A Novel Edge-Computing-Based Framework for an Intelligent Smart Healthcare System in Smart Cities

Abstract

The wide use of internet-enabled devices has not left the healthcare sector untouched. The health status of each individual is being monitored irrespective of his/her medical conditions. The advent of such medical devices is beneficial not only for patients but also for physicians, hospitals, and insurance companies. It makes healthcare fast, reliable, and hassle-free. People can keep an eye on their blood pressure, pulse rate, etc., and thus take preventive measures on their own. In hospitals, too, the Internet of Things (IoT) is being deployed for various tasks such as monitoring oxygen and blood sugar levels, electrocardiograms (ECGs), etc. The IoT in healthcare also reduces the cost of various ailments through fast and rigorous data analysis. The prediction of diseases through machine-learning techniques based on symptoms has become a promising concept. There may also be a situation where real-time analysis is required. In such a latency-sensitive situation, fog computing plays a vital role. Establishing communication every time with the cloud is not required with the introduction of fog and thus the latency is reduced. Healthcare is a latency-sensitive application area. So, the deployment of fog computing in this area is of vital importance. Our work focuses on improving the efficiency of the system for the precise diagnosis of and recommendations for heart disease. It evaluates the system using a machine-learning module.

1. Introduction

In the current wireless network communication environment, fog computing and the Internet of Things (IoT) are splendid technologies for intelligent healthcare applications and security network modeling [1,2,3]. At the same time, artificial intelligence (AI) and deep-learning applications have also won much recognition and acknowledgment in recent years. With the onset of the COVID-19 pandemic, the situation was modified to an even greater degree. During the crisis, people witnessed a rapid digital transformation both in rural and smart cities and the adoption of disruptive technologies across distinct industries. Healthcare became one of the beneficiary sectors that received many blessings from deploying disruptive technologies. AI, machine learning, and deep-learning methods have become a crucial part of the world. Deep knowledge of healthcare has had a huge impact and has enabled the sector to enhance patient monitoring and diagnostics [4,5,6,7,8].

Devices were digitized to the IoT in many ways, particularly in business and healthcare. They have continued to share a wealth of essential information for the effective deployment of healthcare systems. The IoT is, after all a force that needs to be fought with in terms of how electronic security frameworks are viewed and the industry as a whole due to the rapid decrease in costs for IoT equipment and the billions of new IoT devices that are anticipated to be introduced in the coming years. The evolution of medical services simply keeps pace with the introduction of newer technologies. People look forward to smart devices with comprehensive sensing capabilities, improved connection protocols, and pervasive data analytics over collected data [8,9,10,11]. For clarity, the concepts of cloud and fog computing are delineated.

(i)

Cloud computing is a kind of computing dependent on sharing computer assets in place of having local servers to address data packages. Through cloud computing, each type of processing and calculation related to sensitive user data is accomplished in the cloud. Previously for computing frameworks, the computer server or every other server has been in use. However, more advanced computing may now be achieved on the cloud. Figure 1 depicts the functionality of the cloud layer, fog layer, and body sensing layer in a standard design of a healthcare application. Currently, each business organization and venture is transforming towards cloud technology with the added advantage of fog computing.

The concept of cloud computing has furnished ease to our computations. Through cloud computing, we spurn the number of servers supplying services to us or the configuration. Data processing may be completed on the computer. The server can also be carried out in a distributed manner between the computer and the server. Through cloud computers, distinct devices for a special type of work can be examined [3]. Cloud computing includes many elements, which include cloud storage and saving applications in the cloud. In cloud computing, programs are used to gather and process energy over the internet. It is a pay-as-you-go cross-carrier. Without owning any computing infrastructure or any information centers, each person can rent access to some programs via storage from a cloud service provider. In this case, the complexity of retaining infrastructure via the use of cloud computing offerings is reduced. In short, cloud computing service vendors can take advantage of large economies of scale by giving equal services to a huge range of clients.

(ii)

Fog computing is a decentralized computing infrastructure or method in which computing resources are located among the data supply and the cloud. Fog computing is a paradigm that provides offerings to user requests at the edge networks. The fog layer gadgets typically perform networking operations along with routers, gateways, bridges, and hubs. Researchers envision these gadgets to act simultaneously on each computational and networking process. However, these gadgets are resource-restricted compared to cloud servers. The geological spread and the decentralized nature help in supplying reliable services with coverage over a vast vicinity. Fog computing is the physical place of the gadgets that are much closer to the customers than the cloud servers. Despite the broad deployment of cloud computing, some systems and services still cannot take advantage of this trendy computing paradigm due to inherent problems of cloud architecture, such as undesirable delay, lack of mobility hold-up, and location awareness. As a result, the fog computing paradigm has developed as a potential infrastructure to provide efficient communiqué and resource consumption.

The cloud is centralized storage situated farther from the endpoints than any other type of storage. Fog acts as a middle layer between cloud and edge and provides the benefits of both. It relies on and works directly with the cloud, handing out data that do not need to be processed on the go.

As the main subject of the current paper is the design of a smart healthcare system, the research outcomes of [9,10,12,13] depicted that smart healthcare is an essential factor of smart living in modern towns. Essentially, healthcare is one of the fundamental needs of human life, and smart healthcare systems are progressively contributing to developing a healthy society [14]. There are several additives of smart healthcare, including the IoT, the Internet of Medical Things (IoMT), clinical sensors used in wireless networks, artificial intelligence, fog computing, cloud computing, mist computing, and Wi-Fi communication [15,16]. Many research works have been published within the last decade, focusing on novel approaches to solving critical diseases such as heart disease, diabetes, cancer, depression, hemorrhoids, lupus, psoriasis, etc.

In the current contribution, a new architecture using fog computing is proposed for better care of heart disease using an advanced deep-learning method in an i-fogsim simulation environment. The result analysis indicates a high rate of accuracy in heart disease diagnosis with comparatively reduced network bandwidth and minimum energy consumption compared with similar approaches.

The current paper has been divided into subsequent sections. Segment 1 presents the preface to the subject. Section 2 describes the literature review. Section 3 portrays this advanced healthcare using an IoT-enabled system. Section 4 presents the mist-enabled smart health system architecture and proposed work for a smart city. Section 5 presents a comparative investigation and a discussion of the results. Finally, Section 6 gives the concluding remarks and future work on the proposed approach.

2. Related Work

IoT-based technology is employed to create a refined smart city with all necessary settings and user comforts. Given that the occupants come from a variety of elegant environments, this provides suitable statistical data analysis for numerous resources. Researchers can build and improve enhanced health-based applications by using advanced town communications to affect people-centric statistical data.

More information about wearable technology can be found everywhere in ecological sensing and creative actuation [17]. Smart health programs can be developed using cutting-edge methods, and end-to-end implementations are also available that can be applied in a variety of settings. In [1], the modular approach for IoT packages is described as much as the context machine for advanced health troubles, enabling the potential to develop with the offered information, use standard-motive system mastering, and decrease the computed unemployment and density. For advanced health, this produces reaction instances for essential conditions, the additional resourceful classification of health-associated environments, and the next acceleration in elegant metropolis surroundings. They display capability with three sets of consistent background-conscious submissions; the retrieval of well-being-related human-beings-oriented context, user presence, consumer motion, air suitability, and position as IoT sensors. Research work [2] has talked about a parasite-inbred sickness that accumulates quickly in diverse areas of the United States.

A new paradigm in elegant health had to be created to draw attention to and prevent this disease. The positive potential of developing IoT technology for continuously changing devices and materials has played a crucial role in the subsequent technology health concern systems for elite patient care to protect the citizens from those types of illnesses [18]. Even so, there can be a requirement for real-time health monitoring to look into the patients and determine whether any early preventive actions or precautions for a healthy existence are necessary [19]. S-Health services IoT has a long way to go before it can recognize and track analog signals. It consists of connected apps, objects (such as gadgets), communication technologies, tracking equipment, and patient knowledge bases.

An IoT-enabled version is provided by research in [2] that uses statistics. By gathering information regarding the factors contributing to the rise in mosquito populations, precautionary steps can be taken. Information is sent to the cloud with the aid of side nodes after the method’s acceptability is verified at the IoT layer. Simulation results suggest that the proposed technique is superior to the ME-CBCCP procedure [20]. A fantastic study in [3] showed how the IoT improves cybercities and AI-based systems to create intelligent designs. Along with healthcare, the IoT has expanded its utility. The healthcare ecosystem is described in Figure 2.

IoT is currently a widely used individual contributor to the world at large. As a result, IoT support for healthcare is leading the way since it enables significantly more accurate supervision of health status. A significant amount of information has historically been produced by the hygiene industry. The introduction of IoT has significantly increased this volume over time. Therefore, to make sense of the data, the health statistics must be carefully maintained. An effective small cloud-assisted IoT-backed e-Health structure is proposed in the research article [4].

People’s physical health is crucial in smart cities, and research work [5] has argued that this is because of the unique challenges of occupancy, accessibility, and ease of access. The IoT offers the most promise in terms of possible technological interference. Wearable critical sensors communicate data into unique software program engines, notably speedy and active integration for successful diagnosis. Researchers in [5] described an IoT-based, fully improved peripheral system for remote fitness examination (RASPRO). RASPRO creates tailored health themes from extensive sensor statistics and clinically significant summaries of population health management (PHMs). The Coded Mark Inversion computes a mixed-criticality rating (CMI) signal machine. The advanced aspect of IoT uses a possible stratified protocol, including the quick, assured communication of signals and PHMs straight to the doctors as well as a first-class attempt to extract the exact information required.

Khansari et al. [21] conducted a comprehensive review of the various most recent security challenges and their countermeasures in the smart healthcare environment. In addition, an artificial intelligence (AI) and blockchain-based secure architecture were proposed as a case study to analyze malware and network attacks on wearable devices in this paper.

Alabdulatif et al. [22] proposed a multilayer perceptron algorithm to assess the accuracy of the model and distinguish private and public hospitals as a novelty approach. Overfitting resulted in finding a reliable MLP model, and the accuracy of the model was acceptable.

The experimental legalization and performance evaluation of the advanced aspect scheme has been completed. The empirical evaluation of 183 patients established that the IoT elegant area is exceedingly valuable in remote tracking, developing caution, and recognizing cardiac situations [8], as measured through three units, decimal (0.87) and F1-score (0.85). Furthermore, overall performance evaluation confirmed enormous reductions in capacity (98%) and power (90%), which means manufacturing is appropriate for budding small-group IoT arrangements. In the exploitation of the coordination in the heart disease organization of the universal health facility, it turned out that its IoT advanced technology aspect helped to grow the availability of physicians, utilizing fifty-nine percent. Hence, IoT-advanced smart gadgets are a widespread step closer to addressing the requirements for international health.

3. Advanced Healthcare Using an IoT-Enabled System

The IoT system’s wireless medium can considerably modify how physical well-being structures are used at quarters or in a healthcare atmosphere [12]. A structure intending to use the gadgets ought to cope with the necessities of using their capability and safe utilization. It is thought that such a structure will assist in developing implementation and effectiveness, final gaps connecting the range of rising physical conditions by IoT structures. The study in [12] provided some requirements to be considered quintessential to the ongoing growth of the digital fitness sensing environment (Health IoT). The cutting-edge landscape of the IoT was considered in terms of these needs and obtainable clarifications that deal with extreme necessities: (1) balancing cellular fitness apps (giving users management and possession over their app as well as records), and (2) making cell apps act and perform like some other aspect in a sensor. The research in [12] suggests functioning and assessing those physical condition sensor necessities to expose how particular hygienic solutions can pressure and affect the design of more fabulous generalized IoT architectures. As COVID-19 spreads through an area, a not-unusual reason for locating a rapid clarification to manipulate the pandemic has brought together scientists, hospitals, authorities, and civilization [23].

3.1. Fog-Assisted Secured Healthcare

A fog-assisted IoT architecture for better health management as well as monitoring has been designed [24]. Web (internet) of factors (Internet-of-things) and system-mastering-oriented structures integrate elegant wearable machinery that is unexpectedly evolving to supervise and manage healthcare and physical performance. It is centered on the intention of a fog-centric Wi-Fi, real-time, advanced body sensor network and IoT device-like structure designed for ever-present well-being in addition to health analysis in refined health group surroundings. The projected structure aims to use resources related to the well-being and strength of manufacturing fixed-frame motion and fitness-connected facts. Such structure is ordinary to help contestants, running shoes, and doctors with the translation of more than one bodily sign and send signals if there is any health risk.

Researchers devised a technique to accumulate and examine exercising factual information, which can be used to calculate workout concentration and its increase in athlete fitness and function as a suggestion machine for athletes. As per the research in [8], IoT technology affords a dependent method to deal with the carrier deliverance factors of healthcare in terms of cellular health. IoT produces an extraordinary number of facts that may be processed using cloud calculation. However, the delay due to shifting realities to the shade and back to the submission is intolerable for instantaneous, remote fitness to observe programs. In this regard, the authors of [8] suggested applying the concept of fog technology at the smart gateway to track distant impacted individual health in smart homes. The recommended version in [8] uses advanced methods and products, including embedded data mining, distributed storage, and notification systems at the community’s edge. To transmit the patient’s real-time data at the fog layer, an event-triggering-based records transmission approach is used. By determining the patient’s temporal fitness index, the secular mining concept is employed to examine the adverse events. Health data from 67 patients in an advanced residential environment powered by IoT were systematically generated for 30 days to determine the system’s validity. Analysis showed that, in comparison to other classification methods, the proposed Bayesian concept network classifier-based model has high accuracy and reaction speed in determining the state of an event. Additionally, decision-making that is solely based on current healthcare facts improves the use of the suggested equipment. The main area of focus and performance parameters used in each research project are highlighted in

Table 1. Illustration of research works focused on ML and AI.

Ref	Artificial Intelligence/Machine Learning Technique	Focus/Contribution	Limitations
[25]	Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement as far as applicable to qualitative evidence syntheses.	This study provides a thorough summary of the ethical concerns raised by LHCS and emphasizes the significance of striking a balance between the social goal of bettering healthcare and the need to safeguard and respect individual participants.	Discussion of challenges with interpretations is limited. This process should ensure as much as possible the reliability and validity of the findings.
[26]	Smart Tissue Autonomous Robot (STAR) against conventional surgical techniques.	The study investigated the augmentation of surgical capabilities to improve outcomes and broadened access to care that may be made possible by the combination of AI and surgical robotics.	The study fails to extensively explain early designs, which are likely to be expensive and have a limited number. Adoption of new technologies where underserved medical contexts provided their clinical efficacy is established and their technology is continuously improved.
[27]	Qualitative research method. It is a method of research that relies on non-numerical and unstructured data; the research is based on secondary data. Support vector machine (SVM), 3D convolution neural networks.	The growing usage of technology in the industry makes information systems project management in healthcare relevant. The Internet of Things, artificial intelligence, augmented reality, and big data are just a few of the technologies that are being deployed.	The pandemic situation is discussed, but no primary data were used, which may have produced better findings.
[28]	Classification trees (decision trees) are visual and easily interpretable models for solving regression and classification problems.	The tasks completed in the project “Development and research of machine learning methods in the issues of diagnosis and assistance of patients with illnesses of the cardiovascular system” are summarized in this article.	Precision is insufficient for machine-learning models to work properly. One field where the results’ interpretability is crucial is medicine. Understanding why the model anticipated a certain result is crucial for clinicians to use the system effectively.
[29]	Qualitative and quantitative factors. Additionally, the Bibliometrix R software package was helpful for the paper.	The study revealed that there is a growing body of literature in this area. It focuses on clinical decision-making, patient data and diagnostics, health services management, and predictive medicine.	There are currently some limitations that will affect future research potential, especially in ethics, data governance, and the competencies of the health workforce.
[30]	They used two machine-learning models, i.e., Support Vector Machines (SVM) and Random Forest (RF) and their results showed a staggering agreement with related works.	The technological advancement in mobile computing, artificial neural networks, robotics, the storage of huge data on the internet, cloud-based machine learning, information processing algorithms, etc., has propelled the use of AI in modern healthcare.	The study needs to address technical difficulties during a surgical procedure. Moreover, complicated surgical interventions need cognitive as well as decision-making skills.
[31]	Random Forest and Simple Logistic Regression methods	Age, urine albumin excretion, low-density lipoprotein, triglycerides, cholesterol, and glomerular filtration rate (GFR) have all been identified as risk factors for DN.Additionally, fasting plasma glucose (FPG), potassium, and changes in FPG and GFR after year 1 were identified as early and late biomarkers of DN, respectively, using ML approaches.
[32]	The work reviewed various deep-learning architectures, such as the convolutional neural network (CNN), recurrent neural network (RNN), graph neural network (GNN), autoencoder (AE), Boltzmann machine (RBM), and deep belief network (DBN), as well as their variations that incorporate transfer learning, attention learning, and reinforcement learning.	A thorough analysis of the previous seven years’ worth of work in the domains of genetics, sensing, internet communication health, electronic health records, medical imaging, and artificial intelligence in health informatics.	The fact that current research is reliant on retrospective data from hospitals that are easily shareable datasets raises the possibility of data contradiction and bias. We see missing value issues due to patient transfers, poor chart maintenance, and procedural tendencies.

, which also provides a comprehensive overview of the research trends and inspiration for other research efforts.

3.2. Healthcare Using Artificial Intelligence and Machine Learning Support

The research-based knowledge in [9] depicted that elegant health concerns are an essential factor of smart and associated living. Physical condition concerns are individually the fundamental support a person requires, and excellent fitness concerns are anticipated as the way to supply quite a few billion dollars in returns in the near future. We can find quite a few additions to advanced fitness concerns, together with the web, web-based IoMT, clinical sensing devices, machine learning, facet calculations, cloud analytics, and the subsequent invention of the Wi-Fi communiqué generation. Numerous papers within the literature address advanced fitness concerns or fitness care as popular investigation areas in IoT and IoMT science, machine learning, fog networks, security in healthcare systems, and scientific computing. Table 1 illustrates research works focused on ML and AI.

In a related work [33], where artificial intelligence and machine learning have been used, the authors mentioned that a Health Smart Home (HSH) is a critical part of an intelligent metropolis. It offers a new sort of remote scientific treatment. It might efficiently improve the absence of scientific sources due to the aging population and assist aged humans to live at home more thoroughly and independently. Activity recognition is a significant point of Health Smart Home. However, building activity recognition models typically requires a large number of labeled records, which enforces a heavy weight on guide marking. In this work, the authors suggested a movement labeling approach based totally on visually rectified semi-supervised learning rules. Such advances can divide the unrefined sensor data collection with label statistics into suitable fragments.

3.3. Healthcare Using Deep-Learning Methods

Deep learning is a part of machine learning, which is inspired by neurons in the human brain: there are tens of millions of neurons in the human brain, and there are more than 100,000 connections between them. The deep-learning method is called an artificial neural network [34]. It is a unique instance of artificial neural networks, a technique for machine learning that was inspired by the human brain [35]. Remarkable improvements in the area have been made as relates to the capacity of machines to comprehend and modify data, including voice, language, and picture data. Due to the enormous amount of data being created, the expansion of medical equipment, and the adoption of digital record systems, healthcare, and medicine, stand to gain a great deal from deep learning [36].

Deep learning employs several methodologies that present both important opportunities and applications with great potential. These fields frequently center on diagnosis, detection, and prediction. Most image processing methods are based on deep learning (DL), most especially on artificial neural networks (ANNs) and convolutional neural networks (CNNs). An advancement of ANNs was used in contemporary methods to increase performance while categorizing pictures [37]. Healthcare terms (“deep learning” OR “neural network” OR “machine learning” OR “artificial intelligence”) in combination with the terms (i) medical imaging, (ii) EHR, (iii) genomics, (iv) sensing and online communication health were used to search across multiple databases and estimate the trend and performance of the algorithms in health informatics. Significantly pertinent publications for each section with the use of DL algorithms were selected from the literature and quickly examined [8]. Figure 3 illustrates some deep-learning applications in healthcare for detecting diseases, analyzing data in healthcare, drug discovery, predicting unforeseen circumstances in healthcare, record keeping, and combating the coronavirus disease (COVID-19).

4. Mist-Enabled Smart Health System for a Smart City

4.1. System Architecture

Mist-enabled innovative health: this platform builds and implements an integrated IoT mist–fog–cloud framework with organized communication and application execution, which is platform-independent. Mist-enabled innovative health integrates a range of IoT sensors, such as healthcare detectors and routers, to communicate information as well as jobs to fog worker end nodes. MistBus broker nodes perform resource management and task execution. Mist-enabled innovative health incorporates blockchain, authentication, and encryption methods to maintain data integrity, privacy, and security, enhancing the cooperative framework’s stability and robustness [38,39]. MistBus connects with the cloud via HTTP RESTful APIs and combines fog installation with the Aneka software package [40].

The mist-enabled smart health model is an IoT-based fog-enabled cloud computing paradigm for healthcare that can better manage patient data and monitor medical status to determine the seriousness of heart disease. Through software components, mist-enabled smart health connects a variety of hardware instruments, allowing for a system with coherent point-to-point incorporation of mist–fog–cloud, considering quick as well as effective outcomes. Figure 4 illustrates this HealthFog framework which is made up of several hardware components as well as software components, which will be discussed later.

4.2. The Hardware Component of the Mist-Enabled Smart Health Model

These hardware elements make up a mist-enabled smart health prototype:

• Network of body area sensor devices: This part is made up of three primary sensing devices: health sensors and activity sensors with environmental sensors. Electrocardiogram sensors, electroencephalogram sensors, electromyography sensors, respiratory sensors, thermocouple sensors, and differential pressure sensors with potentiometric sensors are among the medical sensors available. These components are responsible for transferring the data from the patient’s body through associated gateway devices.
• Gateway: Overall, three kinds of gateway gadgets (cell phones, laptops, as well as tablets) operate even as mist nodes, acquiring data from various sensors and forwarding them to broker/worker units for some further assessment.
• Mist–fog bus modules: The below are indeed the components of the mist–fog bus structure:
  • The broker node collects the task requests and input data from gateway endpoints. Before sending data, the request input module accepts job requests from gateway devices. To enhance overall trustworthiness and data security, the Security Supervisor module enables protected interactions between two units and prevents this acquired data from unwanted access or harmful alteration. The load statistics among all worker nodes supply into the arbitration module (a component of the resource supervisor on the broker node), which then decides in real time whether this is the same end node or a subset of end nodes to send jobs to.
  • Worker node: One module that handles assigned jobs is the broker node’s resource supervisor. Embedded types of equipment, single-board computers (SBCs) such as Raspberry Pis, can be used as worker nodes. Worker nodes in MistBus can also use advanced deep-learning models to analyze and interpret data as well as generate recommendations. Other features such as computerized information, data purification with mining, big data analytics, as well a repository can be added to the worker node. Through the gateway devices, the worker nodes receive input data, produce results, and communicate them with the gateway devices. In this proposed framework, if it requires this, the broker end node can act similarly to the worker node.
• Cloud Data Center: This mist-enabled innovative health utilizes the fog infrastructure or Cloud Data Center resources if this mist framework becomes overcrowded, services are latency tolerant, or the length of incoming data is significantly more prominent than a typical one.

4.3. The Mist-Enabled Smart Health Model Software Components

The relevant software components from the mist-enabled innovative healthcare model are data cleaning and processing, data analytics, resource supervisor, judgment module, deep-learning module, and accumulating module. The relevant software components form the mist-enabled smart health model:

• Data cleaning and pre-processing: Pre-processing data is the first action once submitted. It also involves employing data analytics technologies to filter information. To obtain essential ingredients of data feature vectors that influence patient health conditions, the filtered data are condensed to a smaller size using Principal Component Analysis (PCA) adopting Set Partition [41] and protected using the Singular Value Decomposition (SVD) technique [42]. It immediately concludes the input data, then suggests medicine and appropriate check-ups depending on the continually trained data-related healthcare professionals, and preserves it in a database for re-training as required.
• Resource supervisor: Workload management and arbitration modules [43] are the two elements that define this framework. The workload manager keeps track of job requests and task queues for data processing. It also maintains large amounts of data that must be evaluated. The arbitration component assigns the mist, fog, or cloud resources that have been provided for processing tasks that have been scheduled and handled through this workload supervisor. The arbitration part is embedded with this broker end node and evaluates whether the fog computing node, the fog worker end node, or the Cloud Data Center must give this information to find outcomes [43]. To balance the load and ensure optimal performance, the fundamental purpose is to divide tasks across multiple devices. Users can customize their load balancing and arbitration techniques depending on their application needs with mist-enabled innovative health.
• Deep-learning module: The dataset is used to develop a neural network for identifying data endpoints that will feature vectors created following the pre-processing of this information acquired through this wireless body area network. This anticipates and processes output for this information gained through these intelligent gateway gadgets depending on the task assigned by the resource supervisor.
• Ensembling module: This component takes the results from many models and utilizes polling to measure the output class; that is, whether or not the individual has cardiovascular disease. The feature is located on the task-assigned FogBus node and is responsible for relaying and gathering reliable out-run through various worker end nodes.

4.4. Sequence of Communication

Every hardware device in the mist-aided innovative health model communicates via specified procedures or the three scenarios described earlier: broker only, worker node, and cloud. The gateway transmits a job request to the broker node in every situation. Depending on the circumstance, the broker unit delivers the gateway to either the worker IP address or the master IP address. The broker unit may or may not monitor worker workloads in the broker-alone situation; when all worker units are overloaded or have been corrupted and the cloud is turned off, the broker delivers the gateway endpoints to their IP address without cloud transmission. When there are worker end nodes that are optimal or underloaded, this broker gives the gateway gadget the IP address of the lowest-laden worker node. Expanding the amount of worker end nodes may lengthen the arbitration process by necessitating additional load checks. In the case of an absent cloud layer, the router gadgets deliver the task, such as the supplied information for analysis. After that, the worker/broker node performs pre-processing, runs the forecast model, and gives an output to the router gadgets, because the gateway gadget may not be connected to the VPN, the data collected are sent to the broker end node and transferred to the cloud layer. This protects IoT detectors as well as router gadgets from hostile actors because these are only linked to the LAN with other mist nodes rather than the internet. System performance will likely be faster in the cloud because of resource availability. Still, latency will be maximum because of communication overheads and queue delays at both the broker and cloud layers. This broker provides the gateway router device with the IP address of the least-occupied worker node when there are worker end nodes that are optimum or underloaded, and the worker node to which the data are provided, employing bagging, selects the majority class. The communication sequence in Mist Health is given in Figure 5.

4.5. Proposed Work

The proposed model implements the algorithm in python. It uses deep-learning techniques for data normalization. This will consider the maximum and minimum values of the parameters for the efficient distribution of values. This processing will occur in the mist layer. The proposed algorithm will randomly distribute the input data and predict the value. At the diagnosis time, this will produce the result. Several parameters have been considered for the purpose. Input will be taken from a data set. For the execution of the deep-learning algorithm, three hidden layers will be considered. Finally, the system will generate the output in binary format. This will display yes if the patient has heart disease and no if the patient does not has heart disease.

4.6. Resource Scheduling Algorithm for Mist Layer

The resource scheduling algorithm is responsible for managing the tasks to be executed, denoted as T, and for managing the worker nodes for handling the set of worker loads, $\mathrm{WL}$, where $\mathrm{WL}=\left\{{\mathrm{WL}}_1,{\mathrm{WL}}_2,\cdots ,{\mathrm{WL}}_{\mathrm{n}}\right\}$. The worker nodes are responsible for handling the large amount of data to be processed in the mist layer generated from the IoHT layer and accordingly maintain the worker load and queue the tasks to be executed. In order to balance the load in the mist layer, the fundamental objective of the proposed framework is to optimize the workload being allotted to each worker node in order to minimize the number of resources being used in the mist layer, and this increases the efficacy of the proposed framework. The detailed working mechanism for resource scheduling in the mist layer adopted in this work can be depicted through Algorithm 1.

Algorithm 1 Algorithm for Resource Scheduling in the Mist Layer

Input:$\tau ,T,WL$ Output:$T_{}$ Initialize: $\tau \leftarrow$Set the threshold value for the worker load $T\leftarrow$set of tasks to be executed $T=\left\{T_1,T_2,\cdots ,T_n\right\}$ $WL\leftarrow$Set of worker load $WL= \left\{W{L}_1,W{L}_2,\cdots ,W{L}_n\right\}$ $W\leftarrow$Set of workers $W= \left\{W_1,W_2,\cdots ,W_n\right\}$ if $Min\left\{W{L}_1,W{L}_2,\cdots ,W{L}_n\right\}>\tau$ do           $Send data to Mist Layer$ else         $Transfer T to \arg \underset{W}{\min }\left(WL\right)$ end if return $T,WL,\tau$ exit

4.7. Q-Reinforcement Learning over Mist (QRLM) Algorithm

This work introduces the use of Q-reinforcement learning, defined as a value function-based reinforcement learning model [43]. Q-reinforcement learning is meant to optimize the numerical rewards of a model for mapping the different learning states to individual actions. By exploiting the Q-reinforcement learning algorithm, the proposed intelligent healthcare monitoring system can obtain specific given goals by employing its experience gained during the learning process through interaction with the desired decision-making system.

Following [43], it is evident that if any given task can be defined as a discrete stochastic process, then it is said to satisfy the Markov property, which can simplify the given real-world problem over continuous space. This work leverages the Markov decision process for modeling the proposed Q-reinforcement learning-based training approach. The policy-based reinforcement learning approach is adopted because it offers a more stable solution under function approximation and performs well for continuous action spaces where reasonably sizable real-world datasets are involved. Furthermore, when the action space is broad, policy-based reinforcement learning offers quick convergence [43]. Here, a specific Markov decision process is considered over a tuple, $〈\mathrm{S},\mathrm{A},\mathrm{P},\mathrm{R},\mathsf{\delta }〉$, where $\mathrm{S}$ corresponds to the set of states, $\mathrm{A}$ denotes the set of actions, $\mathrm{P}$ represents the probability of transitioning the present state, $\mathrm{R}$ depicts the set of rewards, and $\mathsf{\delta }$ denotes the reward discount factor. Thus, the elements for the state transition matrix corresponding to the proposed algorithm can be defined as,

$$\mathrm{p}\left(\mathrm{s}\prime |\mathrm{s},\mathsf{\alpha }\right)=\mathrm{p}\left({\mathrm{S}}_{\mathrm{k}+1}|{\mathrm{S}}_{\mathrm{k}},{\mathrm{A}}_{\mathrm{k}}\right)$$

(1)

where ${\mathrm{S}}_{\mathrm{k}+1}$ denotes the set of $\mathrm{k}+1$ states, such that ${\mathrm{S}}_{\mathrm{k}+1}=\mathrm{s}\prime$, and ${\mathrm{S}}_{\mathrm{k}}$ and ${\mathrm{A}}_{\mathrm{k}}$ denote the state and action, respectively, at $\mathrm{k}$ states, such that $\left({\mathrm{S}}_{\mathrm{k}}=\mathrm{s}\right)$ and ${\mathrm{A}}_{\mathrm{k}}=\mathsf{\alpha }$.

Further, $\mathrm{A}$denotes the finite action set, such that $\mathrm{A}=\left\{{\mathsf{\alpha }}_1,{\mathsf{\alpha }}_2,\cdots ,{\mathsf{\alpha }}_{\mathrm{n}}\right\}$; hence, the immediate reward for the finite set of actions can be given as,

$${\mathrm{R}}_{\mathrm{k}}=\mathrm{R}\left({\mathrm{S}}_{\mathrm{k}},{\mathrm{A}}_{\mathrm{k}}\right)$$

(2)

The policy can be stated as a mapping function, $\mathsf{\pi }$, which is responsible for defining how an agent behaves in a given learning environment based on its observations. Specifically, $\mathsf{\pi }$ is responsible for mapping the states and actions to a probability distribution, given as,

$$\mathsf{\pi }\left(\mathsf{\alpha }|\mathrm{s}\right)=\mathrm{p}\left({\mathrm{A}}_{\mathrm{k}}|{\mathrm{S}}_{\mathrm{k}}\right)$$

(3)

Now, considering the start state distribution as ${\mathrm{f}}_0$ and the policy $\mathsf{\pi }$, the expected return for all possible trajectories can be represented as the probability corresponding to some state $\mathrm{k}$ and is defined as,

$$\mathrm{p}\left(\mathsf{\phi }|\mathsf{\pi }\right)={\mathrm{f}}_0\left({\mathrm{S}}_0\right)\times \left[{\prod }_{\mathrm{k}=0}^{\mathrm{K}-1}\mathsf{\pi }\left({\mathrm{A}}_{\mathrm{k}}|{\mathrm{S}}_{\mathrm{k}}\right)\mathrm{p}\left({\mathrm{S}}_{\mathrm{k}+1}|{\mathrm{S}}_{\mathrm{k}},{\mathrm{A}}_{\mathrm{k}}\right)\right]$$

(4)

Hence, given the possible trajectories $\mathsf{\phi }$ and the reward function for our considered model $\mathrm{R}$, the expected return can be obtained as,

$$\mathsf{\theta }\left(\mathsf{\pi }\right)={\int }_{\mathsf{\phi }}\mathrm{p}\left(\mathsf{\phi }|\mathsf{\pi }\right)\times \mathrm{R}\left(\mathsf{\phi }\right)$$

(5)

Solving the above equation, we get the expected return as,

$$\mathsf{\theta }\left(\mathsf{\pi }\right)={\mathsf{{\rm E}}}_{\mathsf{\phi }~\mathsf{\pi }}\left[\mathrm{R}\left(\mathsf{\phi }\right)\right]$$

(6)

The Q-reinforcement learning approach is specifically used to improve the policy in order to maximize the expected return. Therefore, the optimal policy can be computed as,

$${\mathsf{\pi }}^{\mathrm{o}}=\arg \underset{\mathsf{\pi }}{\max }\mathsf{\theta }\left(\mathsf{\pi }\right)$$

(7)

where ${\mathsf{\pi }}^o$ denotes the optimal policy.

Now, for the policy $\mathsf{\pi }$, the value function can be given as,

$${\mathrm{V}}_{\mathsf{\pi }}\left(\mathrm{s}\right)={\mathsf{{\rm E}}}_{\mathsf{\phi }~\mathsf{\pi }}\left[\mathrm{R}\left(\mathsf{\phi }\right)|{\mathrm{S}}_0\right]$$

(8)

$${\mathrm{V}}_{\mathsf{\pi }}\left(\mathrm{s}\right)={\mathsf{{\rm E}}}_{{\mathrm{A}}_{\mathrm{t}}~\mathsf{\pi }}\left[\sum _{\mathrm{k}=0}^{\infty }{\mathsf{\delta }}^{\mathrm{k}}\mathrm{R}\left({\mathrm{S}}_{\mathrm{k}},{\mathrm{A}}_{\mathrm{k}}\right)|{\mathrm{S}}_0\right]$$

(9)

Further, if the agent acts as per the policy $\mathsf{\pi }$, then we can compute the Q-value as follows,

$${\mathrm{Q}}_{\mathsf{\pi }}\left(\mathrm{s},\mathrm{a}\right)={\mathsf{{\rm E}}}_{\mathsf{\phi }~\mathsf{\pi }}\left[\mathrm{R}\left(\mathsf{\phi }\right)|{\mathrm{S}}_0,{\mathrm{A}}_0\right]$$

(10)

$${\mathrm{Q}}_{\mathsf{\pi }}\left(\mathrm{s},\mathrm{a}\right)={\mathsf{{\rm E}}}_{{\mathrm{A}}_{\mathrm{k}}~\mathsf{\pi }}\left[\sum _{\mathrm{k}=0}^{\infty }{\mathsf{\delta }}^{\mathrm{k}}\mathrm{R}\left({\mathrm{S}}_{\mathrm{k}},{\mathrm{A}}_{\mathrm{k}}\right)|{\mathrm{S}}_0,{\mathrm{A}}_0\right]$$

(11)

In Algorithm 2, the proposed QRLM algorithm for facilitating learning in the mist layer is presented. The algorithm considers the setups discussed in Section 4.7 to deduce the working principle for the model. The healthcare dataset considered in this study is initially processed using the QRLM framework, which considers processing the dataset in the mist layer by exploiting the Markov property. The QRLM algorithm considers five tuples, namely $〈S,A,P,R,\delta 〉$, to model the considered healthcare dataset. The probability distribution for mapping the states and actions incorporated by the learner over the data is achieved through $\mathrm{p}\left({\mathrm{A}}_{\mathrm{k}}|{\mathrm{S}}_{\mathrm{k}}\right)$. The Q-value and the value function are further computed by recursively solving the expressions in Equations (8)–(11) for $\mathsf{\phi }$.

Algorithm 2 QRLM Algorithm

function QRLM$\left(S,A,P,R,\delta \right)$ Input: $S,A,P,R,\mathsf{\delta }$ Output:$V_{\mathsf{\pi }}\left(s\right),Q_{\mathsf{\pi }}\left(s,a\right)$ Initialize: $S\leftarrow$Set of states $S=\left\{s_1,s_2,\cdots s_n\right\}$ $A\leftarrow$Set of actions $A=\left\{a_1,a_2,\cdots a_n\right\}$ $P\leftarrow$Probability of transitioning the present state $R\leftarrow$Set of rewards $\mathsf{\delta }\leftarrow$Reward discount factor Compute:      $\mathsf{\pi }\left(\mathsf{\alpha }|\mathbf{s}\right)=\mathbf{p}\left({\mathbf{A}}_{\mathbf{k}}|{\mathbf{S}}_{\mathbf{k}}\right)$ for $\mathsf{\phi }$ do if $\mathsf{\phi }~\mathsf{\pi }$ and ${\mathbf{A}}_{\mathbf{k}}~\mathsf{\pi }\left(.|{\mathbf{S}}_{\mathbf{k}}\right)$ do      ${\mathbf{V}}_{\mathsf{\pi }}\left(\mathbf{s}\right)={\mathbf{{\rm E}}}_{\mathsf{\phi }~\mathsf{\pi }}\left[\mathbf{R}\left(\mathsf{\phi }\right)|{\mathbf{S}}_{\mathbf{0}}\right]$      ${\mathbf{V}}_{\mathsf{\pi }}\left(\mathbf{s}\right)={\mathbf{{\rm E}}}_{{\mathbf{A}}_{\mathbf{k}}~\mathsf{\pi }\left(.|{\mathbf{S}}_{\mathbf{k}}\right)}\left[{\displaystyle \sum _{\mathbf{k}=\mathbf{0}}^{\mathbf{\infty }}}{\mathsf{\delta }}^{\mathbf{k}}\mathbf{R}\left({\mathbf{S}}_{\mathbf{k}},{\mathbf{A}}_{\mathbf{k}}\right)|{\mathbf{S}}_{\mathbf{0}}\right]$      ${\mathbf{Q}}_{\mathsf{\pi }}\left(\mathbf{s},\mathbf{a}\right)={\mathbf{{\rm E}}}_{\mathsf{\phi }~\mathsf{\pi }}\left[\mathbf{R}\left(\mathsf{\phi }\right)|{\mathbf{S}}_{\mathbf{0}},{\mathbf{A}}_{\mathbf{0}}\right]$      ${\mathbf{Q}}_{\mathsf{\pi }}\left(\mathbf{s},\mathbf{a}\right)={\mathbf{{\rm E}}}_{{\mathbf{A}}_{\mathbf{k}}~\mathsf{\pi }\left(.|{\mathbf{S}}_{\mathbf{k}}\right)}\left[{\displaystyle \sum _{\mathbf{k}=\mathbf{0}}^{\mathbf{\infty }}}{\mathsf{\delta }}^{\mathbf{k}}\mathbf{R}\left({\mathbf{S}}_{\mathbf{k}},{\mathbf{A}}_{\mathbf{k}}\right)|{\mathbf{S}}_{\mathbf{0}},{\mathbf{A}}_{\mathbf{0}}\right]$ end if end for return ${\mathbf{V}}_{\mathsf{\pi }}\left(\mathbf{s}\right),{\mathbf{Q}}_{\mathsf{\pi }}\left(\mathbf{s},\mathbf{a}\right)$ exit

4.8. Experimental Setup

▪

Samsung Galaxy S7 with Android 9 is taken as the gateway device;

▪

Dell XPS 13 with Intel (R) Core (TM) i5-7200 CPU as the broker node;

▪

Worker Node: Raspberry Pi 3B+, ARM Cortex-A53 quad-core SoC CPU @ 1.4 GHz as a worker node;

▪

Microsoft Azure, 1 GB RAM, as a mist node.

4.9. Data Set

To determine if the patient suffers from heart illness or not [44,45,46,47], represented by an integer value of zero (no presence) or one, we took into account the data of heart patients (presence). Andras Janosi (M.D.) at the Gottsegen Hungarian Institute of Cardiology in Hungary and others carried out this experiment using the Cleveland database [44]. Both the patients’ names and patient IDs are kept private.

To determine the state of the patient’s health, we used the following 14 crucial data attributes:

• Age: the person’s age in years;
• Sex: a binary value (1 = male, 0 = female);
• Cp: the type of chest pain;
• Resting cardiographic data (0 for normal);
• Cholesterol (in mg/dL);
• Blood sugar (>120 mg/dL);
• Blood pressure at rest (mmHg);
• ST elevation (0.05 mv);
• Maximum heart rate;
• ST depression;
• ST segment slope;
• Colored fluorescence;
• Heart disease target;
• Angiography status (value between 0 and 1).

5. Results and Discussions

5.1. Performance Measurement Characteristics

To measure the performance of the model in the mist layer, the following characteristics were observed and analyzed:

• Correctness: 30% of the dataset’s data were used for testing, while 70% were used to train the model. To create the corresponding trained deep-learning framework, all the worker/broker nodes received an equal distribution. Since the volume of mist nodes increases, it would be necessary to send the dataset examples to every node to utilize all training resources. This not only cuts down on training time but also improves exam accuracy. The training and test accuracies were examined to see any such impacts. The proportion of all patients for whom this framework correctly predicts whether or not they have heart disease is known as accuracy. With different mist conditions, we changed the number of edge nodes and whether or not the results are assembled to test accuracy levels.
• Time factor: Arbitration time, latency, execution time, and jitter are examples of the time factor. We compared these temporal characteristics for several mist configurations, such as cloud-only computing architecture, up to two edge nodes (with or without ensembling), and no edge nodes.
• Utilization of network bandwidth: This was examined to ascertain network use in various scenarios because the scenario (broker alone, workers, or cloud) and the number of worker nodes impact network consumption. Like the testing for time parameters, we compared the network bandwidth used for the various mist scenarios. This was done to ascertain how bandwidth utilization and the different mist configurations that HealthMist offers relate to one another.
• Energy consumption: Because energy is such an essential factor in the transition from cloud to mist layer, we looked into energy usage in various environments. We looked at the different HealthMist setups used for diverse individuals.

5.2. Parameter Comparison and Discussion

In prior work, the power of FogBus was presented, along with comparisons to the cloud frameworks, illustrating how FogBus enables more efficient implementation of applications that utilize edge and cloud resources. This work used the MistBus framework to develop a heart patient monitoring application considering delay and reliability with technical simplicity and in a short time to employ edge and cloud resources properly. Depending on that user’s needs, the application deployment system supplied several settings with higher accuracy or latency. Depending on the experimental findings, we suggest that HealthMist be employed in the following setup depending on the desired applications.

Figure 6 demonstrates how training accuracy improves by the number of mist nodes (broker as well as worker nodes). We can see that as the number of worker nodes rises, training accuracy also gradually rises. As every node tries to build a framework for the information it receives, the count of examples collected by each node decreases as this volume of nodes increases. As a result, training the model over multiple epochs causes the samples to fit the model better, which increases training accuracy.

Figure 7 demonstrates the change in test data accuracy as the number of mist nodes increases. As every node receives a lower subset of training data and cannot generalize the model; as a result, test accuracy decreases with increasing node count. Another finding is that ensemble learning almost always results in significantly higher accuracy than using it alone (best or average).

Figure 8 displays the variation in arbitration time near the broker node for various mist circumstances. One broker, one worker node, two worker nodes, and fog are the only configurations. We observed that the time required for arbitration is small when the job is transmitted straight to the master, broker, or fog (almost 115 ms). The arbitration time increases as the number of edge nodes increases because this broker must look at the loads at each worker node to determine which worker has the lowest load and to which to send a job. The broker does not need to do load checking when data are delivered to worker nodes for ensemble learning because one of these worker nodes must make the majority class choice; therefore, arbitration time is comparable to the situation without resembling.

Figure 9 demonstrates the fluctuation in latency, which, according to Figure 3, is the sum of the waiting time and communication time. As single-hop data transfers do all communication, we can see that the latency is essentially the same whether the job is delivered to the broker or any of the edge nodes. The ensemble cast’s delay has slightly increased. The extraordinarily high latency for cloud settings is caused by the several times that data must be sent outside the LAN.

The change in jitter with the fog setups is shown in Figure 10. When jobs are dispatched to worker nodes instead of just the broker, we see that the Jitter is larger. This is due to the broker’s added responsibilities, which include managing resources, arbitrating disputes, and verifying security. As the number of workers increases, the difference in worker loads causes the jitter for two edge nodes to increase compared with one edge node increasing slightly. The jitter is also crucial in ensemble situations. The jitter is extremely severe if jobs are routed to the fog data center.

The alteration in the execution pattern is depicted in Figure 11. Because more resources are available in a fog configuration, the execution time is much less than anticipated. Because HealthMist workers are Raspberry Pis, that is, having CPUs with low clock frequencies, broker execution time is faster than worker nodes. The execution time is also longer if ensemble prediction is active since this worker node now must determine whichever class is the majority among all predicted types.

Figure 12 shows the variation in network bandwidth utilization across all edge nodes for various rundowns. We have found that when this number of worker nodes rises, traffic similarly rises because data transfer, security checks, and heartbeat packets (with fog) are necessary. When using an ensemble, all of the data are sent. Worker nodes use the most bandwidth on the network. Figure 13 represents various scenarios; we also examined HealthMist’s energy consumption characteristics, especially in comparison with broker nodes (microcomputer) or worker nodes; the fog data center uses much more energy (Raspberry Pi). As a result, the fog example consumes a lot more power than the mist case.

6. Conclusions and Future Work

The smart healthcare assistant is a vast and sensitive project. In the current research work, our main objective was to focus on the healthcare system for heart patients and proposed a fog-computing-enabled smart healthcare system that incorporated the latest technology, i.e., a deep-learning ensemble technique that provides an automated diagnosis of heart disease with the utilization of IoT-enabled resources. HealthFog integrates deep learning in edge-computing devices and deployed it for a real-life application of heart disease analysis. Prior works for such heart patient analysis did not utilize deep learning and hence had very low prediction accuracy, which renders them useless in practical settings. Deep-learning-based models with very high accuracy require very high computing resources (CPU and GPU), both for training and prediction. This work allowed the network with a complex deep-learning model to be set in edge computing standards that use the correspondence and model dispersion methods so as to ensure that high latency can be met with low latency. As part of our future work, we propose to extend HealthFog to allow cost-optimal execution given different QoS characteristics and fog-cloud cost models. Currently, HealthFog works with file-based input data, which can be converted to be seamlessly integrated to take data directly from sensors to make it user-friendly. Moreover, the model training strategy used currently uses separate training at each worker node. The trained models at each node have been combined using various ensemble models of bagging. More intelligent ensemble models with advanced security features can be deployed to further improve the accuracy.