A systematic review on artificial intelligence approaches for smart health devices

Artificial intelligence techniques

Artificial intelligence is a field of computer science that aims to develop systems capable of performing tasks that normally require human intelligence (Winston, 1984). These tasks include reasoning, learning, sensory perception, understanding natural language, and interacting with the surrounding environment. The goal of AI is to create machines and algorithms that can emulate or surpass human capabilities in specific domains.

It is divided into two large classes (Wang & Siau, 2019). The first represents weak artificial intelligence, also known as specialized AI, which refers to systems designed to perform specific tasks without necessarily understanding or imitating human intelligence in a general way. Examples include search engines, recommendation systems, and gaming algorithms. The second is strong artificial intelligence which aims to reach a level of intelligence comparable to human intelligence. In contrast to weak AI, strong AI should be able to tackle any intellectual task. However, currently, strong AI is still largely a future goal and subject to debate.

There are four main generic approaches, ML and DL, natural language processing, which focuses on the ability of computers to understand, interpret, and generate human language, and computer vision which is the ability of machines to interpret and analyze visual data, often used in applications such as facial recognition, robotic vision and medical image analysis (Dargan et al., 2020; Adam & Mukhtar, 2024). AI is an ever-evolving discipline with a growing impact on multiple industries, including healthcare, industrial automation, transportation, marketing, and many others.

More specifically, the application of AI in medicine is revolutionizing the healthcare sector, offering opportunities to improve the diagnosis, treatment, and management of patients. The main application areas are medical diagnosis where machine learning algorithms, particularly convolutional neural networks (CNN), are used to analyze medical images such as X-rays, MRIs and computed tomography scans to identify abnormalities and aid in the early diagnosis of diseases like cancer; clinical decision assistance, where machine learning algorithms are used to provide personalized recommendations to doctors, based on clinical data and test results, helping in choosing the most appropriate treatment options; treatment personalization where AI helps identify subgroups of patients with similar genetic or molecular characteristics or analyze historical patient data to predict treatment responses and optimize treatment protocols; and finally healthcare management and continuous monitoring, where AI facilitates the provision of remote medical care, allowing continuous monitoring of patients, management of chronic diseases and virtual medical consultancy, and introduces wearable devices that integrate with algorithms of AI can collect data on vital signs and report any anomalies, improving proactive health management (Kavitha et al., 2023; Al Kuwaiti et al., 2023).

Smart health

E-health refers, generically, to the various processes of digitalization of healthcare and, therefore to the creation of electronic infrastructures for the administrative and not just clinical management of each citizen’s data, smart health focuses more on IoT devices, including smartphones, which implement those same processes and personalized models for monitoring and treating health conditions (Tian et al., 2019).

Smart health is the basis of the so-called 4P medicine: Preventive, Participatory, Personalized, and Predictive. It is a medicine capable of preventing the disease based on the intersection of the different biological profiles of the person, through a simple blood sample and/or the processing of data from IoT devices (Aversano et al., 2022a; Chang, 2023). A medicine capable of making each person no longer a “patient” but aware and involved in their state of health thanks to real-time communication with their doctors. But also to personalize diagnosis and treatment (Aversano et al., 2021; Adam & Mukhtar, 2024) based on the risk factors detected and the information received from the wearables and to predict the severity of the risk factors of pathology before it manifests itself.

Smart health serves to automate procedures and simplify bureaucracy, contain costs, reduce times, optimize flows, make people aware of their health condition through the transparency of data, immediately consultable, and develop “person-friendly” health paths. It is essential for monitoring the quality of air, water, and building safety.

Smart health allows you to be assisted almost everywhere, realizing the famous transition from expectant medicine to initiative medicine, or from an emergency model to a proactive model, of reticular, “care-intensive” assistance, which sees the hospital as a high-intensity centre for acute pathologies, and the “low-intensity” territorial facilities as structures that manage chronic conditions, prevention and social-welfare services (Aversano et al., 2022b; Mshali et al., 2018).

Wearable devices

The term wearable devices refers to a wide range of intelligent devices that can be connected to other electronic devices such as smartphones, wirelessly or via Bluetooth technology, allowing not only the detection but also the storage and exchange of data of different like, immediately and often without the need for human intervention. Increasingly advanced, wearable devices are used for various purposes: from the detection of biometric signals, through sensors positioned on the skin, to quick access to online information (Iqbal et al., 2021; Devi et al., 2023; Tariq, 2024). Among these, we find wearable devices such as heart rate monitors, glucometers, or sphygmomanometers which are now able to offer a valid contribution when used in the healthcare sector. In fact, through sensors, actuators, and software connected from the device to a smartphone or tablet with the cloud, intelligent wearable devices worn by patients allow the collection, analysis, and real-time transmission of personal health data (Babu et al., 2024; Cui, Du & Wu, 2023).

Wearable healthcare is based on the collection of real-time or continuous 24-h data on a person’s state of well-being and physical fitness (Cho, 2019; Escobar-Linero et al., 2023). This data allows you to create a personalized health database. Anyone who wears a smart device capable of detecting health data can therefore use it to allow a healthcare worker to collect their data to obtain remote monitoring and to achieve health, fitness, well-being, and physical fitness objectives that are not necessarily related to the control of pathological conditions.

In healthcare, wearable devices such as healthcare smartwatches are often used to monitor patients’ health conditions and related emergencies, allowing them to be recognized as soon as they occur. The use of wearable technology therefore creates the potential for a proactive approach to healthcare. This can help detect symptoms of an illness before they develop into larger problems and, therefore, can have dangerous health consequences. Even people with known pathologies can benefit from early identification of signs of possible anomalies.

A typical example of a smart (i.e., intelligent) device for health monitoring is smart patches (Segev-Bar, Konvalina & Haick, 2015; Wong et al., 2023). These are thin stickers containing electronic components (including sensors and actuators with appropriate processing, energy storage, and communication capabilities) that collect a patient’s vital signs: heart rate, body temperature, and more, and transfer them to appropriate apps for doctors and patients.

Another example is a mobile ECG monitor (Baig, Gholamhosseini & Connolly, 2013). This device records the electrocardiogram via a wireless electrode that can be applied, for example, to the patient’s chest or finger. The ECG data is sent to the cloud and, then, to a doctor for appropriate evaluation.

Therefore, wearable technologies can represent innovative and highly effective solutions for patient health, while also offering enormous potential benefits to healthcare workers and the entire healthcare system (Tariq, 2024). But, to improve the usability and functions of the devices for practical use, issues such as acceptance by all users, security, ethics, and issues relating to the management of the big data generated still need to be addressed. from wearable devices. There are also elements of concern that arise from a large-scale wearable implementation, such as aspects of data privacy and security, but also supply chain management (Ahmed et al., 2024). When it comes to wearables and security, the medical Internet of Things opens the door to new attack vectors, including targeted malware and distributed denial of service threats (Shanmugam & Azam, 2023). Therefore, to avoid privacy and security issues, it is necessary to establish robust authentication and authorization processes for collecting, sending, and analyzing data.

Goal of the study

The focus of our work on systematically reviewing applications of AI in smart health is motivated by the growing importance of this research area and its significant implications for improving healthcare services and patient well-being. With the increasing adoption of digital technologies in the healthcare sector, the use of AI is becoming increasingly relevant to address complex challenges related to the diagnosis, treatment, and management of diseases. Therefore, our work aims to provide a detailed and up-to-date overview of AI applications in this emerging context.

The evolution of AI technologies, such as ML and DL, has opened up new opportunities to develop innovative and personalized solutions in the healthcare sector. By exploring the applications of these technologies, we aim to identify best practices and emerging trends in smart health. Understanding how AI can be used to optimize early diagnosis, improve chronic disease management, optimize care processes, and provide decision support to healthcare professionals is crucial to promoting the quality and effectiveness of healthcare services.

Our main objective is to conduct a systematic review to thoroughly explore the applications of AI in smart health. Through this review, we aim to not only evaluate the effectiveness and accuracy of existing AI solutions in improving clinical outcomes and optimizing operational efficiency in healthcare services but also to examine the current challenges and limitations in implementing AI in smart health-related privacy issues that need to be addressed.

Research questions and queries

A systematic review summarizes existing work allowing the completeness of the research to be assessed. The main advantage is that it provides information on the effects of the phenomena under observation. About that the purpose of this study is to explore the state of the art in integrating AI with IoT wearable devices in the field of smart health. In this regard, four research questions were defined.

First, the choice of AI methodology to apply to data collected by wearable IoT devices is crucial to the success of smart health applications. In this context, it is crucial to understand the complexity and depth of the AI models used, as well as the amount of data required for their correct functioning. The complexity of AI models, such as DL, involves the use of deep neural networks with many layers and parameters. This complexity allows the model to capture and learn sophisticated patterns in the data, improving predictive and analytical capabilities, but it is widely known that DL models require large amounts of data to train effectively. In contrast, traditional ML methods can work well even with smaller datasets. These methods are often less complex and require less data to provide accurate and reliable results. Therefore, in this study we aim to evaluate which is the most suitable approach for wearable IoT applications in the field of smart health, addressing our first research question:

RQ₁: To what extent research studies adopt ML or DL approaches on smart wearable data?

The field of DL offers a wide range of architectures, each with specific characteristics and advantages. Data collected by wearable IoT devices presents unique challenges, such as temporal variability, the need for real-time analytics, and managing large volumes of data. It is essential to identify which DL architectures are best suited to make the most of this data and provide accurate and useful results. Examining which DL architectures are most commonly used in studies and applications related to wearable IoT devices helps identify trends and best practices in the field. In this regard, we address our second research question:

RQ₂: Considering the scopes of application, which Neural Network architectures are more effective to support IoT Smart Health ?

In healthcare applications, the accuracy of algorithms is critical. Clinical decisions based on incorrect analyses can have serious consequences for the health of patients. Therefore, it is crucial that the algorithms used can provide accurate results. At the same time, the reliability of the algorithms ensures that the results are consistent and reproducible. In a healthcare context, where decisions must be based on robust and reliable data, trust in the results produced by algorithms is essential. Therefore, we address the third research question:

RQ₃: Which algorithms lead to better performances while exploiting Smart Health data collected from IoT device monitoring?

Finally, health and medical data is extremely sensitive and personal. This data includes detailed information about individuals’ physical and mental health, which, if exposed, could cause significant harm, such as discrimination, invasion of privacy, and reputational damage. The management of healthcare data must comply with strict regulations, which impose high standards of data protection. Addressing our fourth research question allows us to identify and implement best practices and technologies to protect sensitive data, thereby improving the security and effectiveness of smart healthcare applications.

RQ₄: RQ4: What privacy and security issues have taken into account while using health data collected from IoT devices?

Once the main study directions had been identified, the above-mentioned research questions were transformed into a specific query. This query was used to examine the selected databases, as described in the next paragraph, to find studies to be considered in the review. Below is the query used:

Q: (Wearable) AND (Artificial Intelligence) AND (Smart Health)

Databases

In this paragraph we report the databases that we queried with the query just described, to extract the articles on which our analysis focused. Below are the databases:

• IEEEXplore (https://ieeexplore.ieee.org/Xplore/home.jsp): a digital library and research database for discovering and accessing journal articles, conference proceedings, technical standards, and related materials in computer science, electrical engineering, electronics, and related fields.

• ScienceDirect (https://www.sciencedirect.com/): which provides access to a large bibliographic database of scientific and medical publications from the Dutch publisher Elsevier.

• The ACM Digital Library (https://dl.acm.org): a repository of published resources in the computing field maintained by the Association for Computing Machinery.

Research and filtering process

Having defined the research directions and the databases from which we have extracted the studies to be considered for the analysis, we can now tackle the search and filtering process that we have conducted. Specifically, this is shown in Fig. 1. It can be noted that the process consists of three substantial phases. The first consisted of applying the query to the chosen databases, while the second involved initial filtering of the selected documents based on the defined inclusion (IC) and exclusion criteria (EC), reported in Table 1. Finally, the third phase consisted of the actual analysis of the studies considered for the review.

As shown in Table 1, the search process involved the selection of articles published in the year range from 2014 to 2023 (IC₂) from the databases listed in “Databases”, using the keywords of inputs consisting of queries (Q) defined in “Research Questions and Queries”. This time horizon was chosen because, before 2014, the literature lacked relevant publications regarding ML or DL approaches in smart healthcare. This first phase of the search produced a total of 292 articles, of which 107 came from IEEEXplore, 121 from ScienceDirect and 64 from ACM.

The research work continued with the filtering process through (i) reading the titles, (ii) scanning the abstracts, and (iii) removing duplicate documents, thus arriving at a total of 105 selected articles. The selected studies had to fully meet the eligibility criteria in the Table 1.

In particular, they had to be written in English (IC₁), published in the period 2014–2023 (IC₂), they had to be focused on an IoT scenario (EC₂), was to be specifically focused on DL or ML approaches (EC₄) in the smart health domain (IC₃) and was to be neither a survey nor a review (EC₁). Furthermore, the search was limited to articles published or in press (IC₁), thus excluding conferences and preprints. This decision was guided by the quality standards we aimed for in our systematic review. We considered that preprints have not been peer-reviewed and that solid and valid conference contributions are usually published in journals in more detail. Therefore, this final filtering step allowed us to retain a total of 58 articles to analyze across the defined research questions.

For completeness, Fig. 2 presents the temporal distribution of the works definitively considered. This shows the years on the x-axis and the number of contributions on the y-axis, showing the total number of contributions with the blue bars, and the percentage with the orange bars. As can be seen from the figure, the number of articles addressing our research topic has become significant since 2017 and has increased over the years, except for 2019.

Furthermore, in Fig. 3, we show the distribution of the articles considered per publisher, always reporting the name of the publisher in abscissa, the number in the ordinate axis. In particular, the blue bars consider the number of articles published by the editor, instead of the orange ones percentage. The analysis of the figure indicates the predominance of the articles of the IEEE publisher related to the topic of our research.

RQ1: To what extent research studies adopt ML or DL approaches on smart wearable data?

Different methodologies and approaches can be used to analyze health data. Among these, ML and DL algorithms and networks are the best-known and most used in the literature. There are the most disparate ML and DL techniques and there is also no shortage of customized implementations of basic algorithms, which makes the range of possible solutions truly wide.

For this reason, first of all, it is important to investigate which methodology is most used in the selected studies to determine the complexity and depth of the model with a distinction between ML and DL approaches or to investigate whether the study investigates both methodologies, therefore considering, in this case, the number of ML/DL algorithms evaluated.

Figure 4 summarizes the absolute number of articles (in blue) and mass percentages (in orange) for each analysis method used in the study (ML, DL, or both algorithm types) plotted on the x-axis. It can be observed that the majority of articles (57%) concern ML methodology, followed by DL algorithms (33%). However, only in 8 studies (10%), the two approaches are used together.

In Fig. 5 we report the boxplot of the distributions of the number of algorithms used in the selected articles, specifically the distribution of the ML classifiers evaluated in the selected articles is highlighted in blue boxplot and that refers to the DL algorithms is highlighted in orange one. In particular, Fig. 5A shows the distributions of the number of ML/DL classifiers evaluated in the studies that only used the ML or DL approach; Fig. 5B presents the distributions of the ML/DL classifiers evaluated in the studies that used both methodologies.

The analysis of the distributions of the number of ML/DL classifiers evaluated in the studies that used only of approach shows that in the majority of articles using ML algorithms, it is preferred to compare multiple classifiers of this type, at least more than $3$ (median of the blue distribution). Instead, in studies where DL methodologies are used, it is preferred to use a single network selected according to the nature of the data. This choice is assumed to be made because DL algorithms require longer execution times and more computational memory than ML classifiers. Even in the case of articles that used the two approaches together, more ML classifiers are used than the DL approach which remains at the level of single network used.

According to this research question, we also explored some aspects related to the data set used in the selected articles. Firstly we examine the possibility of accessing the data, specifically if the data set used is public or private/personal. Figure 6 displays the data access policy adopted by the authors. The majority of the articles analyzed ( $67$%) operate with private/personal data sets, while $33$% of them analyze public data that in the majority of the cases are publicly available or can be released on request.

Relating to the nature of the data used in the considered articles, we also investigated the number of instances/patients involved in collecting the data. Additionally, we also scrutinized the number of features used to train the models used in the collected articles. In Table 2 we disclose the results of this investigation. In particular, in it, we have reported in line with the confidence interval of the average number, the modal value, the minimum and maximum value assumed by the number of instances/patients involved in the data collection (first column), and the features used in the model training (second column).

Referring to the results obtained in Table 2, we notice that for both analyses the minimum and maximum values differ greatly and for this reason, the mean and its relative confidence interval are extremely high; this is also the reason why we decided to study the modal value of the two distributions. In particular, taking this last value into account, we note that in general, the number of instances/patients in the articles is around $40$ cases while the number of features used to train the models in the majority of the articles is $16$.

RQ2: Considering the scopes of application, which neural network architectures are more effective to support IoT smart health?

Neural network architectures play a significant role in IoT Smart Health by enabling the development of applications that improve health monitoring, diagnostics, and overall well-being of users. Furthermore, DL architectures, together with appropriate data pre-processing and model optimization techniques, contribute to the advancement of IoT smart health applications, enabling early disease detection and better management of chronic conditions. In the documents analyzed, we found that there are some key neural network architectures used in IoT smart health concerning specific types of data:

• CNNs are widely used for image-based diagnostics for example analysis of medical imaging data (X-rays, MRI, CT scans), and real-time video monitoring (e.g., for patient movement analysis). They help in identifying patterns, anomalies, and potential health issues from visual data.

• RNNs are often used for time series analysis of health-related data, such as continuous monitoring of vital signs (heart rate, blood pressure)and patient activity tracking.

• Long short-term memory networks (LSTMs), an extension of RNNs, are valuable for processing and analyzing sequential health data, such as medical history, and time-stamped sensor data.

• Densely connected convolutional networks (DenseNet) are a CNN architecture that connects each layer to every other layer in a feed-forward pattern. They are efficient for medical image analysis, especially when dealing with limited data or for transfer learning tasks.

• Dense neural networks (DNNs) are often used in IoT Smart health for remote patient monitoring. DNNs can analyze data from wearable devices, sensors, and other IoT-enabled medical devices to continuously monitor a patient’s vital signs, detect anomalies, and provide real-time alerts to healthcare professionals for early intervention.

Concerning the pathology analyzed, we report in Fig. 7 the distribution of the articles we considered according to the disease having been tackled.

In particular in Fig. 7 the blue bars take into account the number of studies, while the orange bars report the the percentage; the same figure also shows the table with the respective values. The results of this analysis show that the most studied diseases are cardiac and neurodegenerative diseases (24%), followed by lung diseases (20%), following mental diseases such as depression and panic attacks (12%).

In Fig. 8 we have reported the distribution of the pathologies analyzed only with a DL approach. Also in this figure, we have displayed the number of studies in blue and the relative percentage in orange; at the bottom of the image we report the table of the relative values of both. This analysis shows that complex approaches such as DL are most used when dealing with heart and neurodegenerative disease data (20%), but also for mental health and lung disease data (15%). In particular, these diseases are the most difficult to diagnose because they present vast symptoms and therefore many characteristics that need to be managed during the analysis and classification of patients affected by them and all these aspects justify the use of complex techniques such as neural networks.

To understand which architectures are the most implemented in the case of medical data collected by IoT intelligent devices, we analyze the distribution of the most used DL approaches in article selection. In Fig. 9 we present the distribution of the type of neural network architecture used in the selected articles. In particular, the card numbers are shown in blue, while the percentages are considered in orange; furthermore, at the bottom of the figure, there is a table of the values assumed by the coloured bars. The results show that the most used neural network architectures for IoT smart health problems are the convolutional neural network, the deep neural network, and the short-term memory network: in fact, these three different architectures alone cover more than 60% of the architectures used.

For completeness, to understand which neural network architectures are used in IoT smart health considering the pathology, we carried out a cross-analysis between the type of neural network and the pathology, the results of which are presented in Fig. 10. The analysis shows that CNNs are the most used when dealing with clinical data relating to neurodegenerative diseases, mental problems, diabetes, heart problems, and cancer. Furthermore, DNNs are more used for walking problems, lung diseases, and neurodegenerative diseases, while LSTMs are used when the data is related to heart problems, COVID-19, mental health, lung diseases, and walking problems.

RQ3: Which algorithms lead to better performances while exploiting smart health data collected from IoT device monitoring?

ML and DL algorithms play a vital role in improving the accuracy of the results produced and the reliability of healthcare data collected through IoT device monitoring. Algorithms can be trained using historical data to identify patterns and trends in healthcare data. Therefore, their application can be particularly useful for continuous monitoring of vital parameters such as blood pressure, heart rate, or blood sugar levels, or for long-term data analysis, allowing the identification of correlations between different parameters and risk factors. This can help develop personalized treatments and targeted preventive strategies, thus optimizing the effectiveness of treatments and improving patients’ quality of life. For all these reasons, the ML or DL methodologies used to analyze the data must provide truly accurate results.

Consequently, we study the distributions of accuracy metrics obtained with the different approaches proposed in the selection of the analyzed documents to find out which algorithms bring better and more accurate results in the classification of healthcare data collected by IoT devices.

In this regard, we conducted a quantitative analysis, the results of which are reported in Fig. 11, which shows the distribution of the values of the accuracy metrics reported in the selected articles using the different approaches used. In the 58 studies analyzed, the minimum accuracy value obtained is 68% using multinomial logistic regression, while the maximum accuracy obtained is 100% with Cubic SVM. On average, studies achieve an accuracy of 88% (confidence interval of the mean [79.80–96%]).

Furthermore, Fig. 12 reports the averages of the accuracies detected in the different studies that use the ML algorithm or the DL network (expressed in abscissa). The three best classifiers in terms of average accuracy are, in decreasing order: Cubic SVM and Ensambe (LSTM + RNN) (100%), J48 (99%), and custom classifier (97%).

The first classifier, Cubic SVM, is a variant of SVM that uses a cubic kernel function to map data into a high-dimensional feature space. Using a cubic kernel function allows SVM to find more complex and nonlinear decision frontiers in feature space. This can be useful when the data is not linearly separable in the original space and requires a nonlinear representation to be accurately classified, as is often the case in medical/healthcare data.

Ensemble (LSTM + RNN) is a combined neural network i.e. a combination of networks used to improve predictive performance. Both LSTM and RNN are types of neural networks used for sequential processing of data, such as time series or natural language data. Typically, to combine the two networks, one takes an approach of training each model separately on the same dataset and finally combines the predictions using techniques such as averaging or a more sophisticated method such as stacking or fusion.

J48 is a widely used algorithm for creating decision trees. This algorithm requires a labelled dataset, meaning that each example in the dataset is associated with a class label. To select the best attribute, the J48 algorithm splits the dataset at each node of the decision tree allowing for better information gain; the data set is split to optimize the Gini index or gain ratio. The tree construction process continues until a stopping criterion is met. J48 is widely used due to its simplicity, interpretability, and effectiveness in a variety of industries, including the smart healthcare data space. Finally, a custom classifier is a customization of an already existing algorithm, implemented concerning the classification objective. Customization, however, must be done thoughtfully and it is important to evaluate the performance of the customized model through appropriate validations and testing to ensure improvements and avoid unintended errors.

RQ4: What privacy and security issues have taken into account while using health data collected from IoT devices?

The use of health data collected by IoT devices raises significant privacy and security concerns that must be carefully addressed to ensure the protection of people’s sensitive information. For this reason, we also focused our analysis on this topic, evaluating in the selected documents whether privacy or security problems had been found and how these had been addressed during the phases of collection, transmission, and storage of health data by IoT devices.

First of all, we found that the two topics, privacy and security, were addressed in only 29% of the works analyzed. This means that less than one in three articles using health/clinical data questions whether the privacy of users using the devices is respected, considering whether sensitive data is recorded and whether storage techniques comply with data security standards.

In all the documents in which these issues were considered, the problems related to user privacy during the data recording phase by IoT devices were taken into account. Only 13.80% of the works also take into account the issue of data security addressed during the data archiving phase.

These studies also provide treatments aimed at solving these problems. In particular, issues relating to user privacy during the data recording phase(s) have been resolved as follows:

• considering only some characteristics extracted from the raw data stored and not using the starting data recorded by the device itself (Ware et al., 2022);

• when the video recordings of the users filming their faces and voices are also included in the original data, these recordings were eliminated from the analyzed data set (Wu et al., 2022; Cai et al., 2018);

• instead of recording user videos, a ‘Radio Frequency Identification’ (RFID) system was used, an automatic digital identification technology that allows the detection of objects, people, and animals, both static and moving, exploiting electromagnetic fields and not video recording (Chen et al., 2022);

• the sampling rate is set to 1 KHz to help address privacy issues since speech becomes unintelligible at this sampling rate (Adhikary et al., 2023).

As regards the studies that take into account the issue of data security also during data archiving during and after data analysis, the main solutions adopted for this problem were:

• the use of edge-fog architectures of the cloud (Chakraborty & Kishor, 2022; Sadhu et al., 2022),

• the utilization of encryption techniques to ensure that data collected from IoT devices is encrypted both during transmission and while stored in databases (Sarmah, 2020; Cheikhrouhou et al., 2021),

• the adoption of high-security protocols of the storage clouds (Yue et al., 2020),

• the conduction of regular security checks and assessments to identify vulnerabilities and risks in the cloud system, plus implement monitoring mechanisms to detect any unauthorized access or suspicious activity (Rahman et al., 2020).