A Machine Learning Solution Framework for Combatting COVID-19 in Smart Cities from Multiple Dimensions

1. Introduction

For over 4 months, the novel coronavirus that was later named Severe Acute Respiratory Syndrome Coronavirus 2 (SARSCoV-2) has caused unprecedented Coronavirus Disease 2019 (COVID-19) all over the world. The first human to human transmission of COVID-19 was first reported to World Health Organization (WHO) on 30th December, 2019. Thereafter, several retrospective studies revealed that many COVID-19 patients started showing pneumonia symptoms in early December. Even though, there are scientific controversies and theories over the date and origin of the SARS-CoV-2, it is widely proven that the novel coronavirus originated from the Wuhan live animal market (Hao, Zhong, Song, Fan, & Li, 2020). Genomic sequences of the early isolates of SARS-CoV-2 from infected patients in Wuhan showed over 88% nucleotide homology with two bat like SARS coronaviruses. Hence, strongly pointing towards zoonotic source and indeed bats serve as reservoir host of the SARS-CoV-2 (Hao et al., 2020). Currently, there are ongoing searches for the possible intermediate host, which might have aided the transmission of the virus to human. The SARS-CoV-2 are droplet borne pathogen which gets in contact with humans when they are exposed to oral or nasal secretions of clinically symptomatically or asymptomatic infected persons (ECDC, 2020).

SARS-CoV-2 has tropism to cells and tissues that express the angiotensin converting enzyme 2. These receptors are mainly found in the respiratory tracts and limited to the extent of the kidney, heart and gastrointestinal tract. The virus dock onto the receptor by means of receptor binding domain (RBD) of its spike glycoprotein. This represent the first step of viral replication and pathogenesis (Rothan & Byrareddy, 2020). As the virus replicates in the respiratory tract, it’s provokes respiratory symptoms which mainly include dry cough, difficulty in breathing and sore throat. Then, it disseminates through the blood to other tissues and organs causing viremia and high fever. Hence, these symptoms together with others such as body weakness and pains represent the cardinal clinical symptoms of COVID-19 (Lin, Lu, Cao, & Li, 2020).

Majority of the SARS-CoV-2 infected persons remain asymptomatic and infection being self-limiting. However, some 5% of infected persons suffer severe COVID-19 (Rothan & Byrareddy, 2020). Major determining factor for severe and fatal COVID-19 include old age (>60 years), underlying cardiovascular, immunological, metabolic and respiratory comorbidities. Based on available scientific reports, the transmission of SARS-CoV-2 revolves around human, animals and the environment (Chakraborty & Maity, 2020). For now, preserving human life and health security is the major concern of most countries and territories. Hence, prompted the legislation, implementation and enforcement of adequate infection prevention and control measures for high priority pathogen (WHO, 2019). Therefore, combatting the COVID-19 need better understanding.

For better understanding of the COVID-19 pattern of spread, accurate and speed of diagnoses, development of new therapeutic methods, identification of the most susceptible people according their distinct genetic and physiological characteristics, machine learning algorithms are required to analyze the COVID-19 large scale datasets (Alimadadi et al., 2020). As a result of that, attempts have been made by scholars to apply machine learning algorithms in combating the COVID-19 in from different perspective. For example, drugs discovery targeting COVID-19 is proposed in (Ge et al., 2020). Machine learning approach in CRISPR based COVID-19 surveillance using genome is proposed in (Metsky, Freije, Kosoko-Thoroddsen, Sabeti, & Myhrvold, 2020). The classification of novel pathogens for the COVID-19 is presented in (Randhawa et al., 2020). Automated deep learning COVID-19 patient detection and monitoring system is reported in (Gozes et al., 2020; Oyelade and Ezugwu, 2020). Machine learning approach is applied to develop system for creating awareness about wash to contain the COVID-19 (Pandey, Gautam, Bhagat, & Sethi, 2020). Generative network is used for the design of COVID-19 3C-like protease inhibitors (Zhavoronkov et al., 2020). The survival of severe COVID-19 patients was predicted based on machine learning approach by (Yan et al., 2020). Deep learning based system for quantifying the volume of lung infections is reported in (Yan et al., 2020).

However, the major issue with the previous studies is that each of the study focuses on a particular aspect of combatting the COVID-19 pandemic leaving other critical aspect that are required to fight the COVID-19 pandemic. (Spectrum, 2020) reported that hospital emergency rooms across the globe are experiencing floods of people infected with COVID-19 for urgent treatment. As a result of that, unprecedented influx of COVID-19 patient into the emergency rooms is experienced, doctors are grappling with the problem of patient’s triage. Therefore, struggling to decide the COVID-19 patients that require intensive care. The condition of the patient lungs has to be assessed by doctors and nurses. However, volunteer doctors and nurses without pulmonary training can’t assess the patient lungs. At the peak of the COVID-19 crises in Italy, doctors were faced with a serious problem of taken decision on the patient that should be given much needed assistance. In view of these difficult COVID-19 cases facing doctors and nurses, a machine learning system can help to provide clinical decision support to play critical role in the COVID-19 crises, assist hospitals to be fully functional and keep COVID-19 patients alive.

A framework with multiple dimension to combat the COVID-19 in different front can provide better automated solutions to the fight against COVID-19. A COVID-19 pandemic involve a lot of measures to combat it via machine learning such as predicting COVID-19 vaccine immunogenicity, detecting COVID-19 severity, predicting COVID-19 mortality, COVID-19 resource allocation, COVID-19 drugs discovery, COVID-19 contact tracing, social distance, detecting wearing of mask, detecting COVID-19 patients requiring ventilator, predicting COVID-19 patient that is beyond medical intervention and triage COVID-19 patients. All these can be integrated into a single framework to work automatically within a city. Therefore, smart city can be repurposed to combat the COVID-19 by applying framework with multiple measures to fight the COVID-19. To provide smart applications in the smart cities, many technologies are connected together including sensor networks, broadband communications service, sensor devices, wireless sensor networks, internet, and cloud services. Sullivan (2008) pointed out that smart cities in 2020 will be embedded with smart structures such as the smart healthcare, smart security, smart mobility, smart building, smart governance, smart citizens, smart infrastructure, smart technology, smart energy and smart education.

To the best of our knowledge, no dedicated framework integrated with machine learning to fight COVID-19 from multiple dimension automatically in smart cities reported in the literature. The framework can address the unique challenges in fighting COVID-19 to ease the work of healthcare workers in saving life’s and provide guide for real world execution in smart cities. (Yu, Wang, Liu, & Zomaya, 2016) pointed out that exploring theory is critical for providing guide for effective applications.

In this paper, we propose a solution framework integrated with machine learning to combat COVID-19 in Smart Cities from multiple dimensions.

The other sections of the paper are organized as follows: Section 2 presents the background information about COVID-19, Section 3 presents the basic concept of smart city including case studies and machine learning, Section 4 presents mathematical model of COVID-19 spread in smart city, Section 5 presents the applications of machine learning in fighting COVID-19 pandemic in smart cities, Section 6 outlined case studies on combatting COVID-19 via machine learning in smart city, Section 7 presents the proposed solution framework for combatting COVID-19 in smart city from multiple front, before the concluding remarks in Section 8.

The summaries of the contributions of the study are provided as follows:

• A detail solution framework integrated with machine learning to fight COVID-19 in smart cities with guide for effective real world execution is proposed.

• Mathematical model of the COVID-19 spread in smart cities is presented for easy understanding.

• Use cases of machine learning in combatting COVID-19 in smart cities in presented in the study.

• As a novel virus, the study outlined cardinal clinical and laboratory features of the COVID-19 for different stages to properly guide the machine learning community in developing healthcare decision support related to COVID-19 to be embedded in smart healthcare.

2. The Novel Coronavirus Diseases: background and cardinal clinical features

Out of the seven Coronaviruses, SARS-CoV-2 is the third highly pathogenic coronaviruses that has afflicted human race. As the incidence and fatality rates of SARS-CoV-2 infection, the etiological agent of COVID-19, continues to rise across more than 210 countries and territories, several preventive and control measures have been adopted to halt the spread of the SARS-CoV-2 and minimize COVID-19 associated death. As at 7:30AM GMT+1, 27^th April 2020, there were over 3 million global confirmed cases of SARS-CoV-2 infection with case fatality rate (CFR) of over 7.0% (Worldometer, 2020). European and American countries appeared to have the worst CFRs associated with COVID-19 and least in Africa. Although, there is no categorical explanation for this variation. Several observers have attributed the low incidence rate of COVID-19 in Sub-Saharan Africa to under-diagnosis probably due to inadequate molecular diagnostic capacity. But variation in the genetics, strains, viral proteins mutations and host immune response could have contributed to SARS-CoV-2 virulence and pathogenesis (van Dorp et al., 2020).

Although, there have been controversies on the origin of SARS-CoV-2. However, several studies have been able to trace the source of this infection to zoonotic origin where COVID-19 patients were exposed to an animal market in Wuhan where live animals were sold. Subsequently, efforts have been made to search for a reservoir host and/or intermediate hosts of SARS-CoV-2 from which the infection might have spread to humans. Initially, two snake species were identified as the possible reservoir hosts of SARS-CoV-2. However, the only consistently identified SARS-CoV-2 reservoirs are mammals and bats (Bassetti et al., 2020). Particularly, genomic sequencing of some SARS-CoV-2 isolates show 88% nucleotide homology with two bat-derived-SARS-like CoVs (Lu et al., 2020). Thus, indicating bats as the most likely reservoir hosts for SARS-CoV-2 (Lu et al. 2020).

Virologically, SARS-CoV-2 is a single stranded RNA virus with positive polarity and variable open reading frames (ORFs) (Cui, Li, & Shi, 2019). It has been shown that two-third of SARS-CoV-2 genome are located within the 1st ORF which translates the pp1a and pp1ab polyproteins. These polyproteins encode 16 non-structural proteins (Cui et al., 2019). Conversely, the remaining ORFs code viral structural and accessory proteins of SARS-CoV-2. The remaining one third genome codes the nucleocapsid (N) protein, spike (S) glycoprotein, matrix (M) protein, and small envelope (E) protein of SARS-CoV-2. Out of these 4 proteins, the S glycoprotein is key because of it is role in host cells attachment and pathogenesis of COVID-19. This protein alongside the viral RNA dependent RNA Polymerase (RdRP) have largely been utilized in the synthesis of primers and antigens for molecular and serological tests of SARS-CoV-2 infection, respectively (WHO, 2020).

Indeed, RNA viruses including SAR-CoV-2 have high mutation rates which are significantly correlated with enhanced virulence and evolvability (Duffy, 2018). At proteomic level, amino acid substitutions have been reported in NSP2, NSP3 and S protein (Wu et al., 2020). Interestingly, another study suggested that NSP2 and NSP3 mutations play significant role in virulence and differentiation mechanism of SARS-CoV-2 (Angeletti et al., 2020). Of interest, is the mutation in S-protein. This has made scientists explore the possible differences of the host tropism and transmission rate of SARS-CoV-2. Worthy to note is the NSP2 and NSP3 mutations in SARS-CoV-2 isolated from many COVID-19 patients in China (Angeletti et al., 2020). These have caused scientists to embark on genomic surveillance of SARS-CoV-2 in order to determine the correlation of these mutations to virulence diversity and their implications on reinfection, immunity and vaccines development (Liu et al., 2020).

Based on available genetic analysis, SARS-CoV-2 is related to SARS-CoV-1 and an early mathematical modelling report revealed an R₀ of 2 - 3 for SARS-CoV-2 (Yan et al., 2020). This could possibly explain why SARS-CoV-2 is more contagious than SARS-CoV-1 and MERS-CoV which are endemic in certain countries. Infection rate can be referred to as the R₀, a measure of the transmissibility of SARS-CoV-2. R₀ predicts how many people an infected person could transmit SARS-CoV-2 in a population with no prior immunity to the pathogen. Generally, the higher the R₀, the more contagious the pathogen. An R₀ of < 1 means that the outbreak dies out while R₀> 1 means the infection will continue to spread (Prompetchara, Ketloy, & Palaga, 2020).

First, it is worthy to note that a single SARS-CoV-2 infected individual has the ability to infect approximately three uninfected persons (Kampf, Todt, Pfaender, & Steinmann, 2020). This occurs through the nano-particles of respiratory droplets which can spread, contaminate surfaces and hands where they remain stable for hours (Kampf et al., 2020). The hands now become a mechanical vector. Thus, a potential site to terminate and prevent it from invading the body. However, if the virus is not eliminated at this stage, it can move towards it is predilection site (i.e. cells of the lungs) where it uses it is spikes to dock and attach angiotensin-converting enzyme-2 (ACE-2) as receptors to gain access into epithelial cells of the respiratory tract. At this stage, SARS-CoV-2 compromises innate lung immunity (Kampf et al., 2020). It then takes advantage of these cells as replication site. The virus regenerates and sheds by disassembling itself and utilizing the machinery of the alveoli cells, precisely Golgi apparatus, to reproduce and repackage itself (Kampf et al., 2020).

The SARS-CoV-2 exists in such a way that as it replicates and disrupts the protective function of the ACE-2 receptor which induces the process of fibrosis (scarring). It has been shown that patients with fatalities associated with SARS-CoV-2 present with a characteristic ground glass effect in their lungs and this sequela impedes efficient oxygenation. As the body tries to compensate for this deficiency, it gradually results to Severe Acute Respiratory Syndrome (SARS), where it becomes impossible for the respiratory system to make available oxygen to the rest of the body (hypoxia) (Kampf et al., 2020). This ultimately results in multiple organ failure. Based on available clinical data, those susceptible to contracting severe form of SARSCoV-2 infection include the elderly (>60 years) population, persons with underlying disease conditions (cardiovascular, metabolic, respiratory and immunological disorders) (Lippi and Plebiani, 2020).

When susceptible individuals get infected by SARS-CoV-2, its either the infected person remain asymptomatic (no apparent illness) or symptomatic. If symptomatic, the disease passes through three stages of severity. At the early infection stage (Stage I), patients present with mild clinical symptoms which includes dry cough, diarrhea, fever and headache. This could last for 3-5 days. These are usually accompanied by lymphopenia (low white blood cell counts), elevated prothrombin time, D-dimer and mild increase in Lactose Dehydrogenase (LDH). Majority (98%) of SARS-CoV-2 infected patients remains at this stage and eventually get cured. However, for those with underlying medical disorders, they may proceed to stage II (Pulmonary Phase), which is predominantly characterized by Shortness of breath and hypoxia (inadequate oxygen supply to the body). This could last from 5 days to 3 weeks). At this stage, patients display abnormal chest radiograph, Transaminitis and declined procalcitonin level. Very few patients (2%) proceed to the severe stage of COVID-19. The stage II (hyper-inflammation phase) is largely characterized by acute respiratory distress syndrome (ARDS), severe inflammatory response syndrome (SIRS), shock and cardiac failure. Majority of patients at this stage eventually die. At this stage COVID-19 patients experience significantly high blood inflammatory markers such as elevated C-reactive protein (CRP), Interleukin-6 (IL-6), D-dimer and ferritin. In addition, affected patients present with increased blood level of Cardiac markers especially troponin and N-terminal (NT)-pro hormone B-type natriuretic peptide (NT-proBNP) (Gao et al., 2020), the summary is presented in Figure 1.

• Download figure
• Open in new tab

Figure 1:

Cardinal Clinical, Laboratory Features and Phases of COVID-19

Diagnostically, the use of viral culture for establishing acute COVID-19 diagnosis is not practicable due to its long turnaround time (3 days) for SARS-CoV-2 to cause obvious cytopathic effects (CPE) on Vero E6 cells. In addition, isolation of SARS-CoV-2 is laborious and requires biosafety level-3 (BSL-3) facilities which are unavailable in most healthcare centers, especially in developing countries. So far, all available serum antigen (such as the S-glycoprotein) and antibody (IgA, IgM and IgG) detection tests have not been validated by the WHO. However, it has been suggested that serological assays could assist in the analysis of an ongoing SARS-CoV-2 outbreak and retrospective evaluation of the incidence rate of an outbreak (WHO, 2019). In some instances, where epidemiological data of suspected cases correlates to SARS-CoV-2 infection, the demonstration of fourfold rising antibody titer between acute and convalescent phase sera could support diagnosis of COVID-19 when RT-PCR results are negative (WHO, 2019). In addition, it has been revealed that a significant proportion of COVID-19 patients had tested RT-PCR negative despite having suitable clinical features and radiologic findings highly suspicious of SARS-CoV-2 infection (Xiao, Wu, & Liu, 2020). In most cases, these are termed false negatives which could have been due to wrong sampling where SARS-CoV-2 might have been present in the lower respiratory tracts rather than upper respiratory samples often collected during laboratory diagnosis. Hence, this poses a challenge in the proper evaluation of SARS-CoV-2 symptomatic patients (Li et al., 2020).

3. Rudiments of smart cities and machine learning

This section discussed the concept of smart city including case studies and provide brief explanation about machine learning. This can help readers new in the domain to comprehend the concept of smart city and machine learning.

4. Mathematical modelling of the COVID-19 spread in smart cities

In this section, we develop the proposed model representing the dynamics of COVID-19 in smart cities. For a comprehensive and detailed model representing the transmission dynamics of the novel COVID-19, we divided the total human population existing within a smart space into various compartments and the various intervention strategies adopted for the control of the transmission and management of the disease.

The first compartment in this model is the susceptible group or population. In this population, there are new entrants through birth only into the population (since we assume that restriction of movement is used as a control strategy with in a smart space), and Λ represents it. Note that due to the model looking at COVID-19 over a long period, natural birth rate and death rate cannot be ignored. Furthermore, any person in this compartment will either be vaccinated or unvaccinated (giving a reason for the quest for vaccines as a control strategy). Also, some people in this group can die from natural causes as they are not yet infected. Thus, the susceptible group is further subdivided into the susceptible vaccinated (S_v) and the susceptible unvaccinated groups (S_uv).

The susceptible group that has been vaccinated initially may lose the immunity induced by the vaccine and, therefore, can either be infected or move straight back to the susceptible group that is unvaccinated. They may equally die naturally represented by μ. For the susceptible group that is unvaccinated, they will on contact with infectious persons get infected at a rate α₁ and then join the exposed group. They may equally die naturally. Note that every vaccinated or unvaccinated person might not develop or contract the disease as far as they do not have contact with an infectious person. The next group is the exposed group.

We need to note here that the exposed group (E) comprises of all persons that have had contact with an infectious person (this is where effective contact tracing comes in as a control measure) regardless of the infection status until confirmed to be negative. Note that this model does not include false positives and false negatives. So, those who become infected or had contact with the infected persons and are in the vaccinated or unvaccinated subgroups move into this exposed group. Since person can be infected, if he has contact with infected surfaces, it means such persons also join this group. From this exposed group, we now either self-quarantine, government-quarantine, or the individual develop the symptoms and move straight to the infectious subgroup (I). The personal-quarantine subgroup is represented by P_q and government-quarantine by G_q. From these quarantines, one is either confirmed free from the virus and moves back to the susceptible group or confirmed to be infected and moves to join the infectious group. At this level, the only death considered is from natural causes. After these subgroups, we now consider the infectious subgroup represented as (I). This group comprises of those who became infectious as previously described and those who became infectious by inhalation of the droplets of the virus from the air either due to close proximity to an infectious person sneezing, coughing or droplets just in the air as a recent investigation has shown that the droplets can last for more than nine (9) hours in the air (Lewis, 2020) and this is done at a rate σ.

When a person is infectious, an assumption is made that such a person must undergo treatment. The person must self-isolate and get treated (P_T) or be government-isolated and treated (G_T). We note that the infectious persons are joined by those declared as infectious through self-isolation (h₂) or government-isolation (h₁). Also, as noted recently in China, some who had hitherto been declared recovered got re-infectious and then moved back to the infectious subgroup (Wu and McGoogan, 2020; Li, Geng, Peng and Meng, 2020). Those on personal-isolation and treatment (P_T) decreases by those who die either naturally or as a result of the disease. Equally, some persons recover. Similarly, those on government-isolation and treatment will decrease by those who recover, or those who die either naturally or due to COVID-19. Finally, we have the recovered group, and this comprises of those who recovered due to treatment either personally or by government. This population decreased by those fully recovered rejoining the susceptible group of getting re-infected and then infectious and rejoin the infectious group. Based on these explanations, we have the assumptions of the model and then the following flow diagram showing the sequence of development and control of the disease and then the model equations for the COVID-19. We show these as:

1. Some susceptible persons are vaccinated.

2. There is contact tracing of all persons that had contact with infected persons and all such persons undergo quarantine to confirm their disease status.

3. Personal and government quarantines exist.

4. There is personal or government treatment, but every infectious person must be treated.

5. It is not only through person-to-person contact that infection can occur.

6. Immunity can be lost for some reason over time hence rendering the vaccinated person susceptible or infected.

7. The COVID-19 sometimes induces death, and so infectious persons can die through COVID-19.

8. Recovered persons can be re-infected or susceptible.

9. One can become infectious by inhalation of the droplets of the virus from the air.

10. Those on quarantine can be confirmed free after the incubation period (1 - 13 days) and can either rejoin the susceptible or the infectious group if re-infected.

11. Some persons may have contact with infected persons and never undergo any quarantine and so they develop the infection and move straight to the infectious subgroup.

12. At the infected level, one can die even before personal treatment or government treatment.

Therefore, using all these assumptions, we have the flow diagram as shown in Figure 2 as follows:

• Download figure
• Open in new tab

Figure 2:

Diagram summarizing the model for COVID-19 in a smart city

Definition of the parameters and Symbols used in the Model

View this table:

• View inline
• View popup

Table 1:

Summary of parameter definition of the proposed mathematical model

The corresponding new diagram summarizing the main structure of the above-described model can be presented in Figure 2, and the resulting system of equations describing the model is expressed in the following form:

Equation (1) describes the change in the populations of the susceptible individuals in the populations over time.

Equation (2) describes the change in the population of the vaccinated susceptible individuals over time.

Equation (3) describes the change in the population of the unvaccinated persons over time.

Equation (4) describes the time-change in the number of persons exposed which includes all persons that had contact with infected persons or surfaces.

Equation (5) describes the time-change in the population of persons under government quarantine.

Equation (6) describes the change in the population of those under personal quarantine over time.

Equation (7) describes the change in the population of the infectious persons over time.

Equation (8) describes the change in the population of those on personal treatment over time.

Equation (9) describes the change in the population of those on government treatment over time.

Equation (10) describes the change in the population of those who recovered over time.

5. Applications of machine learning in fighting COVID-19 pandemic in smart cities

Smart city being an inter-connected urban society and collecting data every moment from several embedded devices, smart cities can effectively work with machine learning approaches during this COVID-19 pandemic. It basically implies that machine learning techniques are hugely dependent on data for better learning and predictive models. Thus, the machine learning techniques have the ability to bring out some intrinsic and useful insights to help smart cities decision makers to take preventive measures during the COVID-19 pandemic. In view of the fact that different machine learning techniques are embraced by Artificial intelligence (AI), which is the model with the ability to self-learn, can also support immensely. It is important to discuss the role of the AI along with the machine learning to combat COVID-19 because first of all the data availability is too limited and we have to deal with real-time streaming of the data. Thus, significance of self-learning systems becomes much more desirable in smart cities. Figure 3 demonstrates the overall flow of how AI & machine learning approaches can help in fighting COVID-19 pandemic in smart cities.

• Download figure
• Open in new tab

Figure 3:

Application and flow of AI & machine learning in combatting COVID-19

As shown in Figure 3, there are several types of data generated from the information and communication technology equipment embedded in the smart cities. These data are as follows:

1. First kind of data is the statistical data that normally contains the daily statistics of number of identified cases, no of positive cases, number of deaths, number of recovered cases, etc. This would help in the prediction of future cases to prepare for emergencies.

2. Second type of data is the epidemiological data, which majorly contains all of the clinical tests data concerning test results of different medication, various drug trails, patient’s medical history, patient’s response to several medications, etc.

3. Third type of data is the real-time surveillance data generated from sensors and cameras in the smart cities. One of the initial detection of COVID-19 is based on symptoms is the fever. People body temperature as per facial recognition and other personal information are the type of data which are also useful to prevent spreading of COVID-19.

The data is processed and analyzed through the machine learning approaches for extracting insights in various applications. Notice that deep learning is shown in Figure 3 inside the machine learning block. It’s a type of machine learning algorithm which is much more relevant in adapting to large scale new data, learn itself from the data and update to recognize patterns. The applications of the machine learning in different aspect for combatting COVID-19 are discussed as follows:

6. Case studies on combatting COVID-19 via machine learning in smart cities

In this section, we present use cases where machine learning techniques are applied to help in combating COVID-19 in smart cities. The use case can help readers really understand exactly the way machine learning can provide aid in the era of fighting COVID-19 pandemic. As such, other nations can share the expertise in fighting the COVID-19 by applying the machine learning approaches. The use cases are discussed as follows:

7. Propose smart city framework integrated with machine learning for combating COVID-19

To address the multiple dimensional challenges outlined in Section 1, as a result of COVID-19, a proposed framework is required within the smart city context to allow decision makers to take a crucial decision on the best way to combat COVID-19 from multiple dimensions. The framework consists of multiple components as shown in Figure 4. Each component has a major impact on enhancing the quality of the analytics to combat COVID-19.

• Download figure
• Open in new tab

Figure 4:

Propose smart city framework integrated with machine learning for fighting COVID-19

7.4 Analytics

Recent years have received an increasing trend in image processing in healthcare due to convolutional neural network which has become a significant approach for handling large amount of images generated from the smart cities. Allam and Jones (2020) discuss the universal data sharing standards coupled with AI to benefit urban health monitoring and management. The proposed framework has the potential to incorporate various machine learning and deep learning algorithms to develop the analytical model. These algorithms range from traditional shallow approach such as neural network, Decision Tree, Naive Bayes, K-nearest neighbor to more sophisticated algorithms such as Convolutional Neural Networks, Deep Generative Modeling, Deep Belief Network and Deep Recurrent Neural Networks. These algorithms can be applied to run on COVID-19 dataset to help in the combatting COVID-19 in hospitals in smart cities across the world. These applications include detect, prevent the spread of COVID-19, forecast next epidemic, diagnose cases, monitor COVID-19 patient, suggest vaccine development, track potential patient, help COVID-19 drug discovery and provide better understand of the COVID-19 virus in smart cities.

7.4.1 Social media information verification

The COVID-19 pandemic has come with the challenges of fake news associated with it including conspiracy theories. Since the COVID-19 pandemic start spreading across the globe, there has been a lot of fake news regarding the COVID-19 origin, cure, mode of spread, treatment, and many other myths. This is very common on the social media platforms such as Facebook, Twitter, Instagram, YouTube, etc. In the smart city, citizens voice out opinion on social media regarding COVID-19, as such generate a lot of unstructured data. It is reported by (Obeidat, 2020) that no systematic quantitative study has been conducted to ascertain the magnitude of the myths regarding COVID-19 on the social media but certainly the figures of the misinformation about COVID-19 is significant. The fake news regarding COVID-19 can come in the form of manipulated content, misleading content, satire, false context, malicious account, fabricated content, false connection and imposter content. Therefore, machine learning or deep learning algorithms can be applied to detect fake news regarding COVID-19 on the social media and alerts the citizens living in smart cities.

7.4.2 Prediction of Future pandemic

Although being the worst phases in recent times, this pandemic has come in the times of digital age. Therefore, every aspect of analysis is now being captured including macro-level analysis, logistics, biological etc., in terms of data which will definitely be fruitful for predicting the new and unknown (possibly sources) future pandemic. For example, with the help of machine learning approach (random forest), (Eng, Tong, & Tan, 2014) were able to predict the possible zoonotic strains of influenza i.e. some virus which could only affect animals can also be dangerous to humans. It only implies that machine learning can help in predicting future pandemic which may come from any species.

The only limitation being the data could be from different domain i.e. the source of COVID-19 is possibly from ‘bats’ and in the past, sources of pandemic are different (say, different gnome structure etc.). And traditional machine learning requires data distribution to be from the same domains in training and testing. However, Transfer Learning (TL), a part of machine learning, can effectively handle such situations where training and testing data could be from different data distribution. That is, the knowledge learned from the past pandemics could be used in other situations with new domain even with fewer amounts of data. Such scenario is shown in Fig. 4a, where the pre-trained model from current COVID-19 pandemic (with large data and labels) could be used to predict future pandemics, prepare the smart city for that situation and help overcome quickly in case of the spread (with very less amount of data and labels).

• Download figure
• Open in new tab

Figure 4a:

Typical transfer learning model for future pandemic predictions

7.6 Containment framework of COVID-19 in smart cities: Contact tracing and basic preventive measures

In this section, we display the proposed framework. As at the time of this study, there are no defined set of procedures or approved drugs that can provide complete cure to patients affected by the recent COVID-19 pandemic. However, as temporary preventative measures, usage of face masks and shields, limiting the movement of the hand to the face, social distancing, washing of the hands vigorously with soap, water, and alcohol-based sanitizers have all been recommended for the general population. Also, for people who have been exposed to COVID-19 or display symptoms, measures like usage of face masks, using personal items alone, self-quarantine, and government quarantine are some of the measures put in place. Similarly, new machine learning techniques and application such as the use of facial recognition systems embedded in different location in the smart spaces e.g. bus station, subway station, airports, etc can be developed to detect those with facial mask, or gesture recognition system to enforce social distances. However, it is crucial to note that due to the scale of the pandemic and the lack of fore-knowledge to most of the populace, obtaining adequate containment measures can be cumbersome. In this regard, scientists have made several efforts with the help of valuable artificial intelligence strategies to curtail possible escalation of the pandemic, through continuous monitoring of the COVID-19 spread. In addition, machine learning based algorithms have equally been developed and used in making efficient classifications and predictions that could suggest preventive measures of the COVID-19 virus from further spread. A typical example is the application of machine learning techniques in COVID-19 contact tracing (CT), such as facial recognition and entity resolution analysis, to track persons of interest across various signal-based media and social network platforms.

The CT has been used as a method of monitoring and controlling the pandemic contagious diseases, such as Ebola (Classen et al., 1999; Browne, et al., 2015; Webb et al., 2015), Tuberculosis (Begun, Newall, Marks, & Wood, 2013), and Middle East Respiratory Syndrome (MERS) virus (Breban, Riou, & Fontanet, 2013). CT is the process of identifying and following up on people who may have come into contact with an infected person (Fisman, Khoo, & Tuite, 2014; Salathé et al., 2020; Webb et al., 2015). In Figure 6, we present an explicit illustration of a typical CT process flow diagram for inhibiting the spreading of COVID-19. Accurate modelling of CT requires precise information about the disease transmission pathways from each individual, and hence the network of contacts. Since contact tracing takes place as a process over the system of interactions between hosts, it is natural to consider network-based models for this process. The CT model presented in this paper utilizes the peer-to-peer network architecture that centers on the query routing strategy for identifying infected individuals or potential carriers of COVID-19. Similarly, CT can in this case be supplemented with machine learning algorithms to develop technologies that alerts to higher levels of immediate risk and supports the rapid facilitation of practical interventions such sanitization equipment and deep cleans.

We present a peer-to-peer network-based framework for implementing an effective CT process in a smart city. In particular, the main contributions of this section is the design of an intelligent machine learning based CT strategy. The proposed framework has three main goals:

1. It takes into account the dynamic characteristics of the human network when performing CT process

2. Communication and query referral between communities hosting suspected COVID-19 persons is considered so efficient CT can be performed and tracked, and

3. The strategy demonstrates scalability, making it suitable for adoption into other related pandemic cases that would require both human and financial resources on a large scale.

7.6.1 Peer-to-Peer Network Model

In smart spaces relative to smart city, networked system of places and individuals provide us with resourceful platforms that focus on global information sharing, searching, and replicated categorization of directories (Khambatti, Ryu, and Dasgupta, 2002; Khambatti, et al., 2003). In this instance, much of the interactions among interconnected peered network is directed at forming an overlay network that is based on cluster identifiers and using the overlay network to perform searches on a content addressable space. In the current proposed smart city framework, we extend the aforementioned peer-to-peer network concepts to allow the notion of smart communities or connected society. Smart communities are useful in structuring the information storage space, discovering resources, and pruning the search space. It also aids in better dissemination of helpful information (Khambatti et al., 2003).

A peer-to-peer network consists of a large number of nodes that can join or leave the system at any time. For illustration, consider people living in smart spaces that comprises of households, neighborhoods, and communities. Here, each peer (or node) connects to a relatively small set of neighbors, who, in turn, are connected to larger peers (or smart spaces of community) such as the case of peer C in Figure 5. Similarly, the neighbors of peer A are peers’ B, and C. Formation and discovery of peers belonging to some specific communities are significantly dependent on how peers declare and use their common “interests” (Khambatti et al., 2002). Firstly, we define attributes as a method of declaring an interest, and then we use these attributes to discover affected communities. Secondly, we also define another set of attributes as a method of identifying peers infected with COVID-19 viruses, and we use these attributes to trace all possible contacts made by an individual peer within the designated smart spaces of community. Therefore, two things are involved in the peer-to-peer network machine learning based CT model, first is the identification of communities where COVID-19 pandemic has occurred, and second is the identification of infected COVID-19 virus peers together with all their contacts.

• Download figure
• Open in new tab

Figure 5:

Regular connections amongst peers of human network

The adopted machine learning based smart city CT strategy conforms with the architectural perspective of the well-known Hop-Count Routing Indices (HRI) search mechanism for peer-to-peer network systems. The HRI algorithm discussed in this paper is assumed to have a deep learning features and thus our notion for machine learning based CT concept. Therefore, the hybrid machine learning based HRI algorithm is different from the classical HRI algorithm. In Figure 6, we present a model for the HRI contact tracing strategy that unveils the dynamic interactions of the susceptible and infected COVID-19 populations based on smart space community or regional distribution of the COVID-19 virus. The machine learning based HRI algorithm incorporates the main features of contact tracing strategies, which include; the number of contacts per identified infectious case of COVID-19, the likelihood that a traced contact is infected, and the efficiency of the contact tracing process.

• Download figure
• Open in new tab

Figure 6:

Proposed machine learning based CT model outlines for smart spaces in a smart city.

The initial identification of communities where peers could fall into based on their common interest and the community classification into either COVID-19 free or epicenter zones is vital to the implementation of the proposed HRI-contact tracing algorithm. The identification process would help in pruning the search space during the CT process. Since it would only be the affected communities or epicenter zones that would be contact traced during the CT process.

7.6.2. Attribute-Based Community Networks

In this section, we present a community formation framework that determines the membership of a peer in one or more smart spaces or communities. For identification as part of a community, a peer is required to possess or declare attributes that are usually identified with members of that community. For example, in the smart spaces, several common interests and attributes exists in the formation of a community. These shared interests could come from the following: culture, religion, peer or age group, social activities, tribe, etcetera. For instance, a peer is expected to have originated from one ethnic group by birth. The ethnic group has its unique cultural backroad and probably its accepted religion or beliefs, as is the case with most tribal communities in some countries.

A similar use of peer-attributes in network-formulation is detailed in (Khambatti et al., 2003). (Khambatti et al., 2002) stated that attributes could be either explicitly provided by a peer or implicitly discovered from past queries. The authors further went on to classify peer attributes into three categories, namely, personal, claimed, and group. The main reason for these classifications is as a result of privacy and security concerns. A full set of peer attributes refers to a collection of personal peer attributes, and subsets of the peer’s personal attributes are referred to as the peer claimed attributes. The personal attributes disclose complete attributes information about the peer, while the asserted attributes only disclose those attributes that the peer wishes to make public. The group attribute is a location or affiliation oriented and is needed to form a physical basis for communities. Table 3 provided examples of these three attribute classes.

View this table:

• View inline
• View popup
• Download powerpoint

Table 3:

Peer attribute formation

The group-based attribute is relevant since it associates a peer to a specific community. In other words, every peer belongs to one pre-determined group and has a group attribute that identifies the peer as a member of that community.

7.6.3. Population Structure

The population structure for the peer-to-peer community network model presented in this paper follows a group hierarchical population structure in (Kasaie, Dowdy, & Kelton, 2013). The fabric covers the following settings;

1. The infected peer’s immediate family members

2. The infected peer’s neighbors or the immediate community hosting the infected peer’s together with the household members

3. Finally, the communities that contain the family members, neighbors, and other extended neighboring communities.

The machine learning based CT process usually would first start from the COVID-19 infected peer household, move on to the closest neighbors, before proceeding to neighbor communities for some possible contacts. The sizes of the mixing group peers (or community) are chosen arbitrarily in such a way that it would enable the modelling of different disease transmission routes associated with each mixing peers. The peers presented in Figure 6, only provides a realistic representation of the population structure, which would aid the simulation implementation of the proposed framework.

Note that the notion of smart spaces is used here to denote spaces occupied by households or family members, neighbourhood, community, etc. The COVID-19 life cycle considered in this paper follows the mathematical model sequence presented in section 4 above, which consists of the following compartments;

1. “S” the number of susceptible individuals, who can be infected

2. “E” the number of exposed individuals, who have been infected but not yet infectious

3. “I” the number of infectious cases in the community, who are capable of transmitting the disease

4. “H” the number of individuals who are in the hospital, and

“R” the number of individuals removed from the chain of transmission (cured or dead and buried), respectively. A person is assumed to be born in full health and Susceptible to COVID-19. Upon successful transmission of the disease, the person enters the Latent or incubation state, which varies between 1 to 13 days, in severe forms of the disease, death may occur within an average of 14 days after onset of illness (Ivorra, Ferrández, Vela-Pérez, & Ramos, 2020; Kucharski et al, 2020).

7.6.4 Smart City Contact Tracing Network Structure

In modelling the smart city CT strategy, a three-layer contact network in connection with the population structure referred to as close, casual, and random contact is considered for the peer-to-peer hop-count routing indices algorithm. The three contact network structures represented in the model presented in Figure 7 illustrate the social relationships that could be established or exist between any peer with the rest of the peers’ population. Following the three identified group-mixing, that is the family member group, neighborhood group, and community group mixing, we defined each of the contact types to fall under each of them. For instance, the close contact type will fall under the family member group mixing. In contrast, the causal contact type will fall under the neighborhood group mixing, and the random contact type will fall under the community group-mixing. Structuring the population sizes of each group mixing will differ from places to places or country to country, depending on some parameters that might be linked to either social, cultural, or religious beliefs. Table 4 shows three contact types for the three group mixing.

• Download figure
• Open in new tab

Figure 7:

Cluster-based peer-to-peer smart city CT architecture

View this table:

• View inline
• View popup
• Download powerpoint

Table 4:

Peer-to-peer contact types within a smart space

In view of the possible individual network contact types within a confined smart space, the close contacts type defines the first level, which is the most recurrent interaction type among family members within a confined smart space. On the one hand the casual contacts which is the second level type of contact, represent a social relationship that often occurs among friends at a smart recreational center, sit out spaces, neighbors at a shopping malls, or children at learning institutions such high schools and universities. However, these types of contacts are less regular and cherished as compared to the first level type of contact. More so, it more restricted to a specific network of related individuals that are connected via a predefined smart spaces, including friends, coworkers, etc. (Classen et al., 1999; Raffalli, Sepkowitz, & Armstrong, 1996). For the proposed machine learning contact tracing model, we limit the domain of level one type of contact to each neighborhood’s residents and assume a limited period of twelve months for the duration of the interactions. Finally, the third type of contact or interaction, which is the random contacts are used to represent interactions among individuals at smart spaces, for example, at smart bus or train stations, institutions, etc. This type of contact network has a very limited duration of interaction that would last for a period of one month. This type of contact usually accounts for the potential risk of COVID-19 transmission among non-related people in the whole smart community.

We anticipate the proposed machine learning based CT framework to be an essential tool in slowing down and stagger the spread of the current COVID-19 pandemic or any other pandemic. The structure, when implemented, can be used as an efficient and effective process of identifying people who may have come into contact with an infected person and then amassing further information about the people who are potentially exposed. Indisputably, the new machine learning based CT model, if adequately implemented, can curtail the spread of infection in a vulnerable population. It can be employed by the government and policymakers to help people be aware of their infection status by being alerted. It also offers a diagnosis for a potentially vulnerable population, gain insights into the path of the spread of the disease to aid in further preventive measures. Finally, it allows the uninfected population to move around freely, provided the identified contacts have quarantined themselves.

8. Conclusions

In this paper, we propose a solution framework based on machine learning for integration in smart cities to fight COVID-19. The propose machine learning solution framework integrated different task to fight COVID-19 in the smart cities from different dimension such as predicting COVID-19 vaccine immunogenicity, detecting COVID-19 severity, predicting COVID-19 mortality, COVID-19 resource allocation, COVID-19 drugs discovery, COVID-19 contact tracing, social distance, detecting wearing of mask, detecting COVID-19 patients requiring ventilator, predicting COVID-19 patient that is beyond medical intervention and triage COVID-19 patients. The paper presented a comprehensive guide for implementing the machine learning framework in smart cities. We believed that the solution framework has the potential to automate the measures of fighting the COVID-19 pandemic in smart cities from multiple dimension to ease the fatigue of the healthcare workers due to very high number of COVID-19 patients requiring medical attention simultaneously and provide widespread access to quality healthcare system. In addition, we hope that the proposed smart city machine learning based framework for combatting COVID-19 will serve as essential guide to the research community in developing a more compartmentalized forecasting and analyzing tools with prospect of mitigating the spread of the ongoing COVID-19 pandemic and the reoccurrences of any similar future pandemic disease. For future work, it will be interesting to see the real world application of the proposed framework to further investigates the model practicality and efficiency in smart city environments.