Expert Systems With Applications 210 (2022) 118227

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

A hybrid random forest deep learning classifier empowered edge cloud

architecture for COVID-19 and pneumonia detection
M Hemalatha
Department of Faculty of Artificial Intelligence and Data Science, R.M.K Engineering College (Autonomous), Kavaraipettai, India

A R T I C L E I N F O A B S T R A C T

Keywords: COVID-19 is a global pandemic that mostly affects patients’ respiratory systems, and the only way to protect
Web services oneself against the virus at present moment is to diagnose the illness, isolate the patient, and provide immu­
Healthcare industry nization. In the present situation, the testing used to predict COVID-19 is inefficient and results in more false
Cloud computing
positives. This difficulty can be solved by developing a remote medical decision support system that detects
Deep learning
Heat Transfer Search algorithm
illness using CT scans or X-ray images with less manual interaction and is less prone to errors. The state-of-art
Random forest techniques mainly used complex deep learning architectures which are not quite effective when deployed in
resource-constrained edge devices. To overcome this problem, a Multi-Objective Modified Heat Transfer Search
(MOMHTS) optimized Hybrid Random Forest Deep learning (HRFDL) classifier is proposed in this paper. The
MOMHTS algorithm mainly optimizes the deep learning model in the HRFDL architecture by optimizing the
hyperparameters associated with it to support the resource-constrained edge devices. To evaluate the efficiency
of this technique, extensive experimentation is conducted on two real-time datasets namely the COVID19 lung CT
scan dataset and the Chest X-ray images (Pneumonia) datasets. The proposed methodology mainly offers
increased speed for communication between the IoT devices and COVID-19 detection via the MOMHTS opti­
mized HRFDL classifier is modified to support the resources which can only support minimal computation and
handle minimum storage. The proposed methodology offers an accuracy of 99% for both the COVID19 lung CT
scan dataset and the Chest X-ray images (Pneumonia) datasets with minimal computational time, cost, and
storage. Based on the simulation outcomes, we can conclude that the proposed methodology is an appropriate fit
for edge computing detection to identify the COVID19 and pneumonia with higher detection accuracy.

1. Introduction made a massive amount of people suffer who are affected by this disease
in the rural areas. These techniques also faced severe criticism due to the
The healthcare system plays a major role in predicting one’s emer­ high number of false-negative predictions (Carpenter, Mudd, West,
gency situation and when diagnosed in the initial stage it can even in­ Wilber, & Wilber, 2020). Hence, this arises the need for an efficient
crease the lifetime of the patients. The novel Coronavirus infection that Medical decision support system for COVID19 diagnosis.
developed in Wuhan, China, and was recognized by the World Health Pneumonia is a disease that affects the lungs where the air sacs get
Organization (WHO) in December 2019 is seen as a significant threat to filled with pus which results in chills, fever, and breathing problems.
mankind (Sohrabi et al., 2020).This disease outbreak turned into a This disease mainly occurs due to different bacteria like Streptococcus,
pandemic by March 2020. The symptoms of the disease are fever, dry Staphylococcus, Pseudomonas, Haemophilus, Chlamydia, Mycoplasma,
cough, breathing problems, loss of taste and smell, etc. The symptoms of several viruses, and certain fungi, and protozoans. This disease can be
the virus develop within a time interval of 1 to 14 days. The WHO divided into two forms, bronchial pneumonia, and lobar pneumonia.
mainly suggested the RT-PCR Test which is a real-time reverse The manual analysis of a large number of CT scans and X-rays is error-
transcription-polymerase chain reaction (rRT-PCR) test to identify prone and time-consuming (Das, Kumar, Kaur, Kumar, & Singh,
COVID-19 (Ford et al., 2021). Virus traces from the sick person’s nose 2020). The COVID19 and pneumonia mainly target the respiratory
and mouth are usually taken with a swab to diagnose this condition. To system of the human body, the chest X-rays and CT scan images play a
identify The COVID-19 disease within 30 min, rapid diagnostic tests vital role in diagnosing this disease. Artificial intelligence and machine
were also developed. The high cost and low production of these devices learning techniques tend to be effective to yield insights from the

E-mail address: ahema_me@rediffmail.com.

https://doi.org/10.1016/j.eswa.2022.118227
Received 3 February 2022; Received in revised form 8 June 2022; Accepted 17 July 2022
Available online 21 July 2022
0957-4174/© 2022 Elsevier Ltd. All rights reserved.
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

massive data collected (Hemalatha & Rukmanidevi, 2017; Hemalatha, score.

Rukmanidevi, & Shanker, 2021; Hung et al., 2020; Jose et al., 2021; The rest of this paper is arranged accordingly. Section 2 presents the
Rodríguez-Rodríguez, Rodríguez, Shirvanizadeh, Ortiz, & Pardo-Quiles, summary of the literary works and section 3 elaborates on the proposed
2021; Sundararaj, 2016). The increased volume of data can be efficiently model in detail. Section 4 presents the experimental analysis carried out
handled via edge computing techniques. Edge computing (Xu, Chen, & using different performance metrics and conventional techniques to
Ren, 2017) mainly increases the processing capabilities of edge devices evaluate the efficiency of the proposed methodology. Section 5 discusses
via on-site processing before transferring the data to the cloud. The edge the details of the proposed methodology for COVID19 outbreak pre­
devices have a pre-cloud layer where the computation-intensive IoT vention and section 6 concludes this paper.
devices collect the information and do the substantial processing
without depending on the cloud to send the information to the user. 2. Literature review
Edge computing is suitable for applications that incorporate IoT
devices that are interconnected with the edge devices or cloud. In this EdgeCare, a leveraging edge computing solution built for mobile
way, the number of packets that need to be transferred to the cloud is healthcare systems employing collaborative data management, was
minimized which directly affects energy consumption and increased presented by Li, Huang, Li, Yu, and Shu (2019). They are mainly
storage (Mach & Becvar, 2017). Since the integration of IoT and edge incorporating decentralized and collaborative data management in this
devices offers diverse benefits, in this work the former is applied to work to support the massive volume of global healthcare data, enhance
process the CT scans. These devices also help to perform remote di­ the overall system performance, and minimize the complexity associ­
agnoses in locations far away and the patients cannot afford to get the ated with real-time healthcare applications. The healthcare data is
treatment from a radiologist. Since COVID19 and Pneumonia cause se­ processed and traded using the local authorities who manage the edge
vere health risks to both elderly and younger patients it benefits the server. The optimal incentive mechanism is offered by the Stackelberg
patients when identified at an earlier stage. In healthcare applications, game-based optimization algorithm which assists both the users and
the minimum powered devices are called the edge devices and the model data miners in trading.
designed for these devices needs to consume minimal computation and Vasconcelos, Sarmento, Reboucas Filho, and de Albuquerque (2020)
energy without minimizing the performance of the application. The presented an improved edge-cloud architecture using artificial intelli­
main aim of the proposed work is to develop a cloud edge computing- gence techniques for brain CT image analysis. In this work, the authors
based application which forwards the computationally intensive oper­ are mainly identifying the stroke disease with the help of the Computed
ations to the cloud and sends the results to the edge devices. The Tomography (CT) scan result. Even though Magnetic Resonance Imag­
incorporation of edge computing in our proposed work helps to mini­ ing (MRI) is preferred by most authors, CT scan is taken by the majority
mize the latency for delay-sensitive applications and cloud computing of people due to the minimal cost and time. The applications developed
solves the memory-related issues by offering additional storage for the Internet of Things (IoT) devices need to support low computation
capacities. and low storage costs. Hence the authors proposed an adaptive analysis
The emerging usage of deep learning techniques for natural language of the Brain Tissue Densities (Adaptive ABD) model using edge
processing, image processing, speech recognition, and object identifi­ computing devices for the excellent processing capabilities they offer.
cation has inspired us to incorporate a deep learning classifier for the This methodology offers a low computational cost with an average
prediction of COVID19 and pneumonia (Jamshidi et al., 2020; Khema­ execution time of 0.087 s per sample. However, the efficiency of this
suwan, Sorensen, & Colt, 2020). Since the Deep Learning (DL) classifier method is evaluated in a small CT scan dataset.
does not involve manual training for feature selection and classification Singh and Kolekar (2021) identified the COVID19 via CT images
it helps to minimize the time taken for classifying the normal and using collaborative edge cloud computing. They mainly used a fine-
abnormal samples. The reliable classification results are provided by the tuned MobileNet V2 model since it is often complex to implement the
deep learning classifier due to the usage of the increased number of deep learning model in resource-constrained devices. To support the
hidden layers which does intricate processing (Shone, Ngoc, Phai, & Shi, mobile and edge devices even further, the MobileNet V2 architecture
2018; Lane & Georgiev, 2015). was even optimized in terms of complexity and size. The experiments
This paper presents a hybrid Random Forest Deep learning (HRFDL) were conducted on a real-world CT scan images dataset and the classi­
classifier for COVID-19 and pneumonia disease diagnosis in IoT-based fication accuracy of this model was 96.40%. However, the optimal
edge devices. The Random Forest classifier is mainly selected due to hyperparameters for the Mobile V2 model are not optimized.
its ability to handle the high dimensional data and rapid convergence. Akkaoui, Hei, and Cheng (2020) employed an EdgeMediChain
However, the results obtained from the random classifier have a high framework (PMLA) which utilizes the collaborative and distributed data
false-positive rate. To overcome this problem, the MLP is used in this management framework supported by blockchain and edge computing.
paper which offers a high True Positive Rate and a low false-positive rate The main problem they are trying to overcome is the inability of cloud
(FPR). This is the main reason to hybridize the RF and MLP algorithms computing architecture to process the massive data generated from the
using a majority voting rule to yield an improved detection accuracy. body sensors. The main advantages offered by this methodology are the
The hyperparameters of the Deep learning architecture are optimized higher throughput and minimal execution time which is nearly equal to
using the Multi-Objective Modified Heat Transfer Search (MOMHTS) 84.75% for a total of 2000 simultaneous transactions. However, this
which is formulated by integrating the three modes of the standard Heat methodology does not offer real-time disease diagnosis results which
Transfer Search algorithm (HTS). help in efficient decision making.
To prevent the local optima trapping present in the heat transfer Pustokhina et al. (2020) utilized both cloud computing and edge
search (HTS) algorithm and achieve an efficient trade-off between the computing techniques to monitor the vital signs of the patients on a daily
exploration and the exploitation phases of this algorithm, this paper basis. To achieve this objective, they presented an approach known as
adds a new step called the synchronous heat transfer to the conventional the Effective Training Scheme for the Deep Neural Network (ETS-DNN)
HTS algorithm. In this way, a MOMHTS algorithm is formulated which for timely data collection and processing from the internet of medical
helps to minimize the energy and space complexity associated with the things (IoMT) devices to identify the internal patterns that exist in the
resource-constrained edge devices when deploying the HRFDL classifier. data. The patient’s data is captured via the IoMT sensors and sent to the
The performance of the model is evaluated using two real-world medical edge computing platform enabled with the ETS-DNN technique for
datasets namely the COVID-19 lung CT scan dataset and the Chest X-ray processing. A Hybrid Modified Water Wave optimization algorithm for
images (Pneumonia) dataset via different performance metrics such as tuning the DNN parameters. The hybrid algorithm is formed by inte­
sensitivity, specificity, ROC curve, accuracy, confusion matrix, and F1- grating the Modified Water Wave Optimization algorithm with

2
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm. The report different sizes and complexity, the machine learning algorithms need to
generated by the ETS-DNN model is forwarded to the cloud server which be optimized. In the field of computer science, deep learning has shown
provides access to the healthcare professionals. revolutionary improvement due to its efficiency in accessing different
Muhammed, Mehmood, Albeshri, and Katib (2018) developed a datasets, powerful parallelization, and specialized hardware develop­
personalized ubiquitous cloud and edge-enabled networked healthcare ment. The processing power of the edge devices is too low to deploy the
system known as UbeHealth which integrates the IoT, edge computing, deep learning architectures. The proposed architecture is mainly
deep learning, big data, and High-Performance Computing (HPC). They implemented to support end devices with power constraints and perform
are mainly focusing on optimizing the Quality of Service (QoS) param­ computationally efficient operations.
eters, energy consumption, and bandwidth. The improved QoS is offered The diseases such as COVID-19 and pneumonia are monitored in our
to its users using the cloudlet, mobile, network, and cloud layers. proposed framework shown in Fig. 1 using the MOMHTS optimized
Mohammedqasim and Ata (2022) enhanced the accuracy of COVID19 HRFDL architecture. The proposed three-tier architecture comprises
diagnosis via an optimized deep learning architecture. They are opti­ three layers: data generation, edge, and cloud. The IoT layer is mainly
mizing the deep learning architecture via grid search to overcome the formed via the sensors, actuators, and information exchange devices.
unbalanced data problem. In this way, they are minimizing the pro­ These are the IoT nodes that acquire the data from the physical world.
cessing time. Based on the collected data these devices provide the results (COVID 19/
Bhatia, Manocha, Ahanger, and Alqahtani (2022) handled the Pneumonia positive) to the users after processing the values obtained.
COVID19 outbreak via an artificial intelligence technique. They are The storage and computing devices are in the cloud layer and the
mainly identifying the COVID19 outbreak with the help of wearable intermediary devices are in the edge layer. The sensor node needs a
sensors and the Radio Frequency Identification Device (RFID). The small powerful processing unit to efficiently connect with other layers.
infection degree and disease outbreak are analyzed with the help of the The data acquisition layer mainly collects the crucial physiological
J48 decision tree and Temporal Network Analysis. Ghosh and Ghosh symptoms, personal information, users’ contact number, etc. Every user
(2022) identified COVID and pneumonia via a Deep residual neural initially obtains their unique identifier (UID) by entering their phone
network-based chest X-ray image enhancement (ENResNet) approach. numbers and other personal details. The primary symptoms such as
they have built the ENResNet using eight residual modules and the body temperature and blood pressure are captured via the Wireless Body
model is trained using residual images. The residual images are gener­ Area Networks (WBAN) and sent to the user’s mobile device via Blue­
ated via batch normalization. tooth. The data obtained is sent to the edge server via the WiFi or 4G/5G
Ding, Li, Li, Wang, and Zhang (2019) utilized a Tinier YOLO algo­ network. The advanced symptoms are obtained via X-ray and CT scan
rithm for diagnosing the upper gastrointestinal disease in real time via images. After the symptoms of the users are identified and updated, the
the cloud-edge collaborative framework. This work improved the data is immediately transferred to the cloud.
sensitivity and specificity of upper gastrointestinal screening. They are The disease prediction and monitoring framework are deployed in
integrating the cloud and edge platform to provide real-time lesion the bottom layer. Initially, the data is collected and then sent to the edge
identification in the upper gastrointestinal tract. Ortiz, Zouai, Kazar, layer which consists of edge devices and the proposed methodology is
Garcia-de-Prado, and Boubeta-Puig (2022) presented a study to analyze also installed in this layer. If any individual is diagnosed positive
respiratory illness by solving the issues by integrating the three tiers (COVID19 or pneumonia), then the edge devices alert the concerned
(fog, cloud, and IoT) via a two-way communication protocol. The two- authorities (Patient, guardian, medical institution, or hospital) with an
way communication improves the real-time decision-making perfor­ alert message. The progression of the disease is analyzed by the cloud
mance by acquiring the proper contextual and location information. layer based on the patient’s ID and their location derived. This infor­
mation helps to prevent the progression of the disease in a certain area.
2.1. Research gap The detailed description of each layer is presented below:

The COVID19 is an infectious disease that is rapidly spread world­ 3.1. Layer-1: Data generation
wide and causes serious health risks. This arises the need for early health
monitoring and prevention. The state-of-art disease risk prediction sys­ The data regarding the illness (COVID19 and Pneumonia) is acquired
tem is capable of identifying the diseases up to a certain extent. How­ from this layer. To diagnose pneumonia and COVID19, we retrieved the
ever, the accuracy of the results is mainly affected by the diversity and CT scan results of the patients. To help the government and other NGO
incomplete data collected from the IoT sensors. The restricted IoT agencies to provide medical aid, resources, and services to the patient.
resource capabilities also limit the processing of large amounts of data The data acquired is transformed into the edge layer for further pro­
on a timely basis. The existing system faced major challenges such as low cessing where different edge devices are present including the patient’s
accuracy and F1-score, deployment on imbalanced datasets, limited edge device.
dataset availability, used pre-trained architectures with fixed input size
which is not applicable in realtime, overfitting issues, focused on binary 3.2. Layer-2: Edge layer
classifications, and a minimal number of samples used for training
resulting in increased false-positive rates. To address these challenges, The edge layer is mainly used for processing and classification of the
this paper presents a novel MOMHTS optimized HRFDL classifier for CT scan images acquired. It acts as an intermediate between the cloud
monitoring COVID19 and pneumonia via edge computing. Section 3 and the physical layer. The edge layer minimizes the network traffic and
presents the details of the disease monitoring in detail along with the latency. The proposed MOMHTS optimized HRFDL model is imple­
improvements made in the existing artificial intelligence models by mented in the edge layer with resource-constrained edge devices. The
integrating edge computing technology. resource-constrained devices can also act as IoT devices. To increase the
computational efficiency different hardware accelerators are also
3. Proposed methodology attached to the end devices. If any abnormalities exist in the patient’s CT
scan, then the disease diagnosed is sent as a message to the user. To
The state-of-art works provide a reliable medical decision support further notify the healthcare institutions and the officials the patient ID
system to aid in efficient decision making by the healthcare pro­ along with the location is sent to the cloud for further processing.
fessionals. Several kinds of research have been conducted to make the
machine learning algorithms support edge devices and minimize their
latencies along with a security improvement. To support edge devices of

3
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

Fig. 1. Outline of the proposed framework.

3.3. Layer-3: Cloud layer Since the information is accessible by the public, in this way both the
public and the civic bodies are alerted.
The cloud layer performs various complex tasks to overcome the • Offering model updates: After the emergence of the novel Corona­
limitations associated with edge devices such as data storage and low virus, there was only limited data available to train the deep learning
computational power. The cloud layer performs centralized operations model. Since the training of edge devices is very complex, the pro­
over multiple Virtual Machines (VM). The data stored in the centralized posed model is uploaded to the cloud and the result updates obtained
data warehouse can be accessed via government authorities and are sent to the edge devices. The on-the-air updates are essential for
authorized medical practitioners. The cloud layer normally does the every health-related application since we have included pneumonia
following operations: disease. The edge layer identifies the diseases as shown in algorithm-
1.
• Resource management for disease outbreaks: The increase in the
number of COVID19 and pneumonia patients leads to a shortage of Algorithm-1: Disease diagnosis at the edge layer.
medical equipment. To identify these diseases different medical Input: CT images of chest area affected with COVID19 and pneumonia
types of equipment are needed such as testing devices, respiratory Output: Disease class whether normal or abnormal (COVID19 or pneumonia)
Step-1: From the data generation layer, retrieve the chest CT images and create a
devices, ventilators, and oxygen cylinders. Since ventilators and
unique patient ID for each record
respiratory devices are crucial in treating the disease, appropriate Step-2: The images obtained from the data acquisition layer are preprocessed and
equipment handling is necessary. Hence a tradeoff between these then the size of the image is altered to match the proposed model
types of equipment, supply, and demand needs to be achieved. The Step-3: The image is given as an input to the proposed model and the classification
places that face an increase in COVID19 patients need a more supply result is taken as output
Step-4: If Class predicted = Abnormal (COVID19/Pneumonia)then
of this equipment. Hence with the help of the cloud, one can track a
Output the classifier result to the user along with their patient ID for future
disease outbreak, the number of people affected, and offer resource admissions
optimization. Send both the location and patient ID to the cloud to alert the government/health
• Managing the patient’s data: The records related to the number of organizations to conduct the patient checkup
Else
patients being infected, number of mortality, number of active cases,
Output the classified result to the user with the date for the next checkup
and the number of recoveries. Since we need regular updates about Erase the existing patient ID and their location from the cloud storage
the disease outbreak, this information is continuously monitored. End If
Based on the information, necessary precautions are taken to control Step-5:Exit
the virus spread and treat the patients.
• Tracking the outbreak: Since the infection rate and disease spread is
high, it is necessary to control the disease outbreak and identify the
locations of potential danger. Both the local information and the
newly identified disease cases are uploaded to the cloud periodically.

4
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

3.4. Formulation of MOMHTS algorithm with each other to establish thermal equilibrium. This is governed by the
Stefan–Boltzmann law of thermodynamics. The equation for thermal
3.4.1. Heat transfer search radiation mode is given in eq (5) and eq (6).
Using the basic principles of heat transfer and thermodynamics, the ′ ( )
HTS algorithm is formulated (Kumar, Tejani, Pholdee, & Bureerat, Mx,μ = Mx,μ + prob × (My,μ − Mx,μ ), if F(Mx )〉F My ; if iter⩽itermax /RDF
( )
2021). The algorithm mainly implies that a thermal system can achieve
′
Mx,μ = Mx,μ , +prob × (Mx,μ − My,μ ), if F(Mx )〈F My ; if iter⩽itermax /RDF
thermal equilibrium with different modes of heat transfer (conduction, (5)
convection, and radiation). These heat transfer modes are also the
( )
search procedures in the HTS optimizer. The equal selection probability ′
Mx,μ = Mx,μ + probβ × (My,μ − Mx,μ ), if F(Mx )〉F My ; if iter > itermax /RDF
value of each mode is assigned a random number ρ in the range [0,1]. ′ ( )
Mx,μ = Mx,μ , +probβ × (Mx,μ − My,μ ), if F(Mx )〈F My ; if iter > itermax /RDF
For the conduction mode, the value of ρ is in the range 0–0.333 and in
the radiation mode, it is in the range 0.333–0.6666. The range of con­ (6)
vection mode is 0.6666–1. Based on the ρ value, the results are updated The updated molecules in the thermal radiation model is indicated
as per the heat transfer in each iteration. asMx,μ ; x = 1, 2, 3, …, n; µ ∈ (1, 2, …, m); x ∕
′
= y; y ∈ (1, 2, 3, …, n); the
randomly selected molecules are indicated as y and the current iteration
3.4.1.1. Thermal conduction mode. Molecules exchange heat from is indicated as iter; Prob ∈ [0.3333,0.6666] is probability variable. The
higher energy levels to lower energy levels when it contacts each other randomly selected number that lies under the range of 0 to 1 is given
this process is said to be conduction. The conduction mode in the HTS asp2 and pβ and also known as radiation parameters of Stefan–Boltzmann
algorithm is divided into two segments. equation; meanwhile the variation of temperature between the sur­
Segment 1: when iter⩽itermax /RDF roundings and molecules of the system is given as Mx andMy . In order to
′ ( ) ( ) maintain the balance between diversification and intensification, we
Mx,μ = My,μ + − ρ2 × My,μ , if F(Mx )〉F My
′ ( ) ( ) (1) have set the RDF (radiation factor) as 2. The variables belonging to the
Mx,μ = Mx,μ + − ρ2 × Mx,μ , if F(Mx )〈F My
iteration are changed during the iteration process of radiation.
Segment 2: whenif iter > itermax /CDF
( ) ( ) 3.4.2. Multi-objective modified heat transfer search (MOMHTS) optimizer
The metaheuristic optimizer is mainly applied to generate new so­
′
Mx,μ = My,μ + − ρβ × My,μ , if F(Mx )〉F My
( ) ( ) (2)
′
Mx,μ = Mx,μ + − ρβ × Mx,μ , if F(Mx )〈F My lutions that are better than the previous ones or search for a global
optimum in a feasible search space. Another feature of the metaheuristic
The molecules that are updated are represented asMx,μ ; x = 1, 2, 3, algorithm is that the optimizer should prevent local optima trapping. If
′

…, n; the randomly selected solutions are indicated as y; x∕=y; y ∊ (1,2,3, the above features are integrated together, the excellent performance of
….,n); the design variable index is denoted as µ and is selected randomly the algorithm can be derived. To achieve this objective one needs to
and can be given as µ ∊ (1,2,…..,m); the current iteration is given as iter achieve a tradeoff between two phases namely the exploration and
and the total number of iteration is denoted asitermax ; the temperature exploitation phases. The exploration is known as diversification and the
change of the molecules is given as Mx and My and are called conduc­ exploitation is known as the intensification process. The exploitation
tance parameters. CDF is the conduction factor and to maintain the process improves the convergence rate whereas the exploration process
balance between intensification and diversification CDF is indicated as minimizes the convergence rate as per the postulates of some popular
2. In each iteration, there will be one design variable modification that metaheuristic algorithm and their convergence behavior. An increased
happens. exploration phase also helps to identify the global optima but with less
efficiency and in some cases increased exploitation leads to premature
3.4.1.2. Thermal convection mode. Convection mode always removes convergence. In these scenarios, an accurate tradeoff between the
the energy level indifference between the system and surrounding by exploration and exploitation phases needs to be achieved which is an
convection heat transfer. The system molecules Sm and system Sur­ unresolved issue in optimization.
rounding Ss associate with each other to establish thermal equilibrium. In Heat Transfer Search (HTS) algorithm, the system molecules
Equation for thermal convection mode is given in eq (3) and (4). interact with their adjacent molecules to minimize the thermal imbal­
ance and transfer heat. The instant energy transfer process is achieved
(3)
′
Mx,μ = Mx,μ + ρ × (Ss − Sm × TCF) using any one of the three HTS modes. For polynomial functions, the
( ) radiation process tends to be more effective and the convection and
TCF = abs( ρ − nβ) , if iter⩽ itermax /CF
(4) conduction processes are effective when solving the non-linear func­
TCF = round 1 + nβ , if iter > itermax /CF tions. The HTS model transfers continuously to speed up the thermal
balance.
Mx,μ represents the updated molecules; x = 1, 2, 3, …, n; µ ∊ (1, 2, …,
′

To overcome the limitations of the HTS algorithm, a multi-objective

m), which indicates design variable index; the function estimation is Modified Heat Transfer Search (MOMHTS) is presented and it integrates
represented as iter; Prob is the probability variable and can be given as P the three modes of the standard HTS algorithm is integrated. The HTS
∈ [0.6666, 1]; itermax is the maximum number of iteration;pβ is an arbi­ algorithm mainly offers premature convergence since the new solution
trary number varying from 0 to 1; the newton’s law for the cooling created is mainly based on the randomly selected solution, best
convection parameters can be given as p2 andpβ ; the surrounding tem­ solution, and average solution due to the high proximity in between
perature isSs and has been considered as a constant reference; meanwhile them. To achieve a tradeoff between the exploration and exploitation
the mean temperature of the system can be denoted asSm , it will change phases, a new step known as synchronous heat transfer is included in
during the convection process; the tradeoff between the exploration and the standard HTS algorithm.
exploitation is represented by TCF in the convection phase; and CF is set
to be 10. 3.4.2.1. Synchronous heat transfer. In the HTS algorithm, the energy is
mainly interacted in the form of heat to achieve thermal equilibrium
3.4.1.3. Thermal radiation mode. Due to the heavy temperature, heat between the system and adjacent molecules. Since the three heat
transfer occurs in the form of electromagnetic waves due to radiation transfer modes are integrated into the model, heat transfer mostly occurs
released. The system moleculesSm and system surroundingsSs associate since for every generation there is an equal likelihood. The synchronous

5
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

heat transfer speeds up the search process and the heat transfer possi­ For an instance, if there is a normal CT scan that resembles an
bility mainly relies on the probability factor of conduction (PC), proba­ abnormal sample, then if two models classify the sample as normal and
bility factor of convection (PV), and probability factor of radiation (PR). another model classifies this sample as abnormal which is correct means
The values of the PC, PV, and PR values range from 0 to 1 and. the simple voting rule mainly identifies the sample as benign. To over­
When the conduction mode is activated during heat transfer, the first come this issue, a rule-based majority voting scheme is proposed in this
one-third of the molecules in the population are updated. The second paper which utilizes the benefits of the minority classification to mini­
and third one-third solutions are updated during the radiation and mize the false alarm rate. The majority voting scheme mainly offers a
convection mode. To obtain a synchronized form of optimization, the higher probability value to the classifier that identifies the normal
conduction, convection, and radiation modes are integrated to support sample well. To find a reliable classifier, a priority rule is added in
nonlinear, linear, and polynomial functions. Hence the value of PC , PV , identifying the normal samples. A simple majority voting rule is applied
and PR are assigned as 0.333 respectively to achieve an equal probability for the samples not classified using the priority rule. Based on the
value. The three modes are simultaneously implemented by substituting probabilistic density, a decision-making algorithm is provided to iden­
the probability variables such as ρ1 (Conduction probability value), ρ2 tify the abnormal samples.
(radiation probability value), and ρ3 (Convection probability value). In real-time, it is often complex to identify the abnormal samples
The outline of the MOMHTS algorithm is presented in Fig. 2. accurately and in these scenarios, a single machine learning classifier is
not adequate to offer a higher true positive rate and the low false-
3.5. HRFDL architecture formation positive rate at the same time. Hence to improve the machine learning
model in terms of detection rate we add a deep learning classifier. To
The disease detection rate from the CT image dataset is improved improve the detection rate of the deep learning classifier we are using
using the hybrid model which comprises of three architectures. Initially, the MOMHTS algorithm. The initial and optimized deep learning model
the feature extraction is performed and in the next stage, the classifi­ is presented in Table 1.
cation is conducted using the optimized classifiers (Yoo, Kim, Kim, &
Kang, 2021). To improve the detection performance a step further, a
voting methodology is used in the last step. The classifiers with high true Table 1
positive rates are selected to enhance the voting effects. The disease is Optimized DL values using MOMHTS algorithm.
mainly identified by the standard machine learning model with a scoring Parameters Initial Values Optimized Values
range of 0 to 1. The detection process is identified using the random Number of 250 205
forest and Multi Layer Perceptron (MLP) algorithm. In the feature nodes
extraction stage, the features from the CT scan images are extracted and Epoch 512 364
Batch size 2048 Greater than 200
the classification is done using deep learning and the random forest al­
Hidden layer 16 12
gorithm. At last, a rule-based majority voting scheme is deployed to Optimizer Stochastic gradient Root Mean Squared Propagation and
obtain the final decision values. The majority voting is mainly included descent Adamax
to identify the disease accurately and to ignore the classifier that pro­ Dropout rate – 0.5
vides erroneous output.

Fig. 2. Outline of the MOMHTS model.

6
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

The parameters used for optimization are the number of nodes, Table 3
epochs, batch size, hidden layer, and dropout. The optimal parameters Integration of different rules.
are found by varying the values from low to high. The optimal value Index Rule applied
identified where sequentially applied to every hyperparameter. The
1 Majority Voting
random weight, cross-entropy, and activation function were set by 2 Majority Voting + 1
default. A random uniform interval is used for random weight initiali­ 3 Majority Voting + 2
zation for the first layer in the range [-0.05, 0.05]. In the second layer, 4 Majority Voting + 1,2
the Xabier uniform initialize is used which draws random samples in the
range [-limit, limit]. The bias value is initialized as zero and the rectified
linear unit (ReLU) is the activation function used in the hidden layer. ∑
m

The ReLU is a non-linear function and the softmax function is the acti­ Overall Service cost = Pj (7)
vation used in the output layer.
j=1

In the above equation, Pj mainly represents the service cost applied

3.5.1. Decision-making algorithm by the cloud to the consumer to finish their task in a sequential manner.
To decide on the MLP and RF models we are using the majority rule The MOMHTS algorithm is used to optimize the overall service cost by
for decision making and based on the results the random forest achieved allocating the optimal budget and number of cloud service providers.
higher classification results but with a high false-positive rate for the Since the increase in cloud providers increases the communication cost,
normal samples. The false-positive rate is minimized via the additional the number of cloud service providers selected is also optimized.
voting technique presented as shown in Table 2. The results obtained by
combining the two rules are presented in Table 3. The hybrid classifier 4. Experimental results and analysis
used in this paper offers a low false positive and high true positive rate.
The MOMHTS algorithm-optimized Deep learning classifier either offers The experiments were conducted on an Inspiron 24–5000 desktop
a high true positive or a low false-positive rate. The rule-based majority equipped with a Windows 10 OS and 11th Generation Intel® Core™ i3-
vote technique overcomes the higher false positive issues. 1115G4 Processor (6 MB Cache, up to 4.1 GHz). The proposed Hybrid
The rule-based majority model follows both rule policies and also model was implemented in Matlab. A detailed description of the ex­
applies a majority rule. Even though every single optimizer is optimized periments conducted, the dataset used, etc is provided in this section.
means it does not specify that the ensemble of classifiers is optimized.
Our model offers the best performance when applied both rules shown in 4.1. Dataset description
Table 2. When majority voting is applied, the proposed model offers the
highest true positive results for more than 90% of cases. The added rules The COVID19 lung CT scan dataset (LuisBlanche, 2020) and Chest X-
are shown in Table 3 return low false-positive results. Based on our re­ ray images (Pneumonia) (Mooney, 2018) are the datasets used in our
sults achieved the random forest classifier offered higher classification proposed methodology. The CT scan images in the COVID19 lung CT
efficiency for the abnormal samples and the MLP classifier offered scan dataset are obtained from different COVID19 related articles such
higher classification efficiency for the normal samples. The optimized as JAMA, Lancet, bioRxiv, medRxiv, etc. Based on the figure captions,
decision-making algorithm is formulated as shown in algorithm-2. The the abnormalities are identified. A total of 349 COVID19 cases and 397
proposed framework is presented in Fig. 3. non-COVID19 cases were present in the dataset. The pneumonia dataset
Algorithm 2: Decision-making algorithm formulated using rules consists of 5,232 chest X-ray images where the 1,349 samples belong to
shown in Table 2. normal classes and the remaining 3,883 samples belong to the abnormal
Rule-based Majority Vote (MLP, Random Forest) classes. In both datasets, 70% of images were used for training and the
Input: CT images from pneumonia and COVID19 datasets remaining 30% of images were used for testing. After no improvement is
Output: Normal/Abnormal (Pneumonia or COVID19)
noticed in the loss and accuracy after 100 epochs, the training process is
If (RF<=0.5 and MLP<=0.5) then
Return Normal automatically stopped. Both datasets used are publicly available
Else COVID19 datasets. Fig. 4 and Fig. 5 present the samples present in both
Return Majority Vote (MLP, Random Forest) the COVID19 lung CT scan dataset and Chest X-ray images (Pneumonia)
dataset respectively. The images obtained were of different sizes and
they were cut down to a resolution of 224 × 224 × 3 before giving it as
3.6. Web service based implementation input for training.

The proposed model is implemented as a web service in the cloud 4.2. Performance evaluation metrics
environment. A web service is mainly a task that is completed by the
cloud and it is based on different non-functional requirements such as The performance of the classifier is mainly evaluated using different
cost, security, etc in this work we are focused on the cost of the web performance metrics such as accuracy, sensitivity, specificity, and F-
service. Each service provider such as Amazon EC2 and Google appli­ score. The main aim of the different classifiers is to minimize the number
cations has its own processing requirements and it changes as per the of false-positive and false-negative results during classification. The
usage of the consumer per hour, per GB, per MB, etc. The service cost of diagnosis of a novel disease via the CT scan and X-rays is very important
the website is represented as shown below: to society. A brief description of the different performance metrics used
is presented in this section.

Sensitivity (S1): The true positive rate of the classifier is mainly

Table 2 determined by the sensitivity metric which mainly evaluates the
Decision-making results.
efficiency of the model to predict the true positive results correctly.
Rule Output Specificity (S2): The ability of the classifier to distinguish the true
Number
negative rate is represented as specificity.
1 If Results of Random Forest <=0.5 then output “normal” Precision (P): The positive prediction capability of the model is
2 If Results of Random Forest < 0.5 and MLP <=0.5 then output evaluated using prediction which is also known as the Positive Pre­
dictive Value (PPV).
“normal”

7
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

Fig. 3. Proposed MOMHTS optimized.

Fig. 4. Samples images from the COVID19 lung CT scan dataset (a)-(b) Normal CT scan results, and (c)-(d) Abnormal CT scan results.

Fig. 5. Samples images from the Chest X-ray images (Pneumonia) dataset (a)-(b) Normal results, and (c)-(d) Abnormal results.

F1-Score: A tradeoff between precision and recall is obtained via F1- minimal latency of 0.07 s which is relatively low when compared to the
Score. standard HRFDL algorithm which consumes more than 0.08 s. From the
Accuracy: It mainly identifies the efficiency of the classifier to predict results obtained we can conclude that the proposed model shows
the samples correctly. improved performance in terms of latency.
The proposed MOMHTS algorithm-optimized HRFDL model is
mainly developed to minimize the diagnosis time taken in edge devices.
4.3. Results
The time is computed to classify the images in the COVID19 lung CT scan
dataset. The results are shown in Fig. 7 and based on the results the
The latency of the standard HRFDL algorithm and the optimized
performance of the proposed model is higher than the existing Adaptive
HRFDL algorithm for the disease diagnosis healthcare application is
ABD (Vasconcelos et al., 2020), ETS-DNN (Pustokhina et al., 2020),
presented in Fig. 6. Based on the results we can observe that the latency
PMLA (Akkaoui et al., 2020), MobileNetv2 architecture (Singh & Kole­
value increases with an increase in the number of patients. The HRFDL
kar, 2021), and UbeHealth (Muhammed et al., 2018). Based on the re­
offers minimal latency and the optimized HRFDL algorithm offers an
sults we can confirm that the proposed model is four times faster than
improved latency. The proposed optimized HRFDL algorithm takes a

8
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

the existing techniques.

The results of the proposed optimized HRFDL algorithm are
computed using the F1 measure and the results obtained are plotted in
Fig. 8.
Adaptive ABD (Vasconcelos et al., 2020), ETS-DNN (Pustokhina
et al., 2020), PMLA (Akkaoui et al., 2020), MobileNetv2 architecture
(Singh & Kolekar, 2021), and UbeHealth (Muhammed et al., 2018) are
the different state-of-art techniques taken for comparison. The results
obtained by the proposed model show the model’s efficiency to classify
the normal and abnormal results accurately. An F-score of 99% is
accomplished by our proposed methodology. The UbeHealth and
MobileNetV2 architecture obtained the next F-measure values but these
models are not quite effective when it comes to edge device deployment.
The Adaptive ABS and ETS-DNN obtain the lowest F-measure of all.
The specificity values of the proposed MOMHTS algorithm-
optimized HRFDL classifier are compared with the state-of-art classi­
fier and the results obtained are shown in Fig. 9. Based on the results
obtained, the lowest specificity score of 94% is obtained by the Adaptive
ABD model since its efficiency decreases for large datasets. The speci­
Fig. 6. Comparison in terms of latency. ficity score improves as the number of patients increases for our pro­
posed model and our proposed model achieved a specificity score of
99% for a total of 1500 patients.
The overall performance of the proposed model in two datasets in
terms of accuracy and sensitivity is presented in Table 4. The proposed
model is compared with different existing techniques such as Adaptive
ABD (Vasconcelos et al., 2020), ETS-DNN (Pustokhina et al., 2020),
PMLA (Akkaoui et al., 2020), MobileNetv2 architecture (Singh & Kole­
kar, 2021), and UbeHealth (Muhammed et al., 2018) using the COVID19
lung CT scan dataset. The Table 4 achieves the highest accuracy and
sensitivity values when compared to the existing techniques. The results
shown in Table 5 are self-explanatory and it mainly indicates that the
proposed model is capable of identifying the normal and abnormal
classes with higher accuracy and sensitivity values. The accuracy of the
existing techniques is often low due to the lack of parameter
optimization.
The Receiver Operating Characteristic (ROC) curve obtained for the
Chest X-ray images and COVID19 lung X-ray scan is presented in Fig. 10
(a) and (b). The true positive rate is plotted on the X-axis and the false
positive rate is plotted on the Y-axis. A reliable classifier is capable of
minimizing the false positive rate and maximizing the true positive rate
as maximum as possible. For the two datasets, the ROC curve of the
optimized proposed model is near the top left corner which shows sig­
Fig. 7. Comparative analysis using time complexity.
nificant performance.
The storage issues related to the edge devices are optimized even

Fig. 8. Comparative analysis using F-measure.

Fig. 9. Comparative analysis using Specificity.

9
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

Table 4 results. The UbeHealth classifier has an increased application size of 30

Performance evaluation using accuracy and sensitivity. MB which is relatively higher than the existing techniques. The results
Technique Accuracy Sensitivity obtained by the MOMHTS algorithm-optimized HRFDL classifier are
(%) (%) presented in Table 5 along with the input image samples.
Adaptive ABD (Vasconcelos et al., 2020) 92 94 The confusion matrix mainly evaluates the performance of a classi­
ETS-DNN (Pustokhina et al., 2020) 94 95 fier on a dataset whose results are already present beforehand. Tables 6
MobileNetv2 architecture (Singh & Kolekar, 2021) 96.40 96 and 7 present the confusion matrix results obtained for the COVID19
UbeHealth (Muhammed et al., 2018) 89 86 lung CT scan and Chest X-ray images (Pneumonia) dataset.
Proposed MOMHTS algorithm-optimized HRFDL 99 99
classifier
From a total of 349 COVID19 samples, the proposed model is capable
of classifying 346 images correctly and from a total of 397 non-COVID19
samples, the proposed model is capable of classifying the 393 images
correctly. The proposed model offers an accuracy of 99% in the
Table 5
COVID19 lung CT scan dataset. In the Chest X-ray images (Pneumonia)
Prediction results obtained for the proposed classifier for different training
samples.
dataset, the proposed model identifies a total of 3844 pneumonia cases
correctly from the total of 3805 cases. This model also offers an accuracy
Input images Actual class Results obtained Predicted
of 99% for the disease classification.
by our proposed class
model
4.4. Computational complexity analysis
COVID-19 0.004541 COVID-19
(Absence) (Absence)
Energy consumption (EC) mainly measures the amount of energy
spent in transferring the data from the source to the destination. Fig. 12
presents the computational complexity analysis of the proposed model
in terms of energy consumption and it is mainly measured for the pro­
COVID-19 0.99878 COVID-19 posed model with and without optimization. Based on the graph, we can
(Presence) (Presence) conclude that the proposed model’s energy consumption increases
without optimization but decreases with optimization.
The computational time and memory consumed by the proposed
model for storing data of different sizes are presented in Table 8. As per
Table 8, we can notice that the proposed model takes a time of 175 s to
COVID-19 0.99454 COVID-19
store data of 200 MB and a memory capacity of 1,304,578 bytes. The
(Presence) (Presence)
memory and computational time details for the different data size of
20–300 MB is presented in Table 8.

5. Discussion
COVID-19 0.000278 COVID-19
(Absence) (Absence) The proposed model is formed of different components which
minimize the COVID19 outbreak and serves as a reliable model for
COVID19 prevention. Each user in the network has an automatically
generated user ID and each user ID is associated with a number that
defines the severity of the infection. Based on the user ID and their
Pneumonia 0.0005447 Pneumonia
(Absence) (Absence)
current location derived from the devices, the infected users who spread
the disease can be recognized. The relationship that exists between the
unaffected and infected individuals is identified via a tool known as the
Gephi 0.9.1. The COVID19 disease is mainly spread via the air droplets
emitted during coughing and uninfected people can also get infected
when inhaling the cough droplets. Based on the classification results
obtained from the proposed model, the location of the individual is
Pneumonia 0.99458 Pneumonia
(Presence) (Presence)
continuously monitored to identify whether they are in proximity to an
unaffected individual. Government organizations can reduce the disease
progression by closely monitoring the infected and uninfected person
samples classified by the MOMHTS optimized HRFDL classifier. The
high proximity between the affected and unaffected individuals is
monitored using the radio waves generated by the Radio Frequency
Pneumonia 0.99754 Pneumonia Identification (RFID) tags. These tags are placed in the chest area of the
(Presence) (Presence) infected individuals in a certain location and it is activated when any
person approaches the infected. Each user can identify the RFID tag
worn by others via a mobile application mainly designed for this pur­
pose. When an uninfected individual comes near the infected individual,
the application sends an alert to maintain a 1–2 m distance. The contact
information is transferred to the cloud and the user is continuously
monitored via a 5G/4G internet connection. In this way, the contact
further using the flat buffer format and in this way, the proposed model
between the uninfected users with the infected persons is prevented as
can even handle mobile devices with restricted computational capabil­
maximum as possible. The user proximity is identified at regular in­
ities. Fig. 11 provides the application size of different classifiers in
tervals and a 25-second window is used for this purpose.
mobile devices. The findings provided in Fig. 11 illustrate that even with
an application size of 8.3 MB, the proposed model provides considerable

10
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

Fig. 10. ROC curve results. (a) Results for the Chest X-ray images (Pneumonia) dataset and (b) Results for the COVID19 lung CT scan dataset.

number of false positives predicted by the model. The proposed model

yields fast detection and efficient classification of COVID-19 and pneu­
monia in patients living in rural areas. The advantages offered by this
model are cost-efficiency, higher accuracy, and minimized execution
time which is very important when deploying the applications in an IoT-
based edge environment. In this way, it allows the deep learning clas­
sifier to explore even further. The proposed model’s majority voting
system considerably enhances the outcome of the random forest and
MLP classifier. After the disease classes (COVID-19/Pneumonia) are
identified the results obtained are transformed into the cloud server
which can be later utilized by the healthcare professionals and gov­
ernment workers to aid in early diagnosis and notify the people in the
surrounding location to be aware of this disease. The proposed work is
simulated in real-time by conducting a series of experiments and the
results are evaluated using different performance metrics such as
sensitivity, specificity, accuracy, ROC curve, F1-score, confusion matrix,
etc. The proposed methodology offers an accuracy of 99% for both the
Fig. 11. Comparison in terms of application sizes.
COVID19 lung CT scan dataset and Chest X-ray images (Pneumonia)
datasets. The size of the developed application is 8 MB which is rela­
6. Conclusion
tively lower than other models and a latency value of 0.076 is obtained
for a total patient record of 1500. In the future, we plan to optimize the
This paper proposes a MOMHTS algorithm-optimized HRFDL clas­
security of the web services and also identify the different types of
sifier for edge computing IoT-enabled medical environment. Initially,
pneumonia.
the IoT devices such as CT scanners and X-ray machines capture the
chest area and identify if any abnormality is present or not. The pro­
Ethical statements
posed model is optimized using the MOMHTS algorithm which tunes the
hyperparameters associated with the DNN algorithm and minimizes the
Funding: Not applicable.

Table 6
Confusion matrix for the COVID19 lung CT scan dataset.

11
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

Table 7
Confusion matrix for the Chest X-ray images (Pneumonia) dataset.

Authors contributions

MH agreed on the content of the study. MH collected all the data for
analysis. MH agreed on the methodology. MH completed the analysis
based on agreed steps. Results and conclusions are discussed and written
together. The author read and approved the final manuscript.

Declaration of Competing Interest

The authors declare that they have no known competing financial

interests or personal relationships that could have appeared to influence
the work reported in this paper.

Data availability

The authors do not have permission to share data.

References

Fig. 12. Computational complexity analysis using energy consumption for the Akkaoui, R., Hei, X., & Cheng, W. (2020). Edgemedichain: A hybrid edge blockchain-
proposed model with and without optimization. based framework for health data exchange. IEEE Access, 8, 113467–113486. https://
doi.org/10.1109/ACCESS.2020.3003575
Bhatia, M., Manocha, A., Ahanger, T. A., & Alqahtani, A. (2022). Artificial intelligence-
inspired comprehensive framework for Covid-19 outbreak control. Artificial
Table 8 Intelligence in Medicine, 127, Article 102288. https://doi.org/10.1016/j.
Computational Time and memory analysis. artmed.2022.102288
Carpenter, C. R., Mudd, P. A., West, C. P., Wilber, E., & Wilber, S. T. (2020). Diagnosing
Data Size (MB) Time (Seconds) Memory (Bytes)
COVID-19 in the emergency department: A scoping review of clinical examinations,
50 49 1,102,568 laboratory tests, imaging accuracy, and biases. Academic Emergency Medicine, 27(8),
100 57 1,201,456 653–670. https://doi.org/10.1111/acem.14048
150 110 1,295,874 Das, N. N., Kumar, N., Kaur, M., Kumar, V., & Singh, D. (2020). Automated deep transfer
200 175 1,304,578 learning-based approach for detection of COVID-19 infection in chest X-rays. Irbm.
https://doi.org/10.1016/j.irbm.2020.07.001
250 190 1,352,464
Ding, S., Li, L., Li, Z., Wang, H., & Zhang, Y. (2019). Smart electronic gastroscope system
300 230 1,459,875
using a cloud–edge collaborative framework. Future Generation Computer Systems,
100, 395–407. https://doi.org/10.1016/j.future.2019.04.031
Ford, L., Lee, C., Pray, I. W., Cole, D., Bigouette, J. P., Abedi, G. R., … Kirby, M. K.
Compliance with ethical standards (2021). Epidemiologic Characteristics Associated With Severe Acute Respiratory
Syndrome Coronavirus 2 (SARS-CoV-2) Antigen-Based Test Results, Real-Time
Reverse Transcription Polymerase Chain Reaction (rRT-PCR) Cycle Threshold
Human and animal rights Values, Subgenomic RNA, and Viral Culture Results From University Testing. Clinical
Infectious Diseases, 73(6), e1348–e1355. https://doi.org/10.1093/cid/ciab303
This article does not contain any studies with human or animal Ghosh, S. K., & Ghosh, A. (2022). ENResNet: A novel residual neural network for chest X-
ray enhancement based COVID-19 detection. Biomedical Signal Processing and
subjects performed by any of the authors.
Control, 72, Article 103286. https://doi.org/10.1016/j.bspc.2021.103286
Hemalatha, M., & Rukmanidevi, S. (2017). A study on IPV6 prefix matching through
Informed consent DNA computing. Journal of Computational and Theoretical Nano science, 14(8),
3867–3873.
Hemalatha, M., Rukmanidevi, S., & Shanker, N. R. (2021). Searching time operation
Informed consent was obtained from all individual participants reduced IPV6 matching through dynamic DNA routing table for less memory and fast
included in the study. IP processing. In , 25. Soft computing (pp. 3455–3468). Springer.
Consent to participate: Not applicable. Hung, M., Lauren, E., Hon, E. S., Birmingham, W. C., Xu, J., Su, S., … Lipsky, M. S.
(2020). Social network analysis of COVID-19 sentiments: Application of artificial
Consent for publication: Not applicable. intelligence. Journal of Medical Internet Research, 22(8), Article e22590. https://doi.
org/10.2196/22590
Availability of data and material Jamshidi, M., Lalbakhsh, A., Talla, J., Peroutka, Z., Hadjilooei, F., Lalbakhsh, P., …
Sabet, A. (2020). Artificial intelligence and COVID-19: Deep learning approaches for
diagnosis and treatment. IEEE Access, 8, 109581–109595. https://doi.org/10.1109/
Data sharing is not applicable to this article as no new data were ACCESS.2020.3001973
created or analyzed in this study. Jose, J., Gautam, N., Tiwari, M., Tiwari, T., Suresh, A., Sundararaj, V., & Rejeesh, M. R.
(2021). An image quality enhancement scheme employing adolescent identity search
Code availability: Not applicable. algorithm in the NSST domain for multimodal medical image fusion. Biomedical
Signal Processing and Control, 66, 102480.
Khemasuwan, D., Sorensen, J. S., & Colt, H. G. (2020). Artificial intelligence in
pulmonary medicine: Computer vision, predictive model and COVID-19. European
Respiratory Review, 29(157). https://doi.org/10.1183/16000617.0181-2020

12
M. Hemalatha Expert Systems With Applications 210 (2022) 118227

Kumar, S., Tejani, G. G., Pholdee, N., & Bureerat, S. (2021). Multi-objective modified Rodríguez-Rodríguez, I., Rodríguez, J. V., Shirvanizadeh, N., Ortiz, A., & Pardo-
heat transfer search for truss optimization. Engineering with Computers, 37(4), Quiles, D. J. (2021). Applications of artificial intelligence, machine learning, big
3439–3454. https://link.springer.com/article/10.1007/s00366-020-01010-1. data and the internet of things to the COVID-19 pandemic: A scientometric review
Lane, N. D., & Georgiev, P. (2015). February). Can deep learning revolutionize mobile using text mining. International Journal of Environmental Research and Public Health,
sensing?. In In Proceedings of the 16th International Workshop on Mobile Computing 18(16), 8578. https://doi.org/10.3390/ijerph18168578
Systems and Applications (pp. 117–122). https://doi.org/10.1145/2699343.2699349 Shone, N., Ngoc, T. N., Phai, V. D., & Shi, Q. (2018). A deep learning approach to
Li, X., Huang, X., Li, C., Yu, R., & Shu, L. (2019). EdgeCare: Leveraging edge computing network intrusion detection. IEEE Transactions on Emerging Topics in Computational
for collaborative data management in mobile healthcare systems. IEEE Access, 7, Intelligence, 2(1), 41–50. https://doi.org/10.1109/TETCI.2017.2772792
22011–22025. https://doi.org/10.1109/ACCESS.2019.2898265 Singh, V. K., & Kolekar, M. H. (2021). Deep learning empowered COVID-19 diagnosis
LuisBlanche, (2020). COVID-19 Lung CT Scans [WWW Document]. Kaggle, URL https:// using chest CT scan images for collaborative edge-cloud computing platform.
www.kaggle.com/luisblanche/covidct (accessed 12.13.21). Multimedia Tools and Applications, 1–28. https://link.springer.com/article/10.1007
Mach, P., & Becvar, Z. (2017). Mobile edge computing: A survey on architecture and /s11042-021-11158-7.
computation offloading. IEEE Communications Surveys & Tutorials, 19(3), 1628–1656. Sohrabi, C., Alsafi, Z., O’neill, N., Khan, M., Kerwan, A., Al-Jabir, A., Iosifidis, C., &
https://doi.org/10.1109/COMST.2017.2682318 Agha, R. (2020). World Health Organization declares global emergency: A review of
Mohammedqasim, H., & Ata, O. (2022). Real-time data of COVID-19 detection with IoT the 2019 novel coronavirus (COVID-19). International journal of surgery, 76, 71–76.
sensor tracking using artificial neural network. Computers and Electrical Engineering, https://doi.org/10.1016/j.ijsu.2020.02.034
100, Article 107971. https://doi.org/10.1016/j.compeleceng.2022.107971 Sundararaj, V. (2016). An efficient threshold prediction scheme for wavelet based ECG
Mooney, P. (2018, March 24). Chest X-Ray Images (Pneumonia). Kaggle. https://www. signal noise reduction using variable step size firefly algorithm. International Journal
kaggle.com/paultimothymooney/chest-xray-pneumonia. of Intelligent Systems, 9(3), 117–126.
Muhammed, T., Mehmood, R., Albeshri, A., & Katib, I. (2018). UbeHealth: A Vasconcelos, F. F., Sarmento, R. M., Reboucas Filho, P. P., & de Albuquerque, V. H. C.
personalized ubiquitous cloud and edge-enabled networked healthcare system for (2020). Artificial intelligence techniques empowered edge-cloud architecture for
smart cities. IEEE Access, 6, 32258–32285. https://doi.org/10.1109/ brain CT image analysis. Engineering Applications of Artificial Intelligence, 91, Article
ACCESS.2018.2846609 103585. https://doi.org/10.1016/j.engappai.2020.103585
Ortiz, G., Zouai, M., Kazar, O., Garcia-de-Prado, A., & Boubeta-Puig, J. (2022). Xu, J., Chen, L., & Ren, S. (2017). Online learning for offloading and autoscaling in
Atmosphere: Context and situational-aware collaborative IoT architecture for edge- energy harvesting mobile edge computing. IEEE Transactions on Cognitive
fog-cloud computing. Computer Standards & Interfaces, 79, Article 103550. https:// Communications and Networking, 3(3), 361–373. https://doi.org/10.1109/
doi.org/10.1016/j.csi.2021.103550 TCCN.2017.2725277
Pustokhina, I. V., Pustokhin, D. A., Gupta, D., Khanna, A., Shankar, K., & Nguyen, G. N. Yoo, S., Kim, S., Kim, S., & Kang, B. B. (2021). AI-HydRa: Advanced hybrid approach
(2020). An effective training scheme for deep neural network in edge computing using random forest and deep learning for malware classification. Information
enabled Internet of medical things (IoMT) systems. IEEE Access, 8, 107112–107123. Sciences, 546, 420–435. https://doi.org/10.1016/j.ins.2020.08.082
https://doi.org/10.1109/ACCESS.2020.3000322

13