AI augmented edge and fog computing for Internet of Health Things (IoHT)

Motivation

Edge and fog computing have recently become crucial in IoHT. Artificial intelligence (AI), ML-based techniques and algorithms have also been used in a number of elements of smart healthcare system, including load balancing, energy efficiency, and resource management (Kennedy & Eberhart, 1995; Masdari et al., 2017; Houssein et al., 2021; Karaboga & Basturk, 2007; Rani, Ahmed & Rastogi, 2020; Canali & Lancellotti, 2019; Knowles & Corne, 2000; Goudarzi et al., 2021). For this reason, we intend to focus on topics like edge and fog computing in healthcare models, various issues when combining fog and edge computing in IoHT, the significance of AI in IoHT, numerous challenges, and the significance of machine learning-based resource management algorithms to address these challenges. Also in a fog- and edge-based healthcare architecture, resources can be managed and apps can be deployed using AI and ML-based resource management schemes and algorithms. A comprehensive survey is required to identify the potential concerns and future challenges posed by advanced technologies which inturn improve overall healthcare.

The intended audience

Literature search

To guarantee thorough and objective coverage of the literature, a multi-step, methodical strategy was used in the search and selection of publications. To make sure there was a wide range of material available, Google Scholar was used to aid in the initial search.

Popular keywords

Popular terms were chosen to capture the most pertinent studies from the relevant topic on AI augmented Edge and Fog computing for IoHT. Some of the keywords were: “IoHT,” “Fog computing in IoHT,” “Edge computing in IoHT”, “Machine learning,” “Deep learning”, and “ML algorithms for IoT”. These keywords were chosen to capture every aspect of the subject under evaluation.

Refinement of the search

The search was narrowed down to the research domain by rearranging and combining the terms in different ways. To guarantee uniformity in comprehension and elucidation, only English-language articles were taken into consideration.

Selection of the articles

After receiving a lengthy list of publications, their abstracts were carefully examined to ascertain their applicability and value in relation to the subject of the review that we have chosen. Innovation, methodological coherence, and a noteworthy addition to the body of current knowledge were among the selection criteria. Initial search was made in Google Scholar. A list of recent journal information was collected from it and based on that further search was made in the respective journals like IEEE, Elsevier, ScienceDirect etc. Recent manuscripts were selected from the above mentioned journals and shortlisted based on the published year and on their relevance to our topic.

Bias mitigation

Within the parameters of relevance and quality, the selection and search process was designed to be as inclusive as possible in order to reduce any biases. To determine the scientific validity of each article on the shortlist, its approach was thoroughly scrutinized. This was done in order to make sure that the literature review presents a fair and impartial assessment of the current research, without favoring any one methodology or point of view over another.

Final list of articles

A final selection of article was made following this methodical procedure in order to facilitate a thorough and objective examination and analysis.

Search process

Various article from different journals were carefully selected and filtered. Figure 1 gives an overview of the various references collected in different year spans and Fig. 2 gives an overview of different categories of references selected for the review process.

Research questions

The Research questions we used to investigate are as follows

RQ1: What are the various ML and DL techniques used in edge-based and fog based healthcare Models?

RQ2: What are the various challenges for adapting fog and edge computing in the IoHT environment?

RQ3: How various ML-based resource management algorithms help in improving the effectiveness of different environments?

RQ4: How can high security and low delay in communication be achieved in an IoHT framework with the help of current technologies?

All these research questions have been answered in the subsequent sections with appropriate references and illustrations.

The cloud layer

Even though cloud computing provides ample resources and services to conduct intricate analysis, the closest regional cloud facility may be hundreds of kilometers away from the data collection location. The cloud is the centralized IoT deployment platform, as cloud providers offer a variety of pre-built services for IoT operations (Zhang et al., 2021).

The fog layer

The fog layer helps IoT-based systems overcome latency challenges because processing must occur on the fly. Using the fog layer, alerts can be issued in real time, anomalies can be detected, and necessary actions can be taken automatically. It supports local data storage by bringing computational power nearer the edge, decreasing the total response time of the system (Karakaya & Akleylek, 2021). Fog computing settings can produce large amounts of sensor or IoT data spread over vast areas, too large to define an edge (García-Magariño et al., 2019). The fog layer performs operations such as preprocessing sensor data, fusion, data analysis, compression of collected data, filtering, and reducing the load on the cloud. The fog layer improves system performance, bandwidth utilisation, and system Quality of Service (QoS) (Chudhary & Sharma, 2019). Fog computing devices are located next to edge devices, while edge devices are positioned at the network’s edge (Omoniwa et al., 2019; Chakraborty et al., 2017). Latency-aware applications can be implemented using fog computing (Tuli et al., 2020; Daraghmi, Wu & Yuan, 2021; Naha et al., 2018; Bhatia & Kumari, 2022).

Also, the fog layer modifies the task management method which is based on a priority queue and a list of fog nodes as in Karakaya & Akleylek, (2021).

The edge layer

Using processing and storage resources in close proximity to the data source is known as edge computing. Ideally, this deploys computers and storage near the data source at the network’s edge. Devices like embedded automation controllers now have intelligence and processing capabilities because of edge computing. Optimising resource utilisation is just one of the many advantages that edge computing has over traditional architectures. Reducing network traffic and overcoming data bottlenecks are two other benefits of computing at the edge (Varghese et al., 2016).

Protocols for communication

As shown in Fig. 4, short-range communication protocols like IEEE 802.15.1 or 802.15.4 are used for communication between a device and a fog node. An 802.11 wireless protocol connects a sensor node to additional computing devices or cloud services by utilising a sensor node, mobile computing devices, and a cloud service. Bluetooth, or IEEE 802.15.1, is used in several applications as a communication protocol between a medical device and a smartphone that provides calculations. Once the smart device has completed a brief computation, the data is transmitted to a doctor or another server via a mobile connection service such as 4G or 5G (Hartmann, Hashmi & Imran, 2022). Commercially accessible products like the Raspberry Pi, Arduino, and field programmable gate array (FPGA) platforms are used for manufacturing edge gateways; because of their low cost and ease of use, these solutions are quite popular.

The communication protocols are mainly of two types request-reply and publish-subscribe. Message exchange in client server architecture is based on request-reply category. Representational state transfer hypertext transfer protocol (REST HTTP) and common offer acceptance protocol (CoAP) are the two main common protocols fall in this category (Dizdarević et al., 2019). For distributed, loosely coupled communication between data source and destination publish-subscribe protocol is preferred. Data distribution service (DDS), advanced message queuing protocol (AMQP), and message queuing telemetry transport (MQTT) are some of the examples for publish-subscribe protocols. Many works have been carried out to compare the effectiveness of these protocols in fog-to-cloud IoT architecture. Based on latency, bandwidth consumption, energy consumption and security, MQTT and hypertext transfer protocol (HTTP) are considered as most stable protocols to be used by developers for fog, cloud and IoT implementations (Dizdarević et al., 2019). AMQP together with REST HTTP is the main protocol used between IoT and fog layers. Also RESTful HTTP and DDS protocol proposed in Dizdarević et al. (2019) can be used in all layers. Mostly lightweight protocols are preferred between IoT and fog layer but that restriction is not there for communication between fog and cloud Dizdarević et al. (2019). The challenge is portability and interoperability between the protocols. But in many works combination of protocols like MQTT and REST HTTP, REST HTTP and CoAP proved more efficient (Dizdarević et al., 2019). The usage of CoAP and HTTP protocols in IoT, fog, and cloud layers based on client server architecture is shown below in Fig. 5.

Fog based healthcare model

Three critical layers make up the fog computing model for healthcare. The edge, fog, and cloud layer are the layers. End devices like sensors, medical equipment, mobile phones, etc., make up the edge layer. The edge layer has billions of edge devices and all the data they produce is transferred to the fog layer to be processed. It consists of fog devices. Fog machines can perform various activities such as data separation and data analysis, and even deep learning modules can be run in the fog machines (Di Biasi et al., 2021) that help analyse the data and make predictions. Emergency information can be sent to the edge devices after analysing fog, as this reduces the overall response time. By using a fog layer between the edge and the cloud layer, latency is significantly reduced. Only data that needs to be stored permanently and is less sensitive is stored in the cloud. An example of fog-based diagnostics using mobile devices as an edge device is shown in Fig. 7. According to the figure, the mobile app receives the patient’s data and sends it to the relevant fog node for processing. The segregation and data analysis take place there in the fog node, and the information is stored locally in the fog. Only permanently stored information is sent to the cloud. The routers are connected to each other via Local Area Network (LAN) and to the Amazon EC2 Cloud via Wide Area Network (WAN). The “Wireless Access Point” (WAP) function is also integrated into the routers. This allows access to both the Amazon EC2 Cloud and the fog nodes via mobile and smart devices (García-Magariño et al., 2019; Sood & Mahajan, 2018; Vijayakumar et al., 2019; Bhatia & Kumari, 2022; Chudhary & Sharma, 2019).

Edge based healthcare model

Wearable devices such as smartwatches, tablets, phones, and many other embedded systems perform low-level processing and act as edge devices. Edge computing moves some of the data away from the central data centre and stores it close to where the data originated. As shown in Fig. 8, artificial intelligence tasks and data analysis techniques are performed by the edge devices, but the majority of the tedious processing and analysis is performed by the backend machine learning core, which in turn applies useful algorithms for deep analysis (Tuli et al., 2020; Zhang et al., 2018; Abdellatif et al., 2019; Carvalho et al., 2021; Hassan, Yau & Wu, 2019; Ketu & Mishra, 2021). For cloud-based computing models, data transfer speed has become a bottleneck as the amount of data created at the edge rises. The reaction time will be excessively slow if all data must be transferred to the cloud. The patient’s data is received by the mobile app and routed to the relevant fog node, as depicted in the figure for processing. In such situations, the data should be processed at the network interface to reduce response time and network load (Firouzi et al., 2023). Data is processed near the original source in the case of edge computing. Many data mining techniques are performed in the edge layer, which speeds up execution and makes results more accurate. Local data processing takes place at the edge. Since multiple local authorities are involved, the congestion problem is reduced and reliability is increased. IBM’s Watson IoT platforms have been deployed in edge computing models to analyse the analytical results based on the data generated (Li et al., 2019). As the edge devices can only perform fewer calculations, edge-based healthcare models are preferred when the system’s computational overhead is lower and data security is of high importance. Complex calculations are carried out in the fog layer or cloud.

General architecture for edge computing

A user device, a sensor or Internet of Things device, a computationally capable smartphone, and an edge, fog, or cloud computing node make up the general architecture of an edge computing paradigm as shown in Fig. 8. It depicts the basic architecture of edge computing. An important component of the design is the interaction between the edge and the cloud. While the emphasis of the intervention is on speed, the benefits of cloud computing are realized in terms of long-term data (Hartmann, Hashmi & Imran, 2022).

Edge computing includes three layers, namely the terminal, border and cloud layers. By integrating edge devices between the cloud and endpoints, the design brings cloud services to the edge (Cao et al., 2020). Terminal Layer: Mobile phones, IoT devices, and several other devices are included in this layer. The gadgets in the terminal layer act as both data providers and data consumers. The top layer receives and stores all raw data from the terminal layer for processing. Boundary layer: The boundary layer is the most important component of the three-layer architecture. The edge layer is positioned between end devices and the cloud and is made up of edge nodes at the network’s edge. Cloud devices, with their sophisticated computational and storage capabilities, can handle a wide range of data management, analysis, and business decision support tasks (Cao et al., 2020; Abdellatif et al., 2019; Carvalho et al., 2021; Hassan, Yau & Wu, 2019). Variation of response time based on the increasing system load in the cloud-only, edge-only, and fog-only setups is depicted below in Fig. 9. An analysis of some of the related works has been made in Table 1. The works have been analyzed based on the technologies involved, and the techniques used. The technologies include cloud computing, the Internet of Things, edge computing, and fog computing, and the techniques include simulation, alert generation (exists or not), real-time analysis.

Overview of IoHT

Basic stages in IoHT

IoHT for patients includes a range of wearable devices, including fitness bands, smartwatches, and other wirelessly compatible gadgets (such as heart rate monitors, blood glucose monitors, blood pressure monitors, etc.). These state-of-the-art devices are used to set reminders for things like daily calorie intake, activity tracking, blood pressure changes, appointments, and more. Various wearables and sensor-based gadgets are available to monitor the patient’s health when discussing IoHT for hospitals. In case of an urgent medical situation, notification is sent to the doctor and the family members of the patient. IoHT also makes it simple for nurses to track medical supplies such as oxygen tanks, nebulizers, defibrillators, wheelchairs and other equipment (Sneha & Varshney, 2007; Milenković, Otto & Jovanov, 2006; Shnayder et al., 2005; Ketu & Mishra, 2021). The basic infrastructure of IoHT consists of these four steps as shown in Fig. 10. An intelligent IoHT system should be built to collect or process data at one point and generate value at the next. The first step involves deploying a heterogeneous network and devices such as cameras, sensors, displays, and actuators, which are obtained in the second step. The digitized and aggregated data are pre-processed and standardized in the third stage, which is followed by the transfer of refined data to the cloud. The final step involves conducting advanced analytics on the processed data to enable better decision-making (Alamri et al., 2013).

Taxonomy of IoHT

Biomedical sensors or other sensor-based technologies gather crucial physiological data from patients, which is then utilised to identify the various solutions during the processing and communication stages. Figure 11 depicts the four major levels that comprise IoHT systems: the sensor layer, the network layer, the processing layer, and the application layer.

The sensing layer

In IoHT systems, the sensing layer is the foremost layer. The process of acquisition involves a number of factors that includes sensor type, parameters, and method. Single-parameter and multiple-parameter collection can be done according to the needs of intelligent healthcare systems. Perception methods include active and passive perception. Active sensing is the process of obtaining ECG data whereas passive sensing is the process of estimating heart rate using the ECG information (Ketu & Mishra, 2021).

The network layer

The second layer of IoHT systems is the network layer. It helps with both communication and the security of the system. The selection of security mechanisms and the appropriate communication protocol play an important role at this level. For heterogeneous communication, it is important to have a standard protocol for communication. Scheduling is the method we use to control time-based resource communication. This is a vital task of the IoHT system. Data management, job scheduling, and concurrent control are the three core components of the planning task. The two types of scheduling jobs are combined scheduling and publish/subscribe scheduling. While combined scheduling uses a synchronous method, publish/subscribe scheduling employs an asynchronous method. In this layer, security can be ensured at the user level and network level (Ketu & Mishra, 2021).

The processing layer

The third IoHT layer that deals with computation and data management is the processing layer. Storage type, processing medium, and data integration are all critical components of data management. The computational specification is critical in determining the fundamental technique that is integrated and used in applications based on IoHT. Some features, includes real-time remote patient monitoring, and assisting medical professionals in providing patient care. The processing layer may include deep learning or machine learning module to process the input data from sensors and aid in quick diagnosis. The computational capacity of healthcare system has increased thanks to its real-time tracking and monitoring capability. Large-scale and sophisticated computing operations can be completed by integrating with cloud computing (Ketu & Mishra, 2021; Di Biasi et al., 2021).

The application layer

Next to processing layer comes the application layer. Control, actuation, and application are the important aspects of the application layer. An alarm system for patient monitoring is the best example of control/actuation in which when the input signal hits a specific threshold value the alert is activated. The alarm is triggered when the input signal reaches a particular threshold value. IoHT-based systems find use in senior care, smart homes, and hospitals. The requirements determine the application’s level of complexity. It may be an application designed for beginners or for experts (Ketu & Mishra, 2021).

Requirements for adapting fog computing in IoHT

The computing nodes are heterogeneous in nature and distributed in the fog layer. Safety and reliability are important factors to consider when using the fog layer (Mahmud, Kotagiri & Buyya, 2018). Some of the requirements for adapting fog computing in IoHT are listed below.

Structural requirements

The fog infrastructure consists of different components with different configurations. Therefore, in fog computing, using the right strategies for information exchange between nodes and effective resource allocation is crucial (Mahmud, Kotagiri & Buyya, 2018; Iftikhar et al., 2023). The number of layers in the fog framework should be limited depending on the requirements because the number of layers affects the complexity of the network.

Service related requirements

Some of the fog nodes do not support the application requirements because they are not all resource-enriched. Therefore, fog should have an appropriate programming environment for creating distributed applications. Workload balancing, resource provisioning, task scheduling, and resource allocation are some of the service-related challenges. Appropriate policies need to be developed to divide tasks between IoT devices, fog nodes and the cloud. Service Level Agreements (SLAs) in fog are influenced by many factors, and it is difficult to set SLAs in the fog layer (Mahmud, Kotagiri & Buyya, 2018).

Security requirements

Although edge-enabled medical equipment helps patients live better lives and opens up new revenue streams for 5G network operators and healthcare providers, there are serious privacy concerns that will only intensify as these devices are widely used. Current Health Insurance Portability and Accountability Act (HIPAA) regulations are not well-established enough to be applied to cutting-edge healthcare monitoring techniques. The healthcare provider and the network operator would both be subject to legal repercussions in the event of a data breach because many research institutes and insurance firms regard patient information as a valuable commodity (U.S. Department of Health and Human Service, 2017). Patient data retention laws and policies differ by nation and area, which makes matters more difficult (Casola et al., 2016). It is difficult to ensure security in the fog layer because there are more opportunities for attacks. The process of authentication and authorization becomes more difficult at this level. Security mechanisms can affect the QoS (Amin & Hossain, 2021). Areas of concern include access controls, sufficient bandwidth, user adoption, security, lack of visibility, and availability.

The integrity of the medical images is achieved with the help of a hash. Hash is more powerful than the techniques for encryption DES and AES which stands for data encryption standards and advanced encryption, respectively (Rana, Mittal & Chawla, 2020). Patient data should be treated with a high level of security. To further secure the images, reversible data hiding (RDH) and piecewise linear chaotic map techniques were used (Yang et al., 2019). The use of firewalls, virtual private networks (VPN), encryption and masking increases security (Joshi et al., 2021). To secure communication between sensors, a key generation scheme based on multi-biometrics in wireless body area networks (WBANs) was used (García-Magariño et al., 2019). The security of communication between sensors is one of the main factors to be considered in IoT-based healthcare systems. For communication between sensors, several biometric concepts were used to generate a common key. Although attribute- and identity-based encryption helps maintain data confidentiality in the cloud, the problem arises when updating the encrypted document. In addition, the transfer process from local to global leads to data leakage problems when using the cloud (Yang, Xiong & Ren, 2020). The question of the method by which users can securely search for specific keywords in data encrypted in the cloud has been explained in Yang, Xiong & Ren (2020) and Schuiki et al. (2019). Many IoT security related issues in the domain of Internet of Medical Things (IoMT), Internet of Vehicles (IoV), and intrusion prevention system (IPS) has been discussed in Javed et al. (2023).

Requirements for efficient resource management and accuracy

Resource estimation, resource discovery and resource matching are included in resource management (Iftikhar et al., 2023). Resource utilization, resource load, resource lifetime, response time, delay or latency, energy consumption, reliability and security are some of the metrics of fog computing (Aslanpour, Gill & Toosi, 2020). The difficulty of managing computation at a network’s edge is a challenge faced by most applications worldwide. ML has made significant contributions to fog computing and edge processing. Fog computing would benefit greatly from the appropriate implementation of machine learning (Abdulkareem et al., 2019). Many factors such as application context, energy and time influence resource provisioning. Time is described in terms of calculation, communication and deadline. Computation time is the time required to complete a task. This depends on the resource configuration. Resource and power management in fog is done using task calculation time, which shows active and inactive periods of applications. Communication time defines the time required for communication in a fog environment. Communication time helps select appropriate fog nodes by defining the network context. And the deadline is the maximum service delay that a system will tolerate. Application placement plays an important role in the fog framework. To meet the desired QoS criteria, the request should be routed to the most suitable fog node based on the request and the fog resource availability (Abdellatif et al., 2019).

Requirements for processing the big data

In a smart environment, IoT devices produce a lot of data. Processing this data at the edge requires real-time data analysis. Fog computing and machine learning/deep learning effectively support this process. In order to solve the problem of processing resources and reduce cloud usage costs, big data is reduced in the fog layer (Abdulkareem et al., 2019). Transmission latency decreases from 2.84 to 0.13 s when performing distributed computing on a smartphone. Some approaches focus on using data selection to determine whether data is transferred to a server or the cloud for further processing.

Need for the breakdown of fog devices

This happens for many reasons such as software crashes, hardware failures and end-user activities. Multiple fog nodes are connected via wireless networks or Wi-Fi and the Wi-Fi connection is known to be unreliable (H & Venkataraman, 2023). The comparison between cloud, fog, and edge computing based on different parameters is shown in Table 2.

Requirements to overcome the safety issues

In Zahid et al. (2018), many reported data breaches are due to machine theft, manual errors, hacking, ransomware, and abuse, suggesting that basic security measures are still inadequate. Additionally, various hacking techniques such as ransomware, brute force, use of backdoors or C2, operating system commands, forced browsing, spam, and spyware/key loggers were highlighted. Deep Neural Networks, or DNNs, have been used in the fields of image forensics, malware classification, steganalysis, and child abuse investigations in cybercrime investigations. However, no prior research has examined the robustness of DNN models in the context of digital forensics. In order to verify the security robustness of black-box DNNs, a domain-independent Adversary Testing Framework (ATF) was built in the article by K, Grzonkowski & Lekhac (2018). By using ATF, a commercially available DNN service used in forensic investigations has been tested, where published methods fail in control settings. In the article Zargari & Benford (2012), key aspects of digital forensics including forensic data collection, elastic, live and static forensics, evidence separation, proactive preparations, and investigations in virtualized environments in cloud forensics were discussed. Failover, instance relocation, Man in the Middle (MITM), server farming, Let’s Hope for the Best (LHFTB), and address relocation techniques as well as various research challenges related to cloud forensics were discussed.

Requirements for integration between fog nodes and the cloud

The fog framework contains both the cloud and the edge. To execute storage and processing in a distributed environment, cloud architecture and device heterogeneity should be taken into account (H & Venkataraman, 2023). Dynamic mapping of end-to-end fog and cloud integration should be given more importance in any fog computing architecture.

Requirements for adapting edge computing in IoHT

There are numerous challenges associated with controlling peripheral layer devices, includes the heterogeneity of their computational and storage properties, connection, and the enormous number of devices that must be monitored at the same time. Response time, delay or latency, security, energy consumption, availability and reliability are some of metrics of edge computing (Aslanpour, Gill & Toosi, 2020). Nevertheless, mobility and device location are currently the biggest obstacles. Because the edge node position is malleable, the methods and protocols for controlling the node network must be developed in line with the physical edge location. Furthermore, management protocols must be designed to maintain a secure and stable system, with a particular emphasis on four factors: distinction, extensibility, isolation, and reliability (DEIR). As the number of devices and their sensing, computing and communication capabilities continue to increase, bandwidth saturation, energy constraints and associated costs will necessitate new strategies, routines and real sustainability awareness. The future deployment of peripheral devices will depend on network virtualization technologies (Villar-Rodriguez et al., 2023).

Need for collaboration of edges and heterogeneous information

Healthcare requires the collaboration of information between different stakeholders in different areas. To solve this problem, collaborative edge is used (Abdellatif et al., 2019). It improves overall efficiency and reduces latency and communication costs. The edge network receives input from heterogeneous end devices. To enable automatic monitoring and remote monitoring, multimodal data processing techniques must be incorporated to combine information sources. Challenges in the heterogeneous environment include transmitting informative bio signals such as EEG and electromyography (EMG), which consume more energy. The other reason is the fluctuations in the generated signals due to noise, signal offsets and other interference. To overcome this challenge, the multi-access edge computing (MEC) model was used in the S-Health system. It solves the problem of heterogeneous inputs by incorporating multimodal in-network processing methods that emphasize the temporal correlation within each modality as well as the correlation between several modalities. Using MEC, a mobile edge node (MEN) sends only limited functionality to the cloud instead of raw data. Advanced signal processing can be performed at the edge to overcome the problem of artifacts (Abdellatif et al., 2019).

Privacy and security requirements

The highest potential of a healthcare system can only be achieved when the user has confidence in the security of their stored data and health-related information. The first is ownership of the data. One of the best solutions is to protect the data where it was collected and keep it close to the patient, allowing the patient to take responsibility for their own data (Abdellatif et al., 2019). The trade-off between security and QoS is the second challenge. Increased security levels increase computational complexity at the edge even as they increase security (Menaka & Ponmagal, 2018; Omotosho, Emuoyibofarhe & Meinel, 2017; Masood et al., 2018). Therefore, it is important to develop common security and QoS mechanisms that increase both the efficiency and security of the overall system (Abdellatif et al., 2019).

Integration of blockchain technology in smart healthcare for security

Need for integration of edge computing with 5G

Although 5G’s ultra-low latency and increased bandwidth will improve edge computing, the full impact is not yet clear (Carvalho et al., 2021). 5G networks require significantly faster methods of data storage, decision-making, data collection strategies and remote data management, for which pre-processing needs to be improved. Because of 5G networks, the number of IoT devices will expand, which require more data processing, security, and computing resources. Additionally, managing credentials in IoT environments essentially involves huge volume of data. And the collaboration of multiple and diverse entities ensures ubiquitous computing. Innovative mobility options for communication devices could eventually become a target for network attackers. With network function virtualization (NFV) and software-defined networking (SDN), edge computing has become a key technology of the 5G era (Carvalho et al., 2021; Bernardos et al., 2014). Additionally, the convergence of EC with 5G will enable increased response time and computing power, ignoring the physical location and resources of the IoT. The characteristics of 5G present many challenges and increase the complexity of deployment (Carvalho et al., 2021). According to Hassan, Yau & Wu (2019), six important functions for EC in 5G networks are improving security, local data analysis, processing, decision-making, and computation, all performed locally. Edge nodes should also be able to provide some of the capabilities of the cloud.

Mobility requirements

Mobility increases user and application flexibility but also has significant disadvantages. Mobility has a substantial impact on the loss, latency, and bandwidth of connections between edge devices and the edge network, which negatively affects the overall quality of service. User mobility can have a big impact on how many hops there are between a user and their services, especially if that movement happens at the network borders. To adapt to such network changes, edge services must be replicated quickly and dynamically. In addition to the received signal strength, the optimisation algorithms must take into account the movement direction, cost-benefit analyses of alternative networks, and service quality requirements. As the bar for resource availability, resource discovery, workload offloading, and resource provisioning is raised, the mobility of edge nodes must also be considered (Yousefpour et al., 2019).

Reliability requirements

The cloud is used by edge computing devices to improve reliability. Storing data or running apps on cloud servers reduces the risk of data loss on mobile devices (Dinh et al., 2013). Solving issues such as platform, network, user interface, network coverage, individual device failure, and deploying edge devices without support of the cloud plays a very important role. Challenges in implementing such a system also include issues with mobility, device availability, battery limitations, connection fluctuations, and device availability. Additionally, check pointing and rescheduling methods in centralized systems do not function properly at the edge due to their mobility, heterogeneity, dynamics, and latency (Nath et al., 2018).

Need for storage models

Due to problems with data ownership, data format, application programming interfaces supplied by data owners and computing platforms, controlling storage characteristics across all heterogeneous edge devices in edge cloud systems is difficult. Specific data models corresponding to the above concepts could be created for each level of the system to solve this problem. Its implementation is difficult because it must take into account a variety of factors, including automatic data consistency, device discovery, data sharing policies across nodes at different tiers, communication, and network protocols. The transversal data model must be automatic, visible to the user, and accessible at every level, regardless of the data, its location, or its load (Wang et al., 2019). Table 3 gives a brief comparison between our work and different surveys based on different technologies like IoT, fog, edge, 5G, architecture, protocols, security, ML/DL techniques and algorithms used.

Integration of AI with 5G for effective smart healthcare and the associated challenges

AI is used in smart healthcare systems to analyse enormous volume of medical data and to deliver real-time communication and allow for the smooth integration of several procedures in healthcare (Pradhan et al., 2023). The introduction of 5G communication has greatly increased the opportunities for optimizing healthcare, providing low latency, fast data transfer, and the capacity to connect a wide range of systems and devices with unmatched dependability. High-speed networking, real-time data processing, and decision-making intelligence work together to provide enhanced remote monitoring, personalized treatment, better patient care, and efficient use of resources. In Pradhan et al. (2023), effective integration of 5G–Multiple-input Multiple-output (MIMO) technology along with AI in an IoT environment has been discussed. Also the performance of the edge server in 5G network has been analysed based on various parameters. To guarantee the appropriate and efficient application of new technologies, however, considerable thought needs to be given to the ethical, privacy, and security considerations. In Pradhan et al. (2023), some of the challenges linked to the integration of AI with 5G in IoT environment has been discussed. It includes scalability and network capacity issues, interoperability issues and lack of standards, energy efficiency issues, security and privacy problems, QoS and latency issues. Though there are few advantages, 5G technologies still faces heterogeneity and scalability issues of IoT. Also adding 5G technology to an IoT environment introduces additional security and privacy issues.

In order to provide seamless and high-performance connectivity, additional study is required to optimise QoS provisioning in 5G networks for Internet of Things applications. This includes taking into account aspects like latency reduction, traffic prioritisation, and network optimisation. In addition, research is needed to develop economic models, inorder to evaluate how cost-effective 5G-enabled Internet of Things installations are, and look at new business models that take into consideration the particular needs and limitations of various Internet of Things applications (Pradhan et al., 2023). The use of new advancements like chat-bot which is user friendly is a contribution of AI (Mah, Skalna & Muzam, 2022).

Edge-fog-cloud computing and blockchain technology for improving NB-IoT-based health monitoring system

The capacity and power consumption of user devices, particularly in deep coverage, can be greatly improved by integrating SHM systems with narrow band-Internet of Things (NB-IoT), which also enables covering additional categories of IoT devices and services (Daraghmi et al., 2022).

In order to improve NB-IoT security and authentication, Daraghmi et al. (2022) presents a blockchain framework that is applied to a health monitoring system that serves patients in rural and village locations. There are five primary layers in the framework: The master nodes (edges), the blockchain layer, the off-chain layer, the NB-IoT health monitoring devices, the device gateway, and so on are the first five layers. The proposal makes use of both on-chain and off-chain layers since the on-chain data contains all the necessary information to control the suggested smart contracts that are integrated into the blockchain to enhance security, authenticate users, and enable efficient data access while maintaining data integrity. Raw NB-IoT health data from verified master nodes is stored in a cloud-based database called the off-chain data.

Some of the problems faced in NB-IoT are as follows,

1. The significant latency resulting from the lack of delay-tolerant techniques in the majority of NB-IoT frameworks. Furthermore, the massive data sizes generated by the multiple terminals or the healthcare applications—like high-definition images—have an effect on the transmission time because the NB-IoT depends on the user datagram protocol (UDP) to transfer small-sized data in real time. Because the healthcare industry is so vital, it’s crucial to be concerned about delays even if having a huge data set is necessary for fast throughput.

2. Security concerns since patient privacy must be preserved and sensitive patient data must be handled with care. The NB-IoT network is easily breached by hackers.

   Furthermore, the UDP protocol, which is used by the NB-IoT, is open to attack (Daraghmi et al., 2022).

To overcome the above problems, a hierarchical architecture of the network which is divided into three layers: edge, fog, and cloud computing has been proposed in Daraghmi et al. (2022). The system and edge devices can communicate with each other through NB-IoT. Basic classification analysis, classification-based prioritization and an authentication procedure are all included at the edge layer. Advanced analysis, task management, and security are all included in the fog. Security and long-term data analysis are included in the cloud layer. NB-IoT has been used as the primary means of communication between edge devices and other computing layers due to its ability to cover a huge number of devices in large areas while consuming the least amount of power. Between the edge and the fog layer is the base station.

Sorting and ranking data has been carried out to reduce base station congestion and shorten NB-IoT transmission delays. Several IoT authentication protocols has been analysed for secure NB-IoT transmissions and the most effective one has been selected. CloudSim, iFogSim and ns-3-NB-IoT are the simulators used. Authentication protocols like random MAC (RMAC), Light-Edge, and enhanced authentication and key agreement (AKA) has been tested. Access delay for with-edge and without-edge cases and execution time for no-edge, no-fog, no-edge no-fog, and the proposed architecture has been analysed. Also the authentication time taken by RMAC, Light-Edge, AKA protocol in different cases like no-edge, no-fog, no-edge no-fog, and the proposed architecture has been analysed (Daraghmi et al., 2022).

Artificial intelligence

The health sector benefits greatly from the wide variety of applications of artificial intelligence and machine learning which include prediction, classification, and analysis. For example, BlueDot Inc. (Toronto, Canada) created BlueDot: Outbreak Risk Software, which was essential in reducing the SARS-COV-2 viral outbreak (Sheth et al., 2024). A great deal of patient data can be analyzed using machine learning to find patterns and useful information. Thus, these cutting-edge technologies have the ability to anticipate and lessen catastrophic pandemics like COVID-19 and raise awareness in order to stop widespread breakouts (Sheth et al., 2024). An artificial intelligence based model was employed to monitor critical care patients during the pandemic. It is a three-stage model including input, process, and output. Within these three subparts they further divided the managerial and clinical tasks. Clinical, para-clinical, personalized medicine, and epidemiological data made up the input stage. Expert systems, deep learning, machine learning, neural networks, and artificial intelligence are all included in the process stage. Intensive care unit (ICU) decision-making, diagnosis, treatment, risk assessment, and prognosis comprised the final output stage (Sheth et al., 2024).

Deep learning in IoHT

Ahmed et al. (2022) presented an IoHT-driven, deep learning-based system for non-invasive patient discomfort detection. The new method uses an RGB camera device to measure a patient’s level of discomfort instead of wearable sensors or devices. In order to recognize movement and discomfort, the system uses the YOLOv3 model to detect the patient’s body and the Alphapose technique to extract numerous keypoints information of the patient’s body. The identified keypoints are then transformed into six central bodily organs by applying association rules. Pairwise distance between identified keypoints is used to measure the level of pain in patients. Time frame criteria are then applied to define a threshold and examine motions to determine whether they are indicative of normal or uncomfortable conditions. Monitoring of patient’s body is done in multiple frames for distinguishing abnormal conditions from normal (Ahmed et al., 2022).

Advanced attention and transformer models

Blockchain for enhanced security in IoHT

The initial components of any IoT system centered around healthcare are sensors, RFID, and smart tags, all of which have resource constraints. Safe information transformation is crucial when these devices are integrated since they include patient data that is sensitive and might be very harmful if it ends up in the wrong hands. Blockchain-based IoT is a new technology that protects data at the physical layer by fusing the advantages of cryptography and blockchain. In Jayaprakash & Tyagi (2021), elliptical curve cryptography has been used to safeguard data in blockchain network where the data is encrypted and stored.

Application placement algorithms

The fog service placement problem can be used to determine the best arrangement between IoT devices and fog nodes in order to utilise fog resources as efficiently as possible (Ghobaei-Arani, Souri & Rahmanian, 2020). Selimi et al. (2019) provides a micro-cloud infrastructure for community networks. Certain scenarios prioritise application placement requests to streamline application deployment to fog resources and classify the resources (Mahmud et al., 2019). The article explains the module mapping approach for deploying IoT application modules in the fog (Ghobaei-Arani, Souri & Rahmanian, 2020). To deploy the workload, the resources that are available are chosen based on the workload requirements. Fog servers must perform as many tasks as they can to use up as many resources as they can. To achieve this, one method is to use a manager or master entity. After that, the resources are distributed between fog and cloud according to the requirements of the workload (Mijuskovic et al., 2021). The purpose of the detection algorithms is to identify edge resources that can be used to properly distribute processing power between the cloud and fog layers in the future (Javaid et al., 2019; Taneja & Davy, 2017).

The shortest job first algorithm

It executes the tasks according to their highest priority, which is determined by measuring the size of consumer requests (Kaur & Kinger, 2014). The distribution of jobs is based on the regions, fogs and shortest sizes of the VMs. This method is preferred over other methods due to its low latency. The waiting time for the consumer is reduced and their comfort level is not affected. First, the requests with the smallest size are fulfilled, then the appropriate VMs are allocated. The jobs are distributed among different VMs by the scheduler and the scheduling of the jobs is done by shortest job first (SJF) by allocating the minimum completion time. Additionally, it offers faster response times and greater efficiency. Following is the pseudocode for the SJF algorithm.

for i = 0 to i < main queue size

   if the length of task i + 1 < the length of task i then

      add task i + 1 before task i in the queue

   end if

   if main queue size = 0 then

     task i is the last in the main queue

   end if

end for

In the above pseudocode, the length of the tasks are compared and the task with lower execution time is given higher priority and moved forward in the queue. The average response time and average performance time of six different fog setups at six different locations using the ESCE, RR, and SJF algorithms are shown in Figs. 12 and 13, respectively (Javaid et al., 2019). And the virtual machine costs, micro-grids and data transfer costs for the above three algorithms are shown in dollars in Fig. 14.

Offloading algorithms

Offloading focuses primarily on resource provisioning rather than computation. It serves to minimise the overall costs. Latency is caused by determining data storage and reducing transmission costs between the cloud and IoT devices (Wang et al., 2019).

Load-balancing algorithm and techniques

Load-balancing algorithm (LBA) is used to manage the workload distribution in the network to prevent overload, low utilisation, and congestion which in turn improves the efficiency of operations. Some of the LBA algorithms are listed below (Xu et al., 2018).

Hill climbing algorithm

Hill climbing algorithm (HCLB) is a mathematical search optimisation algorithm. Accessible VMs are found using the random solution. The loop of this algorithm runs continuously until the ideal answer to a problem is found. In HCLB, the loop is extended until the closest reachable VM is identified. The best VM is then selected, and requests are sent to it for processing (Zahid et al., 2018).

Figure 15 gives overview of some of the existing algorithms for efficient application placement, offloading, load balancing along with various nature inspired algorithms for resource management.

Nature-inspired algorithms

The service placement problem is used to find a mapping that assigns application components of the cloud to fog nodes, thereby achieving an optimized fitness value. It is similar to matching problems. A fog node can accept a variety of application components when one or more fog nodes are assigned to an application. This assignment can be many-to many. Application limitations, fog resource limitations, and network limitations are some of the limitations. The assignment of components is referred to by mapping, which describes whether an application component is assigned to a cloud or fog node or not. And the decision variables used are binary. For large fog and edge networks that are complex, heuristic or metaheuristic techniques are used to solve the problem. Applications running on IoT devices are interdependent. Directed Asymmetric graph (DAG) is used to model IoT applications that consist of interconnected components that represent a connected graph (Kumar et al., 2022).

Sensors, Internet of Things devices, smartphones, and other end-user devices with limited resources offload computationally intensive activities to remote servers to complete specified tasks (Wang et al., 2020). In edge technologies, computing resources are often heterogeneous, distributed, and resource-constrained with a dynamic topology and limited bandwidth linking these resources. The offloading problem is difficult and non-trivial due to the nature of processing and communication resources (Aazam, Zeadally & Harras, 2018). Few nature-inspired techniques are available to overcome computing offloading and service placement problems.

Swarm intelligence-based algorithms

These algorithms were designed using the collective intelligence of species including ants, birds, honeybees, and insects. There are different types of swarm algorithms available but those that can be applied for computation offloading and service provisioning are discussed below.

Particle Swarm Optimization (PSO): The steps involved in the PSO technique are shown below in Fig. 16. A viable solution to the issue that has been optimized in the N-dimensional search space is called a particle in PSO (Kennedy & Eberhart, 1995). Particles travel in accordance with their global best positions attained by all other particles as well as their local best positions. PSO is a potent optimization technique with a wide range of applications (Masdari et al., 2017). It does, however, have several flaws, such as sliding into local extremes and failing to guarantee accurate solutions when there are many local extremes (Houssein et al., 2021).

PSO-based scheduling: The dimension of the particle in PSO-based workflow planning is the number of tasks, and each position of the particle shows the association between the tasks and virtual machines. For example, the particle for eight tasks is shown in Figs. 17A and 17B, which describe the mapping between five VMs and eight tasks. T1, T2, …., T8 are the tasks, and VM1, VM2, …, VM5 are the virtual machines.

In Wang & Brown (2021), PSO algorithm has been used in the model that detects alcoholism in patients. Using MRI of the brains, scientists have found that compared to normal, healthy individuals, brain of alcoholic patients tend to have less gray matter and white matter. On the basis of this, computer-aided diagnosis strategies for alcoholism detection have been presented in the recent past. In contrast to techniques such as support vector machines (SVM) and CNN, a unique structure for alcoholism detection has been proposed in this article. The proposed structure used a single-hidden-layer neural network trained via PSO as the classifier and gray level co-occurrence matrix (GLCM) as the feature extractor. The proposed structure not only showed a strong performance using experiment datasets, but it also outperformed other strategies in terms of simplicity and speed.

Ant Colony Optimization (ACO): Finding the best routing paths, selecting the right cluster heads, and load balancing are just some of the problems ACO is currently addressing in the IoT space (Kumar et al., 2022). The steps carried out in the ACO technique are shown above in Fig. 16.

Edge mining for big data management in healthcare

Real-time data collection coupled with a large-scale healthcare system ensures that there is a lot of data to analyse and secure. Processing sensory data close to or at the point of perception in order to transform it from a raw signal into contextually relevant information is known as edge mining, or data mining at the edge. Edge mining techniques that reduce the volume of data transported to cloud considerably solve this problem partially; however, more reduction is vital for long-term and continuous data acquired by medical sensors. Often, rather than having to be reduced, this data can be examined in bulk. It can even be up to exobytes in size. This requires the development of new data-feature-based analysis methods (Hartmann, Hashmi & Imran, 2022).

Integration of 5G for mobile edge computing-enabled health care

Current network solutions have significant hurdles when it comes to huge-scale, shared, sensor-based medical monitoring and reporting. The success of numerous supporting technologies is necessary for the transfer of IT and telecom services from a central cloud platform to an edge. Using virtualization techniques such as VMs and containers is one of the essential components. Unlike VMs, container solutions like Docker provides minimum virtualization on consumer devices which helps edge computing devices (Ismail et al., 2015). Similar to this, NFV separates services and network operations from proprietary hardware enabling several service instances to run concurrently on a single VM and reducing overall costs. In cases of overload due to flash crowd events, the operator of a MEC-based healthcare system can use NFV to shift system processes between two edge platforms (Bernardos et al., 2014). Another important technology is the use of SDNs. SDNs, or software-defined networks, are a key technology. The main idea underlying SDN is to separate the control and data planes and create a logically centralised control that allows many virtual network instances to be launched and deployed to users at the edge. In existing network topologies, managing the dynamic provisioning of various services at the network edge is challenging. SDN are anticipated to play a significant role in network connectivity and administration of services across multiple MEC systems (Bernardos et al., 2014; Taleb, 2014). Additionally, network slicing allows a single network to be divided into numerous instances, each tailored to a specific application or use case (Taleb et al., 2017; Afolabi et al., 2017). Various 5G network segments, including mobile broadband, vehicle connections, and extensive IoT may exist. Due to the high capacity requirements of enhanced mobile broadband of 5G, several other related technologies that have been introduced in the Radio Access Network (RAN) would make it possible to provide more bandwidth, pipeline packet processing, fewer TTIs, and effective radio resource control (RRC). These enabling technologies include things like user-centred designs, massive MIMO (mMIMO), millimetre wave spectrum (mmWave) transmission, and others (Hashmi et al., 2018; Hashmi, Zaidi & Imran, 2018; Cetinkaya, Hashmi & Imran, 2018; Bai & Heath, 2015; Andrews et al., 2017; Rappaport et al., 2013).

Need for integration of AI with 5G for the growing need

The next-generation wireless cellular network (5G) is anticipated to meet the needs of next-generation networks. The following are the three main issues that the networks of the previous generation failed to address. First of all, the growing use of mobile devices generates enormous amounts of data. Second, in order to support highly interactive applications that require extremely low latency and high throughput, tight QoS criteria must be met. Third, a heterogeneous environment must be supported to ensure interoperability between a variety of user devices (e.g., smartphones and tablets), QoS requirements (e.g., different levels of latency and throughput for multimedia applications), network types (e.g., IEEE 802.11 and the Internet of Things), etc. (Hassan, Yau & Wu, 2019). 5G is able to solve the majority of these problems. 5G networks will use high-frequency millimetre waves to strengthen communications. Mobile users would be connected to a base station that is connected to the core network via a common 5G mobile network. With low-latency radio interfaces, communication between the end user and the base station would be possible in less than a millisecond, but latency would increase significantly if cloud-based application requests were routed from the core network to the cloud. Fog computing can help shift computing power closer to the end user which is required for 5G networks. Real-time interaction, high data rates, high availability, and local processing are four essential conditions for edge computing to be successfully designed and implemented in 5G. While the rollout of 5G enables widespread MEC-based healthcare infrastructure, in order to offer clients the most pertinent and timely services, it is essential to integrate AI.

To enable rapid diagnosis of health issues, artificial intelligence employs a range of criteria, including user mobility patterns, device utilisation patterns, critical tracking of information from patients, and pre-existing medical disorders. Recent advances in machine learning, particularly deep learning, have enabled advancements in a variety of fields: recognition of faces, healthcare diagnosis, and processing of natural languages are a few examples (Lakhanpal, Gupta & Agrawal, 2015). However, they require an enormous amount of storage space and computing power and require the complex processing of large amounts of data in both central and decentralised data centers. Because the entire concept of moving processing to the edge is based on ultra-reliable low-latency communications (URLLC), distributed, trustworthy, and low-latency edge ML models built on local data. Edge ML offers low cost and minimal latency for mission-critical patient-facing IoT sensor devices. Any combination of the three basic fields of machine learning (supervised, unsupervised, and reinforcement learning) can be employed in a distributed healthcare system with an AI-integrated 5G infrastructure.

A recent survey provides more information about these strategies and how they relate to edge platforms (Park et al., 2019). AI can be deployed at the edge by partitioning helpers and devices. Each device generates a local learning model and transfers it to a helper that gathers all models from multiple devices (Dean et al., 2012; Jin et al., 2016). If one device runs out of memory, a local model can be partitioned and spread among multiple devices. In this case, the intermediate model is moved back and forth between the devices during training. The application of AI and edge computing in 5G includes remote surgery, diagnosis, data and vital sign monitoring for patients. And the doctors can operate remotely (Hassan, Yau & Wu, 2019).

Integration of blockchain technology for enhanced security and privacy

Blockchain technology creates an unchangeable record of transactions that is spread among participants in a ledger to guarantee secure and dispersed transactions. A block containing the transactions is connected to the chain. We clarify blockchain technology, which combines three existing technology such as consensus, distributed ledger, cryptography, and protocols (Ismail, Materwala & Zeadally, 2019).

Integration of IoT with unified communications as a service

Unified communications as a service (UCaaS) is a cloud-based delivery paradigm that provides a wide range of communication and collaboration applications and services. This is an enterprise communication system in which all communications are streamlined via cloud delivery, allowing the industry to be more flexible in terms of human-financial resources and spending. Virtual apps (MS Teams, Zoom, Skype, YouTube, and WebEx) and mobile system apps (video, audio, SMS, and chat) are examples of UCaaS (Mah, Skalna & Muzam, 2022). Devices can be integrated with IoT sensors to transmit alerts or messages directly to communication channels, automate conversations, or initiate video on communication conferences based on sensor data or events using middleware to translate IoT protocols into unified communication systems. Integration of IoT and cloud-based technologies such as UCaaS has brought great transformation (Mehmood et al., 2024). Communication between remote or hybrid environments are made possible with UCaaS on a single platform. Scalability can be increased with the help of UCaaS which will cater to growing business needs. UCaaS solution enables an organization to integrate multiple devices such as laptops, desktops, and mobile phones into their solution. Following the development of UCaaS, many organisations with good hardware sought cheaper, faster, and more convenient alternatives to system integration rather than constructing their own (Son et al., 2019).

Some of the advantages of using UCaaS are reduction in cost of purchasing hardware for communication purposes, more scalability in communication, speed of implementation, availability of services regardless of geographical locations, improved communication and employee productivity, increased technical support by the service providers, unified management of communication systems, and low demand of staff. Though there are many advantages few challenges still exist. Some of them are –

Security-related barriers which includes loss of control over valuable data, unauthorized disclosure of data to competitors, UCaaS non-compliance solutions based on the security policies that have been implemented.

Technical barriers which includes lack of internet connection, network failure, problem in data migration, etc.

Legal barriers which includes problem in negotiating legal terms with the service providers, risk for non-compliance with the guidelines of the service providers.

Psychological barriers which includes reluctance of the employees to implement new communication and collaboration methods due to a lack of confidence, being accustomed to the present communication and collaboration systems, lack of information, and understanding about the importance of UCaaS. Range of solutions available with UCaaS has been depicted in Fig. 18.