IoT-Enabled Framework for Early Detection and Prediction of COVID-19 Suspects by Leveraging Machine Learning in Cloud.

COVID-19 is the repugnant but the most searched word since its outbreak in November 2019 across the globe. The world has to battle with it until an effective solution is developed. Due to the advancement in mobile and sensor technology, it is possible to come up with Internet of things-based healthcare systems. These novel healthcare systems can be proactive and preventive rather than traditional reactive healthcare systems. This article proposes a real-time IoT-enabled framework for the detection and prediction of COVID-19 suspects in early stages, by collecting symptomatic data and analyzing the nature of the virus in a better manner. The framework computes the presence of COVID-19 virus by mining the health parameters collected in real time from sensors and other IoT devices. The framework is comprised of four main components: user system or data collection center, data analytic center, diagnostic system, and cloud system. To point out and detect the COVID-19 suspected in real time, this work proposes the five machine learning techniques, namely support vector machine (SVM), decision tree, naïve Bayes, logistic regression, and neural network. In our proposed framework, the real and primary dataset collected from SKIMS, Srinagar, is used to validate our work. The experiment on the primary dataset was conducted using different machine learning techniques on selected symptoms. The efficiency of algorithms is calculated by computing the results of performance metrics such as accuracy, precision, recall, F1 score, root-mean-square error, and area under the curve score. The employed machine learning techniques have shown the accuracy of above 95% on the primary symptomatic data. Based on the experiment conducted, the proposed framework would be effective in the early identification and prediction of COVID-19 suspect realizing the nature of the disease in better way.

1. Introduction

COVID-19 is a repugnant word across the globe since its breakout from Wuhan City of China in November 2019. COVID-19, the name given by the World Health Organization (WHO), initially erupted as an epidemic, but later turned into a deadly pandemic [1]. In November 2021, the figures of COVID-19-confirmed cases exceeded 257.46 million with 3.7% mortality rate. COVID-19 spread the threat across the globe, as of now it took away around 5.15 million lives. COVID-19 comes under the family of Coronaviridae, which causes illness from common cold to more severe diseases. In 2012, Saudi Arabia was epicenter for MERS-CoV with 35% fatality rate [2]. In 2003, Southern China reported the SARS-CoV, which is also from a family of the same virus. Later, both SARS-CoV and MERS-CoV spread across the globe [3]. COVID-19 from its very inception in November 2019 changed its physical and chemical properties. The novel strains of COVID-19 are more vulnerable and transferable with high risk of infection [4]. WHO proclaimed the new COVID-19 pandemic on March 11, 2020. To stop the spread of COVID-19, most of the countries across the globe have shut down all the traffic including air, railways, and markets. Many countries have also imposed restrictions or locked down the cities.

The virus has wreaked havoc on the whole food chain, revealing its fragility. Due to frontier closures, business circumscriptions, and incarceration measures, the general small-scale businesses, street vendors, vegetable growers, and daily wagers were unable to access their local selling places, including obtaining inputs and selling their goods, disrupting national and global food supply networks, and restricting access to nutritious, safe, and diverse meals [5]. From research perspective, COVID-19 is the most searched term on Internet in 2020. A lot of research related to COVID-19 is currently going on throughout the globe [6]. Medical professionals are trying to come up with an antidote that can prevent corona infection. From the perspective of Internet of things, a vast research is being conducted on the impact of IoT technology to tackle the COVID-19 epidemic [7]. Computer scientists on the other hand are trying to develop models that can detect and prevent the infection. The traditional healthcare system is not sufficient enough to handle the current global prevalent situation. Presently, the only way to avoid COVID-19 infection is to follow the SOPs and get vaccinated with immunity boosters. The advancement and increase in mobile technology such as sensors, smart devices, and other wearables mingled in the healthcare system greatly impact our daily lives [8]. Nowadays, IoT is mingled in every field, with the ability to communicate from anywhere, anytime [9] round the clock. New and advanced powerful devices for monitoring individuals' health came due to IoT [10]. IoT is the integration of physical devices with communication technologies capable of connecting through the Internet. The real-time health parameters are taken from deployed sensors to provide the current status of patients [11]. In the current era, mobile phones have inbuilt onboard sensors that can capture the real-time parameters of patients. The various security mechanisms are employed in sending and receiving data from these smart applications [12]. Smartphones can be used as input devices such as sensing, storing, and computing the results [13]. By the use of technology, it is possible to detect the COVID-19 suspects in early stages to eliminate the spread of infection. Tracking and quarantining of COVID-19-positive and COVID-19-suspected cases can be tracked and monitored with the help of onboard mobile phone sensors [14] and by the wireless sensor technology [15] (WSN). Integration of IoT with other potential technologies such as machine learning (ML) and artificial intelligence (AI) can revolutionize the healthcare sector in near future [16]. As a result, in the face of the pandemic, artificial intelligence (AI) and machine learning (ML) created new potential options for successful therapy. AI and machine learning can be used in the discovery of new drugs, the development of accurate diagnostic processes, and the prediction of disease vulnerabilities. These potential areas are strongly reliant on real-time patient monitoring and information syncing, and the IoT plays a notable role in both of these areas [17, 18]. The future predictions can be made using AI and ML in IoT-based systems for predicting the upcoming infection of coronavirus [19]. The IoT can be used as a data source, and ML is used for data analytics to better further analyze the COVID-19 [20] to get better insights. With the help of IoT, a centralized information system can be created where all activities are stored electronically and can be accessed anywhere and anytime [21]. A vast number of people die because of lack or incorrect and inappropriate knowledge about their health. The use of IoT technology can quickly notify individuals' health parameters through the deployed or wearable sensors [22]. The IoT technology can watch and capture the routine activities of an individual and can generate the necessary alerts if there is any critical health issue [23].

1.1. Motivation and Contribution

COVID-19 has taken millions of lives since its outbreak started in Wuhan, a city in China, from the month of November 2019. A lot of research is going on across the globe to combat this pandemic, but the strategies and procedures for analyzing and predicting the virus are still in its infancy. As the pandemic spread around the globe, healthcare systems collapsed due to the unavailability of smart diagnostic systems. Due to the fast transmission of COVID-19 from person to person, an IoT-based system will help in predicting the onset of infection in a real time, thus in turn help in the prevention of this deadly disease. Healthcare system across the globe is poor due to the lack of integration of technology. The IoT can help the healthcare system to automate many sectors to eliminate the errors made by humans. On the other hand, machine learning can be used for analysis purposes to get better insights and understand the nature of disease. The integration of both the technologies such as IoT, ML, and AI can revolutionize the modern healthcare system. By incorporating machine learning in the domain of health care, most of the things can be achieved such as maintaining accurate data, personalized healthcare facilities, and predictive analytics. IoT can mainly be used as for sensing the environment and actuating accordingly, but machine learning is for high-end analytics. The proposed framework based upon the machine learning and IoT will act in a proactive and preventive manner rather than in reactive manner as used in traditional approaches of prevention.

This article proposes a layered architecture and early detection and monitoring system of COVID-19 suspects. Using IoT devices, a real-time symptomatic data are collected to identify COVID-19 potential cases. By deploying IoT-based sensors, there are mainly three potential advantages: firstly, continuous monitoring anytime and anywhere. Second, frequent symptomatic parameters are collected on regular basis. Thirdly, the regular symptomatic data are collected in a particular time frame. To detect the COVID-19 suspect at an early phase, a set of parameters (symptoms) are required for effective results, which is impossible in a single visit to the clinic. To overcome the cons of the traditional healthcare system, a novel COVID-19 detection and prediction and monitoring system in early stage are proposed. The proposed framework contributes by 1. early detection of COVID-19 suspects, 2. analyzing the symptomatic collected data using machine learning techniques, 3. disease diagnosis (whether COVID-19-positive or COVID-19-negative), and 4. maintaining health record of patients for future use. The main aim of the proposed system is to eliminate the expansion of the coronavirus infection and detect the COVID-19 in the early phase, and the disease can be better understood from the collected data by further analysis.

Lastly, our proposed framework has been tested through a novel dataset collected from SKIMS, Srinagar. Distinct ML algorithms have been employed on the dataset to validate the system. Our system has achieved above 95% accuracy. The proposed system is cross-validated using various performance matrices such as accuracy, precision, recall, F1 score, root-mean-square error, and area under the curve.

The proposed framework comprised of four main components: 1. user system: in which sensors are collecting real-time symptomatic data, 2. data analytic center: various machine learning algorithms are employed to collect data, 3. diagnostic system: healthcare (physicians) experts check the calculated parameters, and 4. cloud system. The aim of the framework is to eliminate the death rate by early detection and eliminate the spread of coronavirus infection. The paper uses five (5) machine learning algorithms namely SVM, decision tree, naïve Bayes, logistic regression, and neural network. The paper tests the proposed framework by experimenting the ML techniques on novel dataset. The experimental results have shown that these algorithms have achieved above 95% accuracy.

The rest of the paper is organized as follows: Section 2 elaborates the detailed relevant literature. Section 3 gives the detailed insights of the proposed system. Section 4 discusses the experimental setup. Section 5 provides the detection and prediction model for potential COVID-19-suspected cases by employing machine learning techniques. Section 6 is the results and discussion of the proposed work. Lastly, Section 7 concludes the work.

2. Related Work

AI and machine learning have opened doors for large array of applications in the medical industry, including statistical data prediction and classification [24]. BlueDot Toronto, for instance, established the first risk-based technique for recognizing the SARS-CoV-2 epidemic, which was developed by IoT 2020 by infectious disease professionals to investigate new solutions for mitigating the initial SARS pandemic. BlueDot's previous SARS research was utilized to include advanced technologies in this impressive demonstration of AI and ML in forecasting illness outbreaks [25]. In [26], the authors have developed machine learning-based framework for diabetes prediction and named it as intelligent diabetes mellitus prediction framework (IDMPF). The authors proposed three machine learning techniques to predict diabetes that are as support vector machine, random forest, and decision tree. They have achieved an accuracy of 83% with low root-mean-square error [27]. In this article, the authors have made the step-by-step review of the artificial intelligence in healthcare domain. The AI is comprised of machine learning and deep learning for prearranged datasets, whereas text mining and natural language processing are for unarranged datasets. The authors have highlighted the challenges and research opportunities by integrating AI in healthcare sector. The authors have discussed in a long the technologies that can combat the pandemic [28]. The authors in this paper have developed machine learning-based efficient automatic disease model based on android application. The model has been tested on three different diseases such as COVID-19, diabetes, and cardiovascular disease. The authors have used logistic regression algorithm for prediction and have comparative analysis. Industry 4.0 has revolutionized the world with the advancements in ICTs in easing human lives [29]. Internet of things (IoT) is one of the main components of Industry 4.0 that has changed the way of thinking. IoT is internetwork of physical objects embedded with sensors, communication technologies, processing abilities, and other technologies [30]. COVID-19 is influenza-type disease, which causes the infection in respiratory system with symptoms such as fever, cough, runny nose, and breathlessness. It spread faster from person to person by coming in contact, so predicting the spread of infection is challenging. The authors proposed a model to diagnose the COVID-19 infection. Three types of techniques have been tested on the Kaggle dataset such as linear regression, multilayer perceptron, and vector autoregression. Reference [31] has made the systematic review of healthcare technologies such as IoT, big data, and cloud computing with respect to Industry 4.0. A lot of literature studies have been surveyed to discuss the main technologies and application of IoT in healthcare. Riazul Islam et al. presents a comprehensive survey of IoT in health care. The authors have reviewed a state-of-art literature about technologies, architectures, and applications of Internet of things in healthcare. Security models have also been discussed and presented as a security model for IoT healthcare [33]. The authors have discussed the possibilities of integration of artificial intelligence with the wireless technologies to combat in pandemic situations. In this study, the authors have proposed an ensemble machine learning model, i.e., random forest algorithm to predict the severity of COVID-19 patients taking under several parameters. The proposed model has performed well in almost all performance measures such as accuracy, F1 score, precision, and recall. The proposed algorithm is compared with other algorithms such as SVM, decision tree, logistic regression, and naïve Bayes. The proposed algorithm surpasses all the algorithms in terms of performance measures. The proposed algorithm achieved an accuracy of 94%, F1 score of 0.86, precision of 1.0, and recall of 0.75. Reference [34] has proposed a cloud-IoT-based framework for student health monitoring. The proposed framework predicts the level of disease by measuring temporal measurements by collecting data from medical IoT devices. The authors of their study have used a dataset of 182 students to test the proposed framework. Various machine learning algorithms have been applied and validated using k-cross-validation methods [35]. A lot of literature studies have been reviewed, and the potential application of IoT has been discussed. The article came under COVID-19 solutions with current applications of IoT such as smart transportation, ambient living, and smart city [36]. The authors have remote asthma patient monitoring system based on IoT technology. The monitoring system is comprised of sensors, android application, and website. The sensors are collecting vital parameters such as blood pressure and glucose level, and the model was tested on some patients [37]. Internet of things is a disruptive technology that can renovate the healthcare system. The authors have made good efforts on how IoT can be implemented to tackle COVID-19. They have given a brief insight of various IoT technologies that can be used during the COVID-19 pandemic [38]. The vaccine is developed by different companies such as BioNTech, Pfizer, and Moderna in India. The vaccines have different effects on the people based on demographic factors. The researchers in this study have analyzed the data collected from vaccine companies to predict the viable persons based on some variables. The variables are age, gender, and others such as state of living. Based on these parameters, the researchers are predicting the best manufacturer for that person. The researchers have employed different machine learning algorithms such as logistic regression, decision tree, random forest, and AdaBoost. The performance measures of these algorithms are contrasted in terms of accuracy. The AdaBoost has surpassed all and achieved 98.1% accuracy, random forest has 97.8% of accuracy, and decision tree and logistic regression are at the same place with 97.3% of accuracy [39]. IoT can be used to eliminate spread of COVID-19. This technology helps in providing more user satisfaction by properly monitoring COVID-19 patients. The authors have explored twelve potential areas of IoT to combat COVID-19. IoT is helpful in identifying the symptoms of COVID-19 suspects to provide better treatments [40]. A cloud-IoT-based platform for disease diagnosis has been presented by the authors. The proposed paradigm forecasts the severity of a potential disease. The suggested framework has been tested using the UCI dataset. To estimate the severity of disease, various machine learning classification techniques were employed to the obtained data. The accuracy, sensitivity, specificity, and F measure were used to calculate the findings [41]. According to the report, the employment of robotics, IoT, and other related innovations has expanded rapidly as a result of the rise of Industry 4.0. The Internet of things (IoT) is a strong solution for a wide range of real-time issues, thanks to the sensors that make it possible. IoT acts as a crucial enabler for Industry 4.0 through device connectivity, enabling better management, customized service, and efficient operation [42]. The authors have developed a cloud-based disease forecast and diagnostic system using various algorithms. The input is collected from IoT wearable devices and then transferred these signals to a server using Internet. The authors first create the feature set from collected data using the proposed hybrid decision-making approach. The authors have also proposed IoT-based framework with flow of instructions in their research paper. Reference [43] discussed a lot of AI techniques used to tackle COVID-19. Medical image processing, data analytics, text mining, and natural language processing are some areas that are discussed in this article. A detailed overview of open COVID-19 datasets is publicly accessible for research purposes. The authors have also discussed the future directions of potential areas of AI that can fight against COVID-19. Siriwardhana et al. [44] present the power of 5G and IoT to combat COVID-19. The authors have discussed several use cases of these technologies that can provide innovative solutions such as contact tracing, telehealth, and education [45]. The present situation has opened the doors for creating new avenues in our daily lives. The authors have a lot of literature studies about the COVID-19 solutions and have identified seven potential applications useful during pandemic [46]. The authors have reviewed the literature on machine learning techniques and IoT in combating COVID-19 pandemic. The medical methods are time-consuming and costly such as RT-PCR and CT (chest) and are putting burden to technologists and radiologists. AI is a potential technology that can eliminate the cost and time to combat the COVID-19 pandemic. The authors have also discussed the challenges of IoT and ML in fighting the COVID-19 pandemic [47]. COVID-19 has affected almost each and every field. In this article, the authors have discussed the literature on IoT and ML to prevent and diagnose the COVID-19 pandemic. The authors have explained the various machine learning techniques for classification and clustering for COVID-19. Reference [48] has highlighted that as a consequence of the COVID-19 problem, several enterprises have closed, and many manufacturing and small merchants will go out of business. They must deal with a myriad of difficulties, such as cost containment and worker sanitation. Several strategies for coping with the pandemic crisis have been presented, with IR 4.0 playing an important role. Reference [49] has proposed a hybrid model to predict the mortality rate on the India in future. They have used statistical neural network (SNN) and nonlinear autoregressive neural network (NAR-NN)-based models to improve the prediction accuracy. The results are compared with SNN-based models such as probabilistic neural network (PNN), radial basis function neural network (RBFNN), and generalized regression neural network (GRNN). The performance of the models is measured using root-mean-square error (RMSE) and R (correlation coefficient). The hybrid model of PNN and RBFNN performed better than all [50]. The authors have suggested the IoT-based identification and control system in real time. The system identifies the potential cases in early stages and tracks their clinical measures. The proposed framework has five main components: data collection, quarantine center, processing unit, cloud computing, and visualization of data to healthcare professionals. The authors have employed various machine learning techniques to detect COVID-19 suspects [51]. IoT is a vital technology that has the potential to combat during pandemic such as COVID-19. The authors in this paper have proposed a four-layer model to predict potential cases of COVID-19. The model has four components: data acquisition, data aggregation, machine intelligence, and services. The model is validated using voice data [52]. The authors have surveyed a lot of literature studies of IoT technologies used in tracing, tracking, and spread of COVID-19. The authors have highlighted the architectures and also future directions of IoT implementations [10]. The authors in this article have highlighted the applications of IoT that can be used in combating COVID-19. The authors have proposed a real-time identification and monitoring system for COVID-19. The model is divided into four components based on cloud technology: the collection of symptomatic data, health center, data warehouse, and health professionals. The authors have tested the framework using machine learning models, and random forest has shown the best results.

3. Proposed IoT Framework

3.1. Proposed Architecture

This section discusses the IoT-cloud architecture of the proposed system, diagrammatically presented in Figure 1: proposed 3-layer architecture. The proposed layered architecture is based on standard IoT architecture, and it has three layers: sensing layer, analysis layer, and cloud layer. The sensing layer or perception layer is accountable for the collection of symptoms from the suspected persons through various deployed sensors, wearables, and IoT devices. There are various types of electronic digital sensors such as temperature sensor, audio-based sensors, motion-based sensors, heart rate sensor, O2 sensor, and other biosensors such as ECG and EEG. Other information such as travel history and other parameters are collected with the help of applications. The sensing layer sends this collected information to the layer above it called as the analysis layer. The analysis layer is responsible for doing analysis of data received from the sensing layer. Numerous machine learning models are deployed in this layer for getting better insights from data. The prediction of suspected cases is made based on symptoms of a person of whether a suspected is COVID-19-positive or not. The resultant data are then sent to the cloud layer for other services. The third layer of the architecture is the cloud layer, which is responsible for storing the data. Healthcare professionals can then use the stored data for further analysis. The data are used to update machine learning models for deriving more accurate results.

Figure 1

Proposed 3-layer architecture.

3.2. Proposed Framework

This section discusses the proposed IoT-based framework to identify and predict COVID-19 suspects in early stages. This framework is also used to eliminate the further spread of infection and get better insights of the disease for future perspective. Figure 2: a conceptual framework for early detection and prediction of COVID-19 suspect, shows the proposed model of the system. The framework has mainly consisted of three main modules with respect to the proposed three-layer architecture: user system, data analysis system, and cloud system.

Figure 2

A conceptual framework for early detection and prediction of COVID-19 suspect.

User System: the main objective of this module is to sense real-time data with the help of sensors and wearables. The collected symptom data are fever, cough, fatigue, rhinitis, breathlessness, myalgia, oxygen saturation, travel history, blood pressure, etc. There are several sensors such as temperature sensor, O2 sensor, motion sensor, proximity sensor, and inertia sensor. The other relevant parameters are collected from user through applications such as travel history through smartphones history. These sensors are connected with IoT gateway to communicate the sensed data through Internet. Sensors are battery-powered so they are not directly communicating to Internet. The communication technology used by sensors to communicate with gateway is low-powered technology such as BLE, infrared, and Wi-Fi. The gateway uses Wi-Fi, mobile networks, 3G, 4G, 5G, etc., to communicate with the cloud system.

Data Analytic Center: this component is responsible for data analysis and hosting of machine learning algorithms. On the basis of collected symptoms accessed from personal health records of cloud system, prediction is made whether a person is COVID-suspected or not. The results are then generated and updated in cloud accordingly. As it is continuously updating the personal health records, the machine learning models are updating also with the help of new analysis made by data analytic module.

Medical Laboratory and Diagnostic System: this module is comprised of health physicians and medical laboratories. The suspected first are sent for laboratory test (RT-PCR/RAT), and if they are found positive, they are checked by medical physicians for health checkup. The clinical investigations are made based on patient's symptoms received from cloud system. This proposed model can predict and eliminate the further spread of COVID-19-suspected cases.

Cloud System: cloud computing is buzz term for last two decades, in which everything is in logical way in a centralized system known as cloud. On-demand services are provided such as storage, databases, and computing resources in a cloud computing environment. In our case, all types of services such as storage and computing resources are taken from the cloud environment. The data sensed by the sensing layer are communicated via communication networks to cloud for storage purposes, updating personal health records, and communicating with other components.

3.3. Flowchart of Proposed Framework

The flow of framework is described in Figure 3: data flow of proposed framework, and the steps are described as follows:

Figure 3

Data flow of the proposed framework.

The system collects data from sensors and wearables deployed through body area network (BAN). The symptoms such as cough, rhinitis, sore throat, breathlessness, O2 saturation, blood pressure, and other related information through smartphone are collected in real time. The collected data are then sent for analysis.

The uploaded data from step 1 is then analyzed for possible COVID-19 infection. The machine learning models are then applied to the collected data and obtained the results. The machine learning models are continuously updating with the real-time data to derive more accurate results. Further, the results are seen by medical physicians to better understand the disease. The COVID-19-suspected cases are predicted and identified using machine learning models.

If a person is COVID-19-suspected, they will be sent for clinical laboratory test (RAT/RT-PCR) for investigation. If suspected is COVID-19-positive, they will be sent to medical physician for checkup. The confirmed positives can then be secluded, and all other previous contacts will also be isolated to eliminate further spread of infection.

4. Experimental Setup

4.1. Data Collection

COVID-19 was declared a pandemic on March 11, 2020, by the World Health Organization. The disease is new in nature and RNA-based and continuously changes its properties. Due to these unpredictable properties, it is hard to derive any concrete solution. The researchers and hospitals give open access to data regarding the confirmed cases. The unpredictable and unknown nature of the disease made it tough to develop any remedy or medicine. Researchers and academicians are trying to develop a vaccine and a solution that can combat COVID-19. The World Health Organization (WHO) and medical organizations made it possible for everyone to contribute to or provide a solution to the COVID-19 pandemic. Researchers from different domains are trying their best to efficiently solve the pandemic. Since the academic fraternity has no prior experience of a pandemic such as COVID-19, none of the solutions is a holistic working solution. As this has become an open challenge, the ongoing research is available on different websites such as Google Cloud, NIH, COVID-19 Data Repository, and other international and national institutes. The available public datasets are simple metadata or confirmed cases of different countries, by which a concrete solution cannot be drawn. The available datasets do not include all the information about patient's symptoms because of the novelty of the virus. The available data are inadequate and insufficient for the use by machine learning algorithms. This research aims to develop an IoT-cloud-based system that can predict the COVID-19 suspects based on patient symptoms. The actual dataset has been collected from the Sher-I-Kashmir Institute of Medical Sciences (SKIMS), Srinagar, Jammu and Kashmir, India, collaborating with the doctors. The SKIMS is a renowned Medical Institute of Jammu and Kashmir, India. During the pandemic, they have received scores of COVID-19-positive patients for medical facilities. The SKIMS Institute has made a separate temporary COVID-19 department. Before starting our work, a round table meeting was held with a team of doctors to discuss the possible symptoms of COVID-19 patients. The symptoms of the COVID-19 patients were already published on various websites; in particular, the set of primary symptoms given by WHO and CDC on their websites are as follows: fever, cough, fatigue, runny nose, breathlessness, etc. The dataset attributes (symptoms) were finalized after consulting a group of senior doctors from the COVID-19 department of the institute. Finally, the proforma of symptoms has been drafted to collect those from COVID-19 OPD clinic and in-patients. The list of attributes or symptoms is given in Table 1: collected symptoms of patients.

Table 1

Collected symptoms of patients.

Collected symptoms
MRD no.	Date	Age	Sex	Residence	Contact history	Fever	Cough
Rhinitis	Sore throat	Shortness of breath	Myalgia	Fatigue	Loss of appetite	Loss of taste	Vomiting
Nausea	Diarrhea	O2 saturation	Hypertension	Kidney disease	Heart disease	Liver disease	Chest disease
Severity of illness	Headache	Body pain	Anxiety	Travel history	Survival	Test type	Test result

There are some other attributes such as travel history, whether a patient is having any other diseases or not, such as diabetes, kidney, and heart, blood group, hemoglobin, headache, anosmia, pulse, BP, respiratory rate, and temperature. The data of these attributes were either inadequate or insufficient to take them as attribute. Thus, the data preprocessing and feature selection must be performed.

4.2. Preprocessing, Feature Selection, and Normalization

The collected data from the SKIMS Institute are preprocessed as follows: in the first phase, the more relevant attributes or features have been selected. The common features such as fever, cough, rhinitis, sore throat, and fatigue have been selected to form a dataset. The other less potential features such as hemoglobin, blood group, comorbidities, anosmia, pulse, and BP have been discarded. Some of the attributes were merged such as loss of appetite with anorexia, because of synonymity of words. After discarding and merging process, less than 25 features were selected. The second phase is preprocessing of data, in which each column is checked for value. There are some missing values for many of the cases written in the database. To overcome that, some of the columns and rows were eliminated. Like values of BP, pulse was missing in most of the cases so these columns were deleted. Likewise, there were some missing values in many rows; many rows were deleted to overcome that. Lastly, our dataset was reduced to 6015 rows and 21 columns as described in Table 2: selected symptoms of patients.

Table 2

Selected symptoms of patients.

Selected symptoms
Age	Sex	Contact history	Fever	Cough	Rhinitis	Sore throat	Shortness of breath
Myalgia	Fatigue	Loss of appetite	Loss of taste	Vomiting	Nausea	Diarrhea	O2 saturation
Hypertension	Other diseases	Travel history	Anxiety	Chest pain	Severity of illness	Survival	Test result

Normalization is another important step to follow after finalizing the attributes of a dataset. Most of the attributes were categorical in nature such as travel history, residence, cough, and sore throat. Some of the attributes were numerical such as fever, pulse, and oxygen saturation. So, to take the dataset into one form, the normalization is needed. In our case, most of the attributes have categorical value, so other attributes are transformed into categorical value. Suppose if fever is above normal range, it is represented by 1, otherwise, 0. Similarly, all other attributes are converted to categorical value to normalize the dataset. Our dataset is a collection of rows and columns, in which each column represents a binary feature, either 1 or 0. The value 1 of a feature represents the presence of a symptom, and 0 feature represents the absence of that very symptom. Table 3: attributes of dataset, displays the attributes of dataset finalized after the above steps and used during the work.

Table 3

Attributes of dataset.

Symptoms used for analysis
Travel history	Contact history	Fever	Cough	Rhinitis	Sore throat	Shortness of breath	O2 saturation
Myalgia	Fatigue	Loss of appetite	Loss of taste	Vomiting	Nausea	Diarrhea	Chest pain
Sex	Other diseases	Severity of illness	Survival	Test result

4.3. Detection and Prediction of COVID-19 Potential Suspect

Machine learning (ML) is a type of artificial intelligence and subfield of computer science by which machines are learning without being explicitly programmed. ML is categorized into three main categories: supervised learning, unsupervised learning, and reinforcement learning. In ML, a learning algorithm takes input from a set of variables known as a training set. The training set of input values together with target labels known as class labels is called supervised learning. The class labels are unknown in unsupervised learning, and reinforcement learning means learning following the action taken for a given situation. Since our dataset is labelled, our focus will be on supervised learning. The preprocessed dataset developed in the previous section is used to build a prediction model to identify the COVID-19 suspects. The function of this model is to predict the possible COVID-19 suspect by analyzing the symptoms of a person. Various ML algorithms have been employed on the dataset to classify them into either positive or negative. Depending on the working, there are different categories of supervised machine learning algorithms, such as regression-based: logistic regression, function-based: support vector machine, Bayes-based: naïve Bayes, tree-based: decision tree, and meta-based: neural network. In this work, various machine learning techniques, such as SVM, decision tree, naïve Bayes, logistic regression, and neural network, are used while performing the task.

Support Vector Machine: SVM is a supervised machine learning classification technique. It takes predefined set of input training examples with a given class label (i.e., positive (1) or negative (0)) as input. SVM is a function-based learning algorithm that divides the instances of each class with the hyperplane. The trained model is then used to predict the label for any new input. In our case, the hyperplane is trained based on a patient's symptoms with the given class label, either COVID-19-positive or COVID-19-negative.

Decision Tree: DT is a supervised machine learning technique. It takes a set of predefined training data with a given class label as input. DT is a tree-based learning algorithm with three types of nodes: root node, leaf nodes, and decision node. The leaf node exemplifies the class label, and the decision node exemplifies the decision to make. DT normally follows the disjunctive normal form (sum of product) to form a tree. It uses many sub-algorithms and follows criteria such as information gain, entropy, Gini index, and gain ratio, also known as vital function.

Multinomial Naïve Bayes: NB is a supervised machine learning technique based on the Bayes theorem, i.e., follows a probabilistic approach. For a given set of training data with predefined labels, it computes model parameters by calculating the probability of each class label. Then, this is used to assign the class label in the coming instance. MNB is an extended version of NB that uses two or more NB variants. MNB uses the concept of term frequency to compute maximum likelihood from the training data based on conditional probability.

Logistic Regression: LG is a supervised machine learning technique borrowed from statistics. A probabilistic model uses a logistic function to determine the binary variable. Mathematically, a logistic function is having dependent variable with two possible values, such as true or false in case of COVID-19.

Neural Network: NN is also known as artificial NN (ANN) and is nature-inspired machine learning technique. ANN is a meta-classifier-based ML technique that mimics how biological neurons are sending signals to one another. NN takes different inputs i.e., neurons, and outputs one single output. NN is also known as multilayer perceptron because many layers are in between, i.e., hidden layers.

5. Results and Discussion

Performance Evaluation: the performance evaluation of the used machine learning algorithms is measured by six different measures. The six measures are accuracy, precision, recall, F1 score, RMSE, and AUC score. These six measures were validated using confusion matrix and cross-validation methods.

Confusion Matrix: the visualization of performance of binary supervised machine learning algorithm is done by creating a 2 × 2 matrix. The column represents the actual class, and the row represents the predicted or computed class. The matrix representation of 2 × 2 confusion matrix is given in Table 4: confusion matrix.

Table 4

Confusion matrix.

True positive (TP)	False negative (FN)
False positive (FP)	True negative (TN)

True Positive: in this predictive model, the number of instances that were as positive is labelled as positive, and in actual, they are positive. In a true positive result, the persons that do have COVID-19 disease are predicted as positive.

True Negative: in this, the model has classified the instances as negative using predictive model, and in actual, they are also negative. For example, in case of COVID-19, the persons that do not have COVID are predicted by model as negative.

False Positive: the model has classified some instances as positive using a predictive model, and in actual, they are negative. In a false positive result, the persons that do not have COVID-19 disease are predicted as positive. It is also known as type I error.

False Negative: in this, the model has classified some instances as negative using a predictive model, but in actual, they are positive. For example, in case of COVID-19, a person having COVID has shown not COVID by our model. It is known as type II error.

After applying machine learning techniques on the novel dataset, the resulted confusion matrices of applied machine learning algorithms are given in Figure 4: confusion matrices of applied machine learning techniques (a, b, c, d, e). Diagonal elements represent good scores, and other (non-diagonal) represent bad scores.

Figure 4

Confusion matrices of applied machine learning techniques. (a) Support vector machine. (b) Decision tree. (c) Naïve Bayes. (d) Logistic regression. (e) Neural network.

The results generated in the confusion matrices above are summarized in Table 5: summary of results of confusion matrices of different applied algorithms. It is clearly visible from the table that the experimentation has been performed on the balanced data that remove the possibility of high bias or variance. A simple look at the value of TP, TN, FP, and FN tells the whole story about the classification results. In case of disease prediction, a classifier should have a smaller number of false negatives as cost is associated with the false negatives. Suppose in case of COVID-19 prediction, if the classifier has predicted any suspected falsely as COVID-19-negative, it will infect others. Otherwise, if the classifier has predicted any value as falsely positive, it will not infect others. On comparing the algorithms based upon the false negatives generated, it has been found that the decision tree performed better than the rest of the algorithms as fewer entries have been falsely predicted as negative. As in the above definitions, there are two types of errors: type I error and type II error. Both the errors are not good for developed model, but in case of disease the type II error is of main concern. Suppose in case of COVID-19, if our model will drop a person in class of FN, it is type II error and it will infect the others. So, in case of disease the model should have low type II error; otherwise, it will make huge cost to our proposed model.

Table 5

Summary of results of confusion matrices of different applied algorithms.

ML algorithm	True positive (TP)	True negative (TN)	False positive (FP)	False negative (FN)	Support
Support vector machine	861	902	2	40	1805
Decision tree	891	888	16	10	1805
Naïve Bayes	889	884	20	12	1805
Logistic regression	884	890	14	17	1805
Neural network	882	894	10	19	1805

From Table 5, the values are clearly shown against each intersection point of the matrix presented in Figure 4. The matrix is divided into four binary classifications; each quadrant is an intersection of actual class and predicted class. In our proposed system, the hold-out method is used, in which the dataset is divided into training set and testing set. The dataset of 6015 rows is divided in the ratio of 70 : 30, 70% for training and 30% for testing, that is, 4210 rows for training and 1805 rows for testing. The dataset is shuffled to eliminate the biases, so that the proposed model will perform well in all situations. In our proposed system, the decision tree has shown best results in terms of false negatives, i.e., type II error. The decision tree has ten false negatives from the rest of the proposed machine learning techniques. The second place has naïve Bayes algorithm with twelve false negatives, and the third logistic regression, fourth neural network, and last place have a support vector machine. From this discussion, the proposed decision tree model has performed well and it can still be enhanced with the data to minimize the false negatives further.

Cross-Validation: it is a statistical technique used to measure the performance of machine learning classification techniques by splitting the training data into two sets. One set that is usually more than half is used for training, and the rest of the data are for testing. The seventy (70) percent is used for training in our model, and thirty (30) percent is used for testing. Each of the six performance measures (accuracy, precision, recall, F1 score, root-mean-square error, and area under curve score) is calculated for all algorithms and summarized in Table 6: summary of results of different performance measures of applied machine learning techniques.

Table 6

Summary of results of different performance measures of applied machine learning techniques.

ML algorithm	Accuracy	Precision	Recall	F1 score	AUC score	RMSE
Support vector machine	0.97673	0.98	0.97	0.97	0.9766	0.1525
Decision tree	0.98560	0.99	0.99	0.99	0.9856	0.1200
Naïve Bayes	0.98227	0.98	0.98	0.98	0.9822	0.1331
Logistic regression	0.98473	0.98	0.97	0.98	0.9821	0.1310
Neural network	0.98393	0.98	0.98	0.98	0.9839	0.1267

The results generated different performance measures from the novel dataset, in which SVM has achieved the lowest of 97% and the rest of the algorithms have achieved 98% of accuracy. In terms of precision, the decision tree has achieved the highest of 99% and the rest of the algorithms have achieved 98%. The decision tree has achieved 99%, naïve Bayes and neural network obtained 98%, and SVM and logistic regression have achieved 97% of recall. The lowest F1 score of 97% is achieved by SVM, 99% is achieved by the decision tree, and the rest of the algorithms achieved 98%. In terms of AUC score, 97% is achieved by SVM and the rest have achieved 98%. The RMSE should be low, DT and NN have achieved 0.12, NB and LR have achieved 0.13, and SVM has 0.15. DT and NN have good value in terms of RMSE.

Our domain is health care, so the proposed model should have good score in all performance measures. The proposed model is to detect and predict COVID-19 suspect in early stage to eliminate the spread and mortality rate of the infection. In this case, the recall of proposed technique should be good so that the best can be achieved. COVID-19 is the repugnant term, so as positive from November 2019. Corona disease spreads from humans to humans by touching the infectious person and by different ways. If the proposed model predicts a person falsely positive, it will not affect the performance of model in our case. If a model detects a person falsely negative, it will infect many, and it is not an effective model. In technical terms, when a cost is associated with false negative, recall is the best measure to check the model.

Accuracy: accuracy is one of the most important performance evaluation measures used to calculate the performance of any machine learning algorithm. It is computed as the total number of correctly classified instances divided by all instances' summation. Mathematically, it is denoted as follows:

Precision: the efficiency of the supervised machine learning algorithm is measured through several performance measures; precision is among them. It is computed using the correctly predicted positive values ratio to the total positive values. Mathematically, it can be represented as follows:

Recall: it is another performance measure for calculating the efficiency of a supervised machine learning algorithm. It is the ratio of correctly predicted positive values to all values of actual class. Mathematically, it is shown as follows:

F1 Score: the performance measure is used to calculate the performance of a supervised machine learning algorithm. It is computed with the help of two measures, i.e., precision and recall. Mathematically given by the harmonic mean of precision and recall, it is calculated as follows:where

Root Mean Square Error: RMSE is another performance measure used to calculate the performance of a supervised machine learning algorithm. Mathematically, it is computed as follows:

Receiver Operating Characteristic: this curve is another performance measurement criterion for measuring the efficiency of machine learning classification algorithm. ROC is drawn by representing the true-positive rate against the false-positive rate. The area under the ROC is known as ROC curve and is used to measure the classifier's efficiency. The better classifier is the one whose area is closer to 1, and Figure 5: ROC curves of applied machine learning algorithms (a, b, c, d, e) shows the ROC curves of different classifiers. The area under the curve (AUC) is another measure to compute the performance of the machine learning technique to distinguish between the labels, and mathematically, it is computed as follows:

Figure 5

ROC curves of applied machine learning algorithms. (a) ROC curve of SVM. (b). ROC curve of decision tree. (c). ROC curve of naïve Bayes. (d). ROC curve of logistic regression. (e). ROC curve of neural network.

AUC-ROC gives us the complete representation of confusion matrices at different points in the graph. A confusion matrix is given at particular point, but AUC gives us graphical representation of confusion matrices at various threshold points. The drawn line should be close to the upper right corner, i.e., 1, the model's good. In our case, almost all the lines of applied machine learning algorithms are close to the upper right corner of the graph. So, the developed models have achieved good in terms of ROC-AUC curve.

Figure 6, shows the performance evaluation of employed algorithms in terms of accuracy, precision, recall, F1 score, root-mean-square error, and AUC of the different classifiers. The results in Table 4 and Figure 6 indicate that models built using these five different machine learning algorithms on our dataset had achieved above 97% accuracy. The decision tree has achieved 98.5%, SVM had shown 97%, and the rest have shown 98% accuracy, and other values are also good for all algorithms. The results have shown that this model will be effective in predicting the COVID-19 suspects in early stages.

Figure 6

Performance evaluation of employed algorithms.

The graphs for different performance measures such as accuracy, precision, recall, F1 score, RMSE, and AUC score are shown above. One graph is corresponding to one performance measure such as the accuracy of all applied machine learning algorithms to clearly visualize the output. Similarly, other graphs have been drawn to visualize the other performance measures. In terms of accuracy, the decision tree has achieved the highest, SVM has achieved the lowest, and the rest have shown equal. The precision, recall, and F1 score of proposed algorithms are the highest of decision tree, and the rest are at the same place. In case of root-mean-square error, the decision tree has the lowest followed by neural network, logistic regression, and naïve Bayes, and the support vector machine has the highest root-mean-square error value. The SVM has achieved low in terms of area under the curve score, and the rest are at the same place.

6. Comparative Analysis

The proposed work is novel, and the dataset used during the experimentation is primarily collected from patients. The data collected are symptomatic, which is to be used to train the machine learning model to detect and predict the COVID-19 suspect in early stages to eliminate the mortality and spread of the infection. The work is compared with the three different papers based on common parameters. The papers [24, 53, 54] used for comparison are the best papers that can be taken as the benchmark in the field of deep learning. The authors have used computed tomography (CT scan) image set as a dataset. Reference [24] has used hybrid deep learning AI models for lung image segmentation such as SegNet, VGG-SegNet, ResNet-SegNet, and NIH. The proposed hybrid model ResNet-SegNet has achieved the highest accuracy of 99% [53]. The authors have proposed the robust and stable inter-variability of CT lung image segmentation of COVID-19 to avoid bias. The study uses two ground truth (GT) annotations of chest images. The three AI models trained are PSPNet, VGG-SegNet, and ResNet-SegNet on GT annotations. The ResNet-SegNet has performed well in comparison with the other two. Reference [54] is a systematic review of AI technologies with respect to ARDS-COVID-19. The dataset of CT images of lungs has been studied to understand the risk of bias (RoB) in a nonrandomized AI trial for handling ARDS using novel AtheroPoint-AI-Bias (AP(ai)Bias). Reference [55] has taken the dataset of positive patients only and has trained the machine learning model. In this study, SVM and decision table have achieved an accuracy of 93.0% and the rest are below them. In terms of ROC area, the decision table has got the highest of 95.5% and the rest have achieved below it. Reference [56] has been used in their work, but with less number of attributes. Table 7: comparative analysis, of the proposed work with the above works already done in these papers, is detailed as follows.

Table 7

Comparative analysis.

Authors	Technique	Dataset	AUC	Accuracy
[24] Suri et al.	AI, deep learning	CT images	✓	✓
[56] Otoom et al.	Machine learning	Only positive data	✓	✓
[54] Suri et al.	AI, deep learning	CT images	NA	NA
[55] Zoabi et al.	Machine learning	Less number of attributes	✓	NA
[53] Suri et al.	AI, deep learning	CT images	✓	NA
Mir et al. (Proposed)	Machine learning	Primary symptomatic data	✓	✓

7. Conclusion

The article proposes the framework to identify and predict the COVID-19 suspect early to eliminate the mortality and spread of infection. The proposed framework collects the data from sensors and IoT devices and employs machine learning to detect and predict COVID-19 suspect. The framework comprises logically connected four components: data collection layer, data analytic center, diagnostic system, and cloud system. The framework is tested using machine learning algorithms on a real dataset collected from SKIMS, Srinagar. The five proposed machine learning algorithms, support vector machine, decision tree, naïve Bayes, logistic regression, and neural network, have been used during our study. The experimental results have shown that all the ML techniques have achieved above 97% accuracy. The support vector machine has achieved 97.67%, the decision tree has achieved 98.56%, and the rest have a round figure of 98%. The decision tree has achieved good in other performance measures such as precision, recall, F1 score, root-mean-square error, and area under the curve score. Keeping all the performance measures under consideration, the decision tree has performed well on our dataset among all proposed techniques. The proposed framework has the potential to eliminate and reduce the spread of infection through early detection and prediction system. The data stored in cloud can easily be accessed by healthcare professionals to further analyze it to get better insights and better understand the nature of disease. In future, our focus will be to propose ensemble approaches such as random forest and various gradient boosting algorithms to train our algorithms. The dataset used in the above work is not so big that it will be good to use ensemble learning or other methods. Furthermore, deep learning techniques will also be experimented for enhancing the performance measures of the model.