𝑥𝑡 represents an observation at time 𝑡, 𝑛 denotes the and the data value is updated by the sigmoid layer, shown as
number of distinct seasonal patterns, 𝑆̂𝑡𝑛 represents the follows:
seasonal effect, 𝑇̂𝑡 represents the effect of trend, 𝑅̂𝑡 represents 𝑖𝑡 = 𝜎(𝑊𝑓 . [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑓 ) (10)
the remainder component of the observation and 𝐻 ̂𝑡 denotes The input gate updates the data that need to be upgraded to
the holiday covariance, which represents the effect of holidays. cell 𝐶𝑡 . A new vector 𝐶̂𝑡 is thus created by the tanh layer to
ARIMA is a simplified type of autoregressive moving determine the amount of data that must be added, and it is
average model that incorporates autoregressive and moving- defined as follows:
average models to develop a composite forecasting model. 𝐶̂𝑡 = 𝑡𝑎𝑛ℎ(𝑊𝑐 . [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑐 ) (11)
The autoregressive model utilises the dependency between the Finally, the state of the memory cell is updated as follows:
observations and several lagged observations, whereas the 𝐶𝑡 = 𝑓𝑡 ∗ 𝐶𝑡−1 + 𝑖𝑡 ∗ 𝐶̂𝑡 (12)
moving-average model uses the association between the
COVID-19 Prediction and Detection Using Deep Learning 175

(iii) The output gate controls how much of the current cell
state is discarded. The data are determined by a sigmoid layer.
The cell state is processed by the tanh layer and multiplied by
the output retrieved from the sigmoid layer to obtain the final
output of the cell, as shown below:
ℎ𝑡 = 𝜎 (𝑊𝜎 . [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑜 ) ∗ tanh(𝐶𝑡 ) (13)
To assess the performance of the implemented forecasting

Predicted death cases

algorithms, we collected the numbers of COVID-19
confirmations, recoveries, and deaths between 30 March 2020
and 5 April 2020 in Australia and Jordan. The predicted values
were then compared with the real numbers, and the results are
presented in figures. 8–13. Figures 8–10 show the predicted
COVID-19 confirmations, recoveries, and deaths in Australia
from 30 March to 5 April 2020 using the PA, ARIMA, and
LSTM algorithms, respectively. Figures 11–13 show the
predicted COVID-19 confirmations, recoveries, and deaths in Date
Jordan from 30 March 2020 to 5 April 2020. The three Figure 10. Predicted COVID-19 death cases in Australia
algorithms were compared in terms of RMSE, accuracy, and
correlation coefficient, as detailed in Tables 3 and 4.

Predicted confirmed cases

Predicted confirmed cases

Date
Figure 11. Predicted COVID-19 confirmed cases in
Jordan
Predicted recovered cases

Date
Figure 8. Predicted COVID-19 confirmed cases in Australia

Date
Figure 12. Predicted COVID-19 recovered cases in Jordan
Predicted recovered cases

Predicted death cases

Date Date
Figure 9. Predicted COVID-19 recovered cases in Australia Figure 13. Predicted COVID-19 death cases in Jordan
176 Alazab et al.

Table 3 details the performances of the three prediction

algorithms in Australia. We calculated the correlation
coefficient, accuracy, and RMSE for each algorithm. PA
delivered the best results, whereas ARIMA delivered the
worst results. Moreover, PA predicted the numbers of
COVID-19 confirmations, recoveries, and deaths and
obtained prediction accuracies of 99.94%, 90.29% and
94.18%, respectively.
Table 4 details the performances of the three prediction (a) Healthy (b) Coronavirus Infected
algorithms in Jordan. We calculated the correlation coefficient,
Figure 14. Chest X-ray images
accuracy, and RMSE for each algorithm. PA delivered the best
results, whereas LSTM delivered the worst results. Moreover,
The COVID-19 detection model is based on VGG16, a type
PA predicted the numbers of COVID-19 confirmations,
of CNN, and it is designed to detect COVID-19 using chest
recoveries, and deaths and obtained prediction accuracies of
X-ray images. VGG16 is one of the highest-quality vision-
99.08%, 79.39%, and 86.82%, respectively.
model architectures currently available. It generates evidence
in favour of stacking convolutional layers with tiny filters
Table 3 Results of three prediction algorithms in Australia
(3×3) rather than using a single layer with larger filter sizes
Prediction Correlation Accuracy RMSE (5×5 and 7×7) as stacked 3×3 filters will approximate bigger
Algorithm Coefficient ones [63].
PA confirmed cases 0.99 99.94% 03.94 The COVID-19 detector was trained and tested on the
PA (recovered cases) 0.99 90.29% 47.83 collected dataset, 80% of which was used for training and the
PA (death cases) 0.98 94.18% 01.55 remaining 20% was used for testing. The weights of the CNN
ARIMA (confirmed 0.99 92.33% 497.55 were randomly initialised, and the batch size was varied up to
cases) 25 and empirically set to 25 to avoid overfitting and to achieve
ARIMA (recovered 0.98 63.52% 260.12 the highest training accuracy. Furthermore, the learning rate
cases) was initially set to 0.1. Figure 15 details the accuracy of the
ARIMA (death 0.98 78.02% 07.26 COVID-19 detector and its loss values for the implemented
cases) detector with augmentation. Figure 16 details the accuracy of
LSTM (confirmed 0.99 94.16% 337.18 the same detector with augmentation and its loss values.
cases)
LSTM (recovered 0.99 86.44% 97.36
cases)
LSTM (death cases) 0.98 92.76% 02.07

Table 4 Results of the three prediction algorithms in Jordan

Prediction Correlation Accuracy RMSE
Algorithm Coefficient
PA (confirmed 0.99 99.08% 03.51
cases)
PA (recovered 0.94 79.39% 11.42
cases)
PA (death cases) - 86.82% 00.78
ARIMA (confirmed 0.95 97.59% 11.30
cases)
Figure 15. Performance of CNN models in the original dataset
ARIMA (recovered 0.93 57.79% 39.26
cases)
ARIMA (death -- 12.87 % 05.59
cases)
LSTM (confirmed 0.85 93.23% 25.02
cases)
LSTM (recovered 0.94 37.33% 11.42
cases)
LSTM (death cases) -- 39.97% 03.04
C. Deep Learning
This study developed a CNN-based COVID-19 detection
model that was tested with both the original and the
augmented datasets. All the chest X-ray images used were
resized to 224×224 pixels while ignoring the aspect ratio.
Figures 14a and b present the chest X-ray images of healthy Figure 16. Performance of CNN models in the augmented
and COVID-19-infected patients, respectively. The collected dataset
dataset was randomly split into a training data subset and a
testing data subset.
COVID-19 Prediction and Detection Using Deep Learning 177

Special attention must be paid to the avoidance of overfitting

in the un-augmented dataset, especially when increasing the
epochs as the validation slowly improves in the beginning and
then stops improving when the epochs are increased, as shown Dataset with augumntaion 99%
in Figure 15. When the augmented dataset is used, the gap
between the training and validation becomes smaller after a
few epochs, as shown in Figure 16. Thus, a greater
improvement is achieved in the training process, and a more
generalised and robust COVID-19 detector is achieved using Original dataset (without
95%
the CNN models when implementing data augmentation on augumntaion)
the dataset. Figures 17 and 18 show the augmented chest X-
ray images of COVID-19 patients and healthy persons,
respectively. 90% 92% 94% 96% 98% 100%

Figure 19. F-measure scores

V. Discussions
This study provided a forecasting analysis of COVID-19
confirmations, recoveries, and deaths in Australia and Jordan.
It further implemented a CNN-based COVID-19 detector to
identify COVID-19 infections using X-ray images. Based on
the study results, the following conclusions were drawn:
 PA delivered the best performance for COVID-19
prediction over 7 days, compared to LSTM and ARIMA.
 The predictions will enable people in both countries to
predict their medical needs for tackling the spread of COVID-
19.
 ARIMA cannot make predictions over the next 1, 2, and 3
days.
 After investigating the number of COVID-19 confirmations,
recoveries, and deaths in various countries, we found that
coastal areas are significantly impacted by the disease because
Figure 17. Augmented chest X-ray images for COVID-19 the numbers of cases in those areas are significantly higher
patients than those in other non-coastal areas. This observation is
medically consistent with the propagation capability of
viruses in areas with higher humidity rates. Thus, the authors
advise healthcare professionals to devote greater attention to
coastal regions.
 The use of chest X-ray images is recommended for
diagnosing COVID-19 because X-rays are easily obtained at
nearby hospitals or clinics fairly quickly and at low costs.
 Our CNN-based COVID-19 detector delivered superior
performance in terms of precision, recall, and F-measure.
 The application of ML techniques for COVID-19 diagnosis
using our CNN-based COVID-19 detector is recommended.
 It is well known that VGG16 (Wu et al., 2017) outperforms
many convolutional networks, such as GoogLeNet and
SqueezeNet, and its feature representation capability is
beneficial for classification accuracy. Hence, VGG16 is a
recommended version of a deep CNN-based algorithm as it
makes training easier and quicker. It was implemented in our
Figure 18. Augmented chest X-ray images for healthy people COVID-19 detector to improve its accuracy in diagnosing
COVID-19 in chest X-ray images.
The CNN-based COVID-19 detector trained on an un- Our COVID-19 detector obtained better results when using
augmented dataset achieved a weighted average F-measure of augmentation. A better training process was achieved as the
95%. The same COVID-19 detector achieved a weighted gap between the training and validation became smaller.
average F-measure of 99% when trained on an augmented Moreover, a more generalized and robust COVID-19 detector
dataset, as shown in Figure 19. Hence, the COVID-19 detector was achieved as the F-measure improved from 0.95 to 0.99.
exhibits superior performance metrics in terms of recall, Thus, the COVID-19 detector trained on augmented data
precision, and F-measure when trained on augmented data. It provides superior performance metrics and is robust for
is therefore sufficiently robust and helpful for rapidly
diagnosing COVID-19 in chest X-ray images.
diagnosing a large number of suspected COVID-19 patients.
178 Alazab et al.

VI. Conclusions and Future Work https://www.technologyreview.com/2020/02/13/844

805/coronavirus-china-app-close-contact-
The rapid spread of COVID-19 across the world and the surveillance-covid-19-technology/
increasing number of deaths require urgent actions from all [4] F. Jiang, L. Deng, L. Zhang, Y. Cai, C. W. Cheung,
sectors. Future prediction of potential infections will enable and Z. Xia, "Review of the clinical characteristics of
authorities to tackle the consequences effectively. coronavirus disease 2019 (COVID-19)," Journal of
Furthermore, it is necessary to keep up with the number of General Internal Medicine, pp. 1-5, 2020.
infected people by performing regular check-ups, and it is [5] Z. Wu and J. M. McGoogan, "Characteristics of and
often vital to quarantine infected people and adopt medical important lessons from the coronavirus disease 2019
measures. Additionally, attention should be given to several (COVID-19) outbreak in China: summary of a report
other factors to curb the spread of COVID-19, such as the of 72 314 cases from the Chinese Center for Disease
environmental effects and the similarities among the most Control and Prevention," Jama, 2020.
affected areas, and careful measures should be adopted. In this [6] T. Ai, Z. Yang, H. Hou, C. Zhan, C. Chen, W. Lv, et
paper, AI-based techniques were proposed for the prediction al., "Correlation of chest CT and RT-PCR testing in
and diagnosis of COVID-19: coronavirus disease 2019 (COVID-19) in China: a
 Prediction models such as the PA, ARIMA, and LSTM report of 1014 cases," Radiology, p. 200642, 2020.
algorithms were used to predict the number of COVID-19 [7] A. Narin, C. Kaya, and Z. Pamuk, "Automatic
confirmations, recoveries, and deaths over the next 7 days. PA Detection of Coronavirus Disease (COVID-19)
delivered the best performance. It predicted the number of Using X-ray Images and Deep Convolutional Neural
COVID-19 confirmations, recoveries, and deaths in Australia Networks," arXiv preprint arXiv:2003.10849, 2020.
and obtained prediction accuracies of 99.94%, 90.29%, and [8] H. S. Maghdid, A. T. Asaad, K. Z. Ghafoor, A. S.
94.18%, respectively. It also predicted the number of COVID- Sadiq, and M. K. Khan, "Diagnosing COVID-19
19 confirmations, recoveries, and deaths in Jordan and Pneumonia from X-Ray and CT Images using Deep
obtained prediction accuracies of 99.08%, 79.39%, and Learning and Transfer Learning Algorithms," arXiv
86.82%, respectively. Meanwhile, investigation into more preprint arXiv:2004.00038, 2020.
sophisticated forecasting and prediction methods is a subject [9] S. U. K. Bukhari, S. S. K. Bukhari, A. Syed, and S.
of a future work. S. H. SHAH, "The diagnostic evaluation of
 A diagnosis model using VGG16 was proposed to detect Convolutional Neural Network (CNN) for the
COVID-19 using chest X-ray images. The model allows the assessment of chest X-ray of patients infected with
rapid and reliable detection of COVID-19, enabling it to COVID-19," medRxiv, 2020.
[10] H. Shi, X. Han, N. Jiang, Y. Cao, O. Alwalid, J. Gu,
achieve an F-measure of 99% using an augmented dataset. In
et al., "Radiological findings from 81 patients with
a future study, we will consider diagnosing COVID-19 in
COVID-19 pneumonia in Wuhan, China: a
chest CT scan images using the VGG-XX versions and descriptive study," The Lancet Infectious Diseases,
compare their performances using larger datasets. 2020.
A further contribution of this study is the analysis of the [11] A. A. Ogunleye and W. Qing-Guo, "XGBoost model
COVID-19 spread and its related statistical data based on its for chronic kidney disease diagnosis," IEEE/ACM
global regional distributions. Thus, two main conclusions transactions on computational biology and
were drawn using our AI-based analysis: (i) the most highly bioinformatics, 2019.
infected areas have similar characteristics, and (ii) the spread [12] C. Feng, A. Elazab, P. Yang, T. Wang, F. Zhou, H.
of the disease in coastal areas is significantly higher than that Hu, et al., "Deep Learning Framework for
in other non-coastal areas. Therefore, extra care and attention Alzheimer’s Disease Diagnosis via 3D-CNN and
should be given to coastal cities. In our future work, we will FSBi-LSTM," IEEE Access, vol. 7, pp. 63605-63618,
investigate the effects of temperature, humidity, and terrain on 2019.
the COVID-19 spread in cities and countries. [13] H. Yin, B. Mukadam, X. Dai, and N. Jha, "DiabDeep:
Pervasive Diabetes Diagnosis based on Wearable
Acknowledgment Medical Sensors and Efficient Neural Networks,"
IEEE Transactions on Emerging Topics in
This work was supported by the faculty of Artificial
Computing, 2019.
intelligence, Al-Balqa Applied University. [14] K. Santosh, "AI-Driven Tools for Coronavirus
Outbreak: Need of Active Learning and Cross-
References Population Train/Test Models on
Multitudinal/Multimodal Data," Journal of Medical
[1] World Health Organization, "Laboratory testing for
Systems, vol. 44, pp. 1-5, 2020.
coronavirus disease 2019 (COVID-19) in suspected
[15] L. Wynants, B. Van Calster, M. M. Bonten, G. S.
human cases: interim guidance, 2 March 2020,"
Collins, T. P. Debray, M. De Vos, et al., "Systematic
World Health Organization, World Health
review and critical appraisal of prediction models for
Organization2020.
diagnosis and prognosis of COVID-19 infection,"
[2] Worldometers. (2020, April. 6). Coronavirus Cases.
medRxiv, 2020.
Available:
[16] A. Alazab, M. Hobbs, J. Abawajy, and M. Alazab,
https://www.worldometers.info/coronavirus/
"Using feature selection for intrusion detection
[3] A. Chen. ( 2020) China’s coronavirus app could have
system," in 2012 international symposium on
unintended consequences. MIT Technology Review.
communications and information technologies
Available:
(ISCIT), 2012, pp. 296-301.
COVID-19 Prediction and Detection Using Deep Learning 179

[17] [M. Alazab, S. Venkatraman, P. Watters, M. Alazab, [30] A. Mesleh, "Feature sub-set selection metrics for
and A. Alazab, "Cybercrime: The Case of Arabic text classification," Pattern Recognition
Obfuscated Malware," in Global Security, Safety and Letters, vol. 32, pp. 1922-1929, 2011.
Sustainability & e-Democracy. vol. 99, C. [31] A. Mesleh, "Support Vector Machine Text Classifier
Georgiadis, H. Jahankhani, E. Pimenidis, R. for Arabic Articles," ed: VDM Verlag Dr. Müller,
Bashroush, and A. Al-Nemrat, Eds., ed: Springer 2010.
Berlin Heidelberg, 2012, pp. 204-211. [32] K. Suzuki, "Overview of deep learning in medical
[18] M. Alazab., S. Venkatraman., P. Watters., and M. imaging," Radiological physics and technology, vol.
Alazab., "Information Security Governance: The Art 10, pp. 257-273, 2017.
of Detecting Hidden Malware," in IT Security [33] Y. LeCun, Y. Bengio, and G. Hinton, "Deep
Governance Innovations: Theory and Research, M. learning," nature, vol. 521, pp. 436-444, 2015.
Daniel, S. Luis Enrique, F.-M. Eduardo, and G. P. [34] C. Rachna. (2020, 15 April 2020). Difference
Mario, Eds., ed Hershey, PA, USA: IGI Global, 2013, Between X-ray and CT Scan.
pp. 293-315. [35] P. K. Sethy and S. K. Behera, "Detection of
[19] A. Alazab, M. Alazab, J. Abawajy, and M. Hobbs, coronavirus Disease (COVID-19) based on Deep
"Web application protection against SQL injection Features," 2020.
attack," in Proceedings of the 7th International [36] E. E.-D. Hemdan, M. A. Shouman, and M. E. Karar,
Conference on Information Technology and "A Framework of Deep Learning Classifiers to
Applications, 2011, pp. 1-7. Diagnose COVID-19 in X-Ray Images.," arXiv
[20] M. Alazab and L. Batten, "Survey in Smartphone preprint arXiv:2003.11055, 2020.
Malware Analysis Techniques," in New Threats and [37] A. E. Hassanien, L. N. Mahdy, K. A. Ezzat, H. H.
Countermeasures in Digital Crime and Cyber Elmousalami, and H. A. Ella, "Automatic X-ray
Terrorism, ed: IGI Global, 2015, pp. 105-130. COVID-19 Lung Image Classification System based
[21] M. Alazab., A. Alazab., and L. Batten., "Smartphone on Multi-Level Thresholding and Support Vector
malware based on synchronisation vulnerabilities," Machine," medRxiv, 2020.
in ICITA 2011: Proceedings of the 7th International [38] S. Wang, B. Kang, J. Ma, X. Zeng, M. Xiao, J. Guo,
Conference on Information Technology and et al., "A deep learning algorithm using CT images
Applications, Sydney, Australia, 2012, pp. 1-6. to screen for Corona Virus Disease (COVID-19),"
[22] V. Moonsamy., M. Alazab., and L. Batten., medRxiv, 2020.
"Towards an Understanding of the Impact of [39] D. Wang, B. Hu, C. Hu, F. Zhu, X. Liu, J. Zhang, et
Advertising on Data Leaks," International Journal of al., "Clinical characteristics of 138 hospitalized
Security and Networks (IJSN), vol. 7 2012. patients with 2019 novel coronavirus–infected
[23] L. M. Batten, V. Moonsamy, and M. Alazab, pneumonia in Wuhan, China," Jama, 2020.
"Smartphone applications, malware and data theft," [40] O. Gozes, M. Frid-Adar, H. Greenspan, P. D.
in Computational intelligence, cyber security and Browning, H. Zhang, W. Ji, et al., "Rapid ai
computational models, ed: Springer, 2016, pp. 15-24. development cycle for the coronavirus (covid-19)
[24] M. Alazab, V. Monsamy, L. Batten, P. Lantz, and R. pandemic: Initial results for automated detection &
Tian, "Analysis of Malicious and Benign Android patient monitoring using deep learning ct image
Applications," in International Conference on analysis," arXiv preprint arXiv:2003.05037, 2020.
Distributed Computing Systems Workshops [41] M. Fu, S.-L. Yi, Y. Zeng, F. Ye, Y. Li, X. Dong, et
(ICDCSW), 2012 32nd, 2012, pp. 608-616. al., "Deep Learning-Based Recognizing COVID-19
[25] Y. Xu, Y. Wang, J. Yuan, Q. Cheng, X. Wang, and and other Common Infectious Diseases of the Lung
P. L. Carson, "Medical breast ultrasound image by Chest CT Scan Images," medRxiv, 2020.
segmentation by machine learning," Ultrasonics, vol. [42] X. Xu, X. Jiang, C. Ma, P. Du, X. Li, S. Lv, et al.,
91, pp. 1-9, 2019. "Deep learning system to screen coronavirus disease
[26] A. Mesleh, "Lung Cancer Detection Using Multi- 2019 pneumonia," arXiv preprint arXiv:2002.09334,
Layer Neural Networks with Independent 2020.
Component Analysis: A Comparative Study of [43] M. Li, Z. Zhang, S. Jiang, Q. Liu, C. Chen, Y. Zhang,
Training Algorithms," Jordan Journal of Biological et al., "Predicting the epidemic trend of COVID-19
Sciences, vol. 10, 2017. in China and across the world using the machine
[27] A. Mesleh, D. Skopin, S. Baglikov, and A. Quteishat, learning approach," medRxiv, 2020.
"Heart rate extraction from vowel speech signals," [44] P. Kumar, H. Kalita, S. Patairiya, Y. D. Sharma, C.
Journal of computer science and technology, vol. 27, Nanda, M. Rani, et al., "Forecasting the dynamics of
pp. 1243-1251, 2012. COVID-19 Pandemic in Top 15 countries in April
[28] A. Mesleh, "Chi square feature extraction based 2020 through ARIMA Model with Machine
svms arabic language text categorization system," Learning Approach," medRxiv, 2020.
Journal of Computer Science, vol. 3, pp. 430-435, [45] C.-J. Huang, Y.-H. Chen, Y. Ma, and P.-H. Kuo,
2007. "Multiple-Input Deep Convolutional Neural
[29] A. Mesleh, "Support vector machines based Arabic Network Model for COVID-19 Forecasting in
language text classification system: feature selection China," medRxiv, 2020.
comparative study," in Advances in Computer and [46] G. Pandey, P. Chaudhary, R. Gupta, and S. Pal,
Information Sciences and Engineering, ed: Springer, "SEIR and Regression Model based COVID-19
2008, pp. 11-16. outbreak predictions in India," arXiv preprint
arXiv:2004.00958, 2020.
180 Alazab et al.

[47] N. Sajid. (2020, April.1). Corona Virus Dataset [61] J. Qiu, B. Wang, and C. Zhou, "Forecasting stock
Available: prices with long-short term memory neural network
https://www.kaggle.com/nabeelsajid917/covid-19- based on attention mechanism," PloS one, vol. 15,
x-ray-10000-images 2020.
[48] A. Buslaev, V. I. Iglovikov, E. Khvedchenya, A. [62] S. Hochreiter and J. Schmidhuber, "LSTM can solve
Parinov, M. Druzhinin, and A. A. Kalinin, hard long time lag problems," in Advances in neural
"Albumentations: fast and flexible image information processing systems, 1997, pp. 473-479.
augmentations," Information, vol. 11, p. 125, 2020. [63] F. Quiroga, R. Antonio, F. Ronchetti, L. C. Lanzarini,
[49] V. Jatana. (2020, April. 1). Coronavirus in Jordan. and A. Rosete, "A study of convolutional
Available: architectures for handshape recognition applied to
https://www.kaggle.com/vanshjatana/coronavirus- sign language," in XXIII Congreso Argentino de
in-jordan/notebook Ciencias de la Computación (La Plata, 2017). 2017.
[50] V. Jatana. (2020, April. 1). Coronavirus in Australia.
Available: Author Biographies
https://www.kaggle.com/vanshjatana/australia-
under-covid-19?scriptVersionId=32280319 Dr. Moutaz Alazab is a computer
[51] M. Alazab., M. Alazab., A. Shalaginov., A. Mesleh., security expert with industry, academic,
and A. Awajan., "Intelligent mobile malware teaching and research experience. He
detection using permission requests and API calls," completed his PhD degree in
Future Generation Computer Systems, vol. 107, pp. cybersecurity from Deakin university,
509-521, 2020. Australia in 2014. He is currently
[52] M. Alazab, "Automated Malware Detection in working as assistant professor in the
Mobile App Stores Based on Robust Feature faculty of Artificial Intelligence, Al-
Generation," Electronics, vol. 9, p. 435, 2020. Balqa Applied University. During his
[53] M. Alazab, "Analysis on Smartphone Devices for PhD, he was an active scholar in the Securing Cyberspace
Detection and Prevention of Malware," Doctor of Laboratory and Network and System Security Laboratory
Philosophy, Faculty of Science, Engineering and (NSCLab). Dr. Alazab has proved the ability in delivering a
Built Environment, Deakin University, 2014. high-quality content for several courses for both level
[54] M. Alazab, S. Venkatraman, P. Watters, and M. (undergraduate students and postgraduate students). He has
Alazab, "Zero-day malware detection based on lectured, coordinated, tutored and moderated in several well-
supervised learning algorithms of API call known universities including in BAU, CCQ, Deakin, RMIT,
signatures," in Ninth Australasian Data Mining CQU and MIT. Dr. Alazab has worked closely in
Conference: AusDM 2011, Ballarat, Australia, 2011, collaboration with industry on several research projects,
pp. 171-181. including BAE Systems, Microsoft and Ericsson. He is
[55] T. Zhang, Q. Wu, and Z. Zhang, "Probable pangolin recipient of number of research grants including research
origin of SARS-CoV-2 associated with the COVID- Incentive Fund (RIF), Zayed University. His research interests
19 outbreak," Current Biology, 2020. include Cybersecurity, Mobile Security, Network Security,
[56] CentersforDiseaseControlandPrevention. (2020, 01). Machine Learning, Digital Forensics, Blockchain, Internet of
Interim Clinical Guidance for Management of Things (IOT) and Big data analytics. He has published more
Patients with Confirmed Coronavirus Disease than 20 peer-reviewed articles in well-known, high-quality
(COVID-19). Available: international journals and conferences.
https://www.cdc.gov/coronavirus/2019-
ncov/hcp/clinical-guidance-management- Dr. Albara Awajan is an associate
patients.html professor at the Cyber Security
[57] AustralianGovernment. (2020, April. 01). Department, Faculty of Artificial
Coronavirus (COVID-19) current situation and case Intelligence, Al-Balqa Applied
numbers. Available: University. He completed his PhD in
https://www.health.gov.au/news/health-alerts/novel- Computer Networks and Multimedia
coronavirus-2019-ncov-health-alert/coronavirus- Applications in 2008 and his MCs
covid-19-current-situation-and-case-numbers Degree in 2003 in Multimedia and
[58] S. Kannan. (2020, April. 10). A drill-down analysis Internet Computing from the
of Covid-19 in India, so far. Available: University of Glamorgan in the UK. He completed his BCs
https://www.indiatoday.in/news-analysis/story/a- from Mutah University in 2001. During his PhD studies he
drill-down-analysis-of-covid-19-in-india-so-far- worked in the Mobile Communication Research Group in the
1665676-2020-04-10 Faculty of Advance Engineering. He is currently Head of the
[59] S. J. Taylor and B. Letham, "Forecasting at scale," Automated Systems Department and Assistant Dean for
The American Statistician, vol. 72, pp. 37-45, 2018. Student Affairs in the Artificial Intelligent Faculty. Dr Awajan
[60] A. Bazila Banu, R. Priyadarshini, and P. Research Areas are in AD-HOC networks, QoS, IoT and
Thirumalaikolundusubramanian, "Prediction of network Security.
Children Diabetes by Autoregressive Integrated
Moving Averages Model Using Big Data and Not
Only SQL," Journal of Computational and
Theoretical Nanoscience, vol. 16, pp. 3510-3513,
2019.
COVID-19 Prediction and Detection Using Deep Learning 181

Prof. Abdelwadood Mesleh is a Professor

at the Autonomous Systems Department,
the Faculty of Artificial Intelligence at Al- Vansh Jatanaand is Data Scientist born
Balqa Applied University (BAU). Prof. in Haryana, India. He is currently rank 13
Mesleh received his BSc and MSc degrees in the world as Grandmaster at Kaggle
in Computer Engineering from Shanghai (Data Science Platform). He has received
University, China, in 1995 and 1998 several international internships and
respectively, and his PhD in Computer Information Systems recognized by NASA. His area of
from the Arab Academy for Banking and Financial Sciences, interests includes machine learning,
Jordan, in 2008. Prof. Mesleh worked as a research assistant natural languages processing and Data Science.
at Hong Kong University of Science and Technology, China,
from 2004 to 2005, and worked as a professor in the Faculty
of Engineering Technology at BAU from 2005 to 2019. His
research interests include artificial intelligence and its
applications.

Dr. Abraham is the Director of Machine

Intelligence Research Labs (MIR Labs),
a Not-for-Profit Scientific Network for
Innovation and Research Excellence
connecting Industry and Academia. As
an Investigator / Co-Investigator, he has
won research grants worth over 100+
Million US$ from Australia, USA, EU, Italy, Czech Republic,
France, Malaysia and China. Dr. Abraham works in a multi-
disciplinary environment involving machine intelligence,
cyber-physical systems, Internet of things, network security,
sensor networks, Web intelligence, Web services, data mining
and applied to various real-world problems. In these areas he
has authored / co- authored more than 1,300+ research
publications out of which there are 100+ books covering
various aspects of Computer Science. One of his books was
translated to Japanese and few other articles were translated to
Russian and Chinese. About 1000+ publications are indexed
by Scopus and over 800 are indexed by Thomson ISI Web of
Science. Dr. Abraham has more than 37,000+ academic
citations (h-index of 91 as per google scholar). He has given
more than 100 plenary lectures and conference tutorials (in
20+ countries). Since 2008, Dr. Abraham is the Chair of IEEE
Systems Man and Cybernetics Society Technical Committee
on Soft Computing (which has over 200+ members) and
served as a Distinguished Lecturer of IEEE Computer Society
representing Europe (2011-2013). Currently Dr. Abraham is
the editor-in-chief of Engineering Applications of Artificial
Intelligence (EAAI) and serves/served the editorial board of
over 15 International Journals indexed by Thomson ISI. Dr.
Abraham received Ph.D. degree in Computer Science from
Monash University, Melbourne, Australia (2001) and a
Master of Science Degree from Nanyang Technological
University, Singapore (1998).

Dr. Salah Alhyari is a visiting lecturer at

the University of Jordan since 2017. He is
a senior of computer operation department
at Jordanian Electric Power Company
(JEPCO) since 2008. He received his PhD
in Business Administration - Logistics
Science and Supply Chain Management in
2015. He received his master’s degree in business
administration and bachelor’s degree in management
information systems. His area of interests includes Artificial
Intelligence, supply chain management, logistics science,
operation management, quality management and information
systems management.

View publication stats