Introduction
The rising number of chronic patients and the aging of the population render the avoidance of diseases an important requirement of healthcare. Prevention is not only defined by regular exercise, nutrition, and periodic preventive controls as a way to sustain a healthier environment but also as a method of keeping serious conditions from becoming worse. The future health sector must tackle an increasing number of chronic problems and the scarcity of treatments to satisfy patient demands [1]. COVID-19 has recently highlighted the importance of quick, comprehensive, and accurate eHealthcare and intelligent healthcare involving different types of medical and physiological data to diagnose the virus.
The use of emerging technology in protective policies and behavioral systems can help identify potential health conditions early and enable the scheduling of appropriate steps, such as concurrently monitoring treatments and preparing new assessments. The world’s smart health market is forecast to reach USD 143.6 billion in 2019, which will expand by an average growth rate of 16.2% between 2020 and 2027 [2].
Smart healthcare refers to platforms for health systems that leverage devices such as wearable appliances, the Internet of Things (IoT), and the mobile Internet to easily enter health documents and link people, resources, and organizations. Intelligent medical treatment includes diverse actors, including physicians, staff, hospitals, and research bodies. It comprises a dynamic framework with many facets, including disease prevention and identification, assessment and evaluation, management of healthcare, patient decision-making, and medical research. Elements of intelligent healthcare involve automated networks like the IoT, mobile Internet, cloud networking, Big Data, 5G, and artificial intelligence (AI), along with evolving biotechnology.
Sensors have been gradually embedded into diverse systems of our lives through computer technology, automation, and automated signal processing. Sensor-produced data can enable clinicians to more quickly and reliably recognize critical situations and help patients become more informed of their symptoms and future treatments. Intrusive and noninvasive tools—ranging from devices to read bodily temperature to dialysis control systems—provide personal and multimedia details and assistance to patients and the health care sector.
Medical signals come in the form of 1D and 2D signals such as electrocardiograms (ECGs), electroencephalograms (EEGs), electroglottographs (EGGs), electrooculograms (EOGs), electromyograms (EMGs), body temperature, blood pressure (BP), and heart rate. A health care monitoring system may use these medical signals to monitor a patient.
The IoT is slowly starting to connect both doctors and consumers through health care. Ultrasounds, BP readings, glucose receptors, EEGs, ECGs, and more continue to monitor patients’ wellness. Conditions like follow-up visits to doctors are critical. Several health care facilities have started to utilize smart beds, which can detect a patient’s movement and automatically adjust the bed to the correct angle and location. The Internet of Medical Things (IoMT) refers to the IoT used for medical purposes. When developing a fully integrated health environment, the IoMT can play an important role.
Sometimes, relying on only one type of medical signal may not fulfill the requirements for a complete diagnosis of a certain disease. In such cases, multimodal medical signals can be deployed for a better diagnosis. These signals can be fused at different levels, including the data level, the feature level, and the classification level [3]. When fusing signals, many challenges may be encountered. These challenges include synchronization when acquiring signals from different sensors, data buffering, feature normalization, and classification fusion [4].
In order to ensure patients’ and stakeholders’ satisfaction, intelligent health care has been revolutionized with the development of AI and machine learning (ML) algorithms in the context of deep learning (DL) and wireless local area network (wLAN) technologies [5]. The medical industry has been able to manage numerous medical signals from the same user—simultaneously improving disease detection and prediction precision—due to these technologies’ high computational performance, high data volume, accommodation of several terminal units, and the introduction of 5G and beyond 5G wireless technology.
In this paper, we present a detailed survey of IoT- and IoMT-based smart health care systems. The survey is limited to academic papers written between 2014 and 2020, located via the IEEE Xplore, ScienceDirect, SpringerLink, MDPI, Hindawi, the ACM Digital Library, and Google Scholar. The survey’s aim is to look at different related research areas such as the state-of-the-art IoT-based smart healthcare, data fusion of IoTs, AI in smart healthcare, cloud- and edge-based smart healthcare, and privacy and issues of IoT-based smart healthcare. At the end of this paper, we give few recommendations and make suggestions of future research directions.
The paper is organized as follows. Section II describes the methodology adopted to select the papers. Section III presents a comprehensive survey of the literature and answers several research questions. Section IV mentions some challenges and offers future research directions in this field. Finally, Section V concludes the paper.
Methods
We used the systematic review process PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) to identify studies and narrow down results for this review, as shown in Fig. 1. In the review process, there are three sequential steps, which are identification, scanning, and eligibility testing. In the identification step, papers are identified through Google Scholar search; after this step we identified 168 papers. In the scanning step, duplicate and non-conforming papers are removed; after this step 132 papers were selected. Then in the eligibility testing step, we removed the papers that were non-healthcare related. After this final step, we selected 110 papers to be included in the survey.
A. Research Areas
The research areas we used to select the articles were as follows: “state of the art regarding IoMT and medical signals for smart health care”; “the techniques of multimodal medical data fusion”; “cloud- and edge-based smart health care”; and “security and privacy of the IoMT”.
B. Search Strategy
Our survey of articles used a combination of keywords and involved formulating a search strategy and selecting data sources. We used the following combination of keywords: a) “Internet of Medical Things”; b) “Fusion medical signals”; c) “Multimodal medical data”; d) “Cloud/edge based smart health care”; and e) “Security and privacy Internet of Medical Things.” The number of papers elicited by each search strategy (item) after searching is shown in Fig. 2.
The search strategy was implemented based on the content of the main research areas. We restricted our selection to papers written between 2014 and 2020, as shown in Fig. 3. To locate appropriate papers, we scanned for related publications in major online research repositories, including IEEE Xplore, ScienceDirect, SpringerLink, MDPI, Hindawi, the ACM Digital Library, Google Scholar,. and other health and engineering journals.
C. Selection of Studies
Our initial search identified 168 papers. The “Internet of Medical Things” keyword got the largest number of papers. After removing duplicate and irrelevant articles, the search was reduced to 110 articles.
D. Data Extraction
The following data categories were collected from articles:
Application or tasks
IoT/IoMT
Features
Classifier
Dataset
Accuracy
Research Areas
The survey is divided into four areas: IoT or IoMT and medical signals; IoMT or medical signals fusion; edge- and cloud-based smart health care; and security and privacy in IoMT-based health care.
A. IoT or IoMT and Medical Signals
The research in [3] used a multi-sensor platform with two-channel pressure pulse wave (PPW) signals and one-channel ECG to estimate BP. From the collected signals, a total of 35 physiological and informative features were extracted. For dimension reduction and to obtain the most promising indicators for each subject, they presented a weakly supervised feature (WSF) selection method. Furthermore, a multi-instance regression algorithm was used to fuse features and enhance the blood pressure model.
Authors in [4] presented a technique for emotion recognition and classification across subjects. It integrated the significance test and sequential backward selection with a support vector machine (ST-SBSSVM) to enhance the precision of emotion recognition. The input modalities used included 32-channel EEG signals; four-channel EOG signals; four-channel EMG signals; and vital signals measuring respiration, plethysmography, galvanic skin response, and body temperature. Ten types of linear and non-linear EEG, EOG, and EMG features were extracted and fused with the vital signals to produce a high-dimensional feature vector. The features were fused and selected using significance tests and a backward selection search. The selected features were then fed into a support vector machine (SVM) classifier. The experiments were performed using two publicly available datasets, namely DEAP and SEED. The proposed method achieved 72% accuracy on the DEAP dataset and 89% accuracy on the SEED dataset.
One of the serious threats to the worker life is the disaster in mine area. Gu et al. [5] proposed a real-time monitoring system to ensure accuracy and reduce the risks to the mine worker. Authors discussed multi-sensor data fusion, situation awareness, and covering theories including the Internet of Things. A random forest (RF) SVM-based model was used to identify the level of the situation and to merge the data. The simulation analysis showed a root mean square error (RMSE) below 0.2 and a TSQ no greater than 1.691 after 200 iterations.
A data fusion enabled Ensemble approach was proposed in [6]. The collected data from body sensor network (BSNs) were fused to and inserted into an ensemble classifier for heart disease prediction. The ensembles were placed in a fog computing environment and the output from the individual predictors were fused. A prediction accuracy of 98% was shown in the result when the number of estimators was set to 40 at a tree depth of 15.
Steenkiste et al. [7] provided a reliable model for improving the performance and reliability of predicting sleep apnea based on sensor fusion method. In order to collect and integrate multi-sensor data, including oxygen saturation, heart rate, thoracic respiratory belt, and abdominal respiratory belt, the proposed approach used backward shortcut connections. To assess robustness and analyzed the performance of the proposed fusion method, both Convolutional neural network (CNN) as well as long short-term memory (LSTM) deep learning base-models were used.
A multi-sensor fusion (HBMF)-based hybrid BSN architecture has been developed by Lin et al. [8] to enable smart medical services. Medical services included data processing technologies, robot, and different sensors. To ensure that the robot make the right decision and to guarantee the quality of medical services, a multi-sensor fusion approach based on an interpretable neural network (MFIN) which used AI technologies has been proposed (see Fig. 5). Reliability and flexibility were improved compared with existing multi-sensor fusion approaches. In [9], seven channels from functional near-infrared spectroscopy (fNIRS) were fused with seven EEG electrodes to improve the detection of mental stress. Simultaneous measurements of fNIRS and EEG signals were carried out on 12 subjects. These measurements were conducted while subjects solved arithmetic problems under two different conditions (control and stress). The performance of the fusion of fNIRS and EEG signals was superior to the performance of each separately.
In [10], a fusion of EEG and ECG videos was proposed using three different transforms to improve video resolution: discrete cosine transform (DCT), discrete wavelet transform (DWT), and hybrid transforms. Both peak signal-to-noise ratio (PSNR) and mean squared error (MSE) parameters were used to measure the fusion effect. This empirical study found that hybrid transforms improved image reconstruction.
Authors in [11] suggested a method of medical image fusion using rolling guidance filtering (RGF). The study used an RGF to filter input images into either low-frequency or high-frequency components. First, the RGF separated the input images into low-frequency and high-frequency components, each of which had its own fusion role. A Laplacian Pyramid (LP)–based fusion rule and a sum-modified-laplacian (SML) based method were used to fuse the structural components and the detailed component respectively. The last step was image reconstruction. The proposed method achieved the best high-frequency information compared with other existing approaches.
A potential field segmentation (PFS) algorithm was presented by Cabria and Gondra [12]. PFS was used to segment brain tumors in magnetic resonance imaging (MRI) scans and the results produced by PFS were fused by ensemble approaches to achieve a fused segmentation. The proposed method was based on the physics notion of potential field and viewed the intensity of a pixel in an MRI scan as a “mass” which produces a potential field. The performance was validated on a publicly available MRI benchmark database called Brain Tumor Image Segmentation (BRATS) and showed that both PFS and FOR were similar methods. However, PFS was an exclusive segmentation algorithm and required fewer parameters.
An approach using particle filtering was suggested by Nathan and Jafari [13] to improve heart rate tracking with existing artifacts and the use of wearable sensors. They estimated heart rate apart from other signal features and to exploit the known steady, they designed observation mechanisms. This has contributed to the fusion of information from various sensors and signal modalities to increase the accuracy of monitoring. The performance of the proposed approach was examined on actual motion objects caused by ECG and PPG data with corresponding accelerometer observations, and results showed encouraging average error levels of less than 2 beats per minute.
A method based on multi-level information fusion was proposed by the authors in [14] to develop a predictive model to calculate BP from ECG sensor data. In this method, the data were fused in five levels (see Fig. 6). Data from multiple ECG sensors were fused and they used different techniques to extract the features from the input data in level one and two respectively. The fusion of output information from seven different classifiers was input into the meta-classifier in level 3. Knowledge from multi-target regression models for each BP type was integrated into level 4, and a single predictor for systolic BP (SBP), diastolic BP (DBP), and mean arterial pressure (MAP) was obtained in level 5.
In [15], the author presented a method based on physiological signals fusion to improve the accuracy of emotion recognition. Its performance was validated by comparing both fused and non-fused physiological signals on two publicly available datasets. A feedforward neural network classifier was trained using both fused and unfused signals. The result of the proposed method showed an improvement in performance on the DEAP and BP4D+ datasets compared with other current methods.
Chen et al. [16] modified an existing real-time system to produce a recognition system for human action. The device obtained data from various sensor types, such as depth cameras and wearable inertial sensors. Low-computation effective depth perception features and inertial signal features were inserted into two computationally powerful shared collaborative representation classifiers (CRCs). The proposed method was tested on a publicly available dataset called UTD-MHAD, and the results showed an improvement in overall classification rate (>97%) compared to using each sensor separately.
A data fusion cluster-tree construction algorithm based on event-driven (DFCTA) was presented in [17]. They designed a data fusion system for intelligent health monitoring in the medical IOT. By calculating the nodes’ fusion waiting time, the minimum fusion delay path was provided, and the fusion delay problem within the network was analyzed. The empirical study showed an improvement in reliability and timely in the proposed method compared with traditional method.
In [18], two procedures built on intrinsic image decomposition (IID) was proposed to address the complexity of complexity in extracting structural and functional information from both MRIs and positron emission tomography (PET) images utilizing the same decomposition scheme. The presented IID was used to decompose both MRIs and PET images into two components in the spatial domain. two algorithms were used, algorithm 1 for extracting the structural information and eliminating the noise from MRI images, while algorithm 2 was used for averaging the color information from the PET image. Based on IID models, three fusion methods were employed. IID+PCA, IID+IIC, and IID+HIS were superior to other existing methods when the planned method was tested.
Guanqiu [19] proposed a framework for medical image fusion that combined two methods: dictionary learning and clustering based on entropy. A Gaussian filter was used to decompose source images into high-frequency and low-frequency components. High-frequency and low-frequency were fused by using dictionary learning and L2-norm based weighted average algorithms respectively. The comparative experiments showed that the proposed method enhanced performance compared with other existing methods.
Baloch et al. [20] presented a layered context-aware data combination tactic for IoT health care applications. It included three phases: situation building, filtering and context acquisition, and intelligent inference. Reliable, accurate, and timely data were gathered from various sources. The aim of the analysis was to resolve issues such as uncertainty, irregularity, restricted range, and sensor deficiency. The drawback of this analysis was that no particular method was used to evaluate the suggested solution.
In [21], a distributed hierarchical data fusion architecture at various levels was employed using complex event processing (CEP) technology to improve decision accuracy and timely. It divided the task of data fusion into three-level processing models (low, middle, and high levels of data fusion). A smart health care scenario was prepared with appropriate IoT network topologies to prove the effectiveness of the proposed architecture. This empirical research found that the proposed solution allowed for effective decision-making at various stages of data fusion and showed an overall increase in the efficiency and response time of primary health services.
Survey Papers on IoMT and Medical Signals: Herrera et al. [22] presented state-of-the-art regarding sensor fusion for hand rehabilitation applications. Authors classified the research on hand rehabilitation into three categories: exoskeletons, hand movements, and serious games for hand rehabilitation. Of the types of sensors used, sensors based on EMG signals were the most common.
Wearable devices play a vital role in long-term health monitoring systems and are currently at the heart of IoMT [23]. In [23], a comprehensive study was presented with the goal of presenting the most important wearable health care monitoring devices, including biophysiological signs, motion trackers, EEG measurement devices, ECGs, BSCs, and so on. Based on expert, authors suggested that the most critical elements in health monitoring are motion trackers, vital signs, and gas detection
In [24], the authors argued that it was complicated to detect and resolve obstructive sleep apnea (OSA), although it is one of the most common diseases. The paper highlighted IoT systems that had supportive technologies and were utilized to diagnose OSA, including FC, smart devices, ML, the cloud, and Big Data. It further considered the improvement in the monitoring of sleep quality and other remote monitoring in AI-based health systems. In addition to the survey, a novel IoMT optimization paradigm was proposed to improve the quality of remote OSA diagnosis. The model showed an enhancement in the sensitivity, accuracy, energy consumption, and specificity of the system of remote OSA diagnosis.
A thorough and systematic analysis of current multi-sensor fusion technologies for BSNs was presented by Gravina et al. [25]. In the context of physical activity, they have presented an in-depth analysis and assessment of data fusion. Furthermore, they presented a systematic categorization by pinpointed specific properties and parameters that affected data fusion design choices at each level of the traditional classification (data-level, feature-level, and decision-level).
A comprehensive overview of different modalities fusing, such as MRI- PET imaging, computed tomography (CT)-MRI, X-ray, and ultrasound, was given by Sumithra and Malathi [26]. The research pinpointed different types of multimodal fusion and found that the exact boundary of the tumor in the brain could be identified by merging both CT frames and MRI slices.
Authors in [27] presented a thorough overview of the application of image fusion technology in tumor treatments and diagnosis, in particular liver tumors. It highlighted the key values of image fusion techniques by considering their limitations and prospects. It further presented an extensive review of the procedures and algorithms used in medical image fusion and concluded with a discussion of the research challenges and trends in medical image fusion. Table 1 presents a summary of the papers described above on the IoT or IoMT and medical signals.
B. IoMT and Medical Signals Fusion
Swayamsiddha and Mohanty [28] discussed different applications of the cognitive IoMT (CIoMT) to tackle the COVID-19 pandemic. Their review showed that the CIoMT was a successful tool for fast detection, decreasing the workload of the health industry, dynamic monitoring, and time tracking.
Yang et al. [29] proposed a combination of point-of-care diagnostics and the IoMT to assist patients in receiving proper health care at home. The proposed platform might reduce national health costs and monitor disease spread.
Singh et al. [30] highlighted the overall applications of the IoT philosophy in tackling the COVID-19 health crisis. This study aimed to decrease costs and improve treatment outcomes by employing an interconnected network for efficient flow and exchange of data. Singh et al. [31] also presented an IoMT concept based on ML approaches to tackle the COVID-19 health crisis. It provided treatments and solutions to issues related to orthopedic patients.
Kaleem et al. [32] discussed ways to actively apply the IoT in the medical and smart health care sectors and provided a method named k-Healthcare in IoT. The proposed method used smartphone sensors to collect and transmit data to the cloud for processing and then to stakeholders.
In [33], an event-driven data fusion tree routing algorithm was presented. The paper discussed the theory of health information and the sports information gathering system, which is divided into terminal nodes and client management systems. The proposed algorithm designed communication mechanisms according to the characteristics of IoT communication and used visual methods for modeling. The outcomes showed an enhancement in accuracy and timeliness compared with other methods.
Chiuchisan et al. [34] provided the design for a health care network to track at-risk patients in smart intensive care units (ICUs) based on the IoT model. It used a series of sensors and the Xbox Kinect to track patient motions and any required adjustments in environmental parameters to notify physicians in real time.
Sharipudin and Ismail [35] proposed a health care monitoring system to manage and process data in the patient monitoring system. The proposed system was combined with health care sensors that measured health parameters. The extracted parameters were then sent to cloud storage for medical staff’s reference.
Dimitrov [36] presented a discussion of IoMT applications and Big Data in the health care field which permitted innovative commercial models and allowed for variations in work progression, customer experiences, and output enhancements. Wearable sensors and mobile applications were used to fulfill numerous health needs and to collect Big Data from patients to advance health education.
Authors in [37] established early warning score systems based on the characteristics of vital signs. The proposed system supported the estimation of a health state by providing a helpful decision and cause for critical care interference. It investigated the most appropriate ML technique to predict the risk associated with input medical signals.
Sanyal et al. [38] proposed a federated filtering framework (FFF) based on the forecast of data at the central fog server using aggregated model from IoMT devices. This framework used models provided by local IoMT devices and then shared with the fog server. It presented a solution for many common issues, such as energy efficiency, privacy, and latency for resource-constrained IoMT devices.
Luna-delRisco et al. [39] addressed recognition, obstacles to implementation, and threats to the usage of wearable technology in the Latin American health care system. Major problems that the authors noted included the training and allocation of human capital in health care, the connectivity of public care, funding arrangements for health programs, and inequality in health. They considered smart wearable sensors in health care to be part of the solution.
Adali et al. [40] used a system where joint independent component analysis (ICA) and transposed independent vector analysis (IVA) were employed to fuse functional MRI, structural MRI, and EEG data. Results were obtained from healthy controls and schizophrenia patients using an audible oddball (AOD) function. The presented system was validated on a private dataset which included 36 subjects. The analysis was performed using the Infomax and entropy bound minimization (EBM) algorithms. The experiment revealed that the joint ICA model could be superior to the transposed IVA model. In the case of joint ICA, a robust ICA algorithm such as EBM was superior to the Infomax algorithm.
Authors in [41] presented a deep CNN model for seizure detection utilizing an excellent cross-patient seizure classifier. The visualization method demonstrates the spatial distribution of the characteristics learned by the CNN in various frequency bands when studying the seizure and non-seizure classes.
Bernel et al. [45] presented a DL method for the fusion of multimodal data to assist and monitor a user in performing multi-step tasks. Furthermore, they extracted deep features from individual data sources by a deep temporal fusion scheme. The Insulin Self-Injection (ISI) dataset consists of motion data captured with a wrist sensor and video data obtained from the wearable cameras of eight subjects. When the performance of the fusion method was evaluated, the proposed method was superior to other state-of-the-art fusion approaches.
Torres et al. [48] proposed a formulation that merged two features from three different modalities to categorize human sleep poses in an ICU atmosphere. Unlike other methods that extract one feature by merging data from various sensors, this method extracted features independently and then utilized them to estimate labels. Various properties and scenes were obtained from different modalities, cameras, and RGB (red, green, and blue) and depth sensors. Both shape and appearance features were extracted and used to train single modal classifiers and generate an estimation of the trust level of each modality.
Using the quantum-behaved particle swarm optimization (QPSO) algorithm, Xu et al. [46] presented an updated pulse-coupled neural network (PCNN) model to solve the problem of PCNN parameters and to improve the efficacy and correctness of medical image fusion. Different metrics, including mutual knowledge, standard deviation (SD), spatial frequency (SF), and structural similarity (SSIM), have been used to determine the efficiency of various methods. The result showed that the proposed algorithm has high estimation Accuracy. The proposed method was validated on five pairs of multimodal medical images from a publicly available dataset [42] and showed an improvement in performance over other current methods.
In [47], an approach based on weighted principal component analysis (PCA) for multimodal medical fusion in the contourlet domain was presented. One of the contourlet transform’s limitations was capturing limited directional information. In this study, the contourlet transform was combined with PCA to overcome this limitation and improve the fusion of medical images. It used max and min fusion rules to merge the decomposed coefficients, and the results showed improvement.
Using a hybrid technique combining non-subsampled contourlet transform (NSCT) and stationary wavelet transform (SWT), Ramlal et al. [49] produced an enhanced multimodal medical image fusion scheme. NSCT was used to decompose the source image into various sub-bands, and SWT was used to decompose the NSCT approximation coefficients into sub-bands. The efficiency of the proposed procedure was assessed through four sets of experiments. The suggested system was compared to other existing fusion schemes and showed improvement in brightness, clarity, and edge information in the merged image.
An improved algorithm based on a fuzzy transform (FTR) for multimodal medical image fusion was presented by Manchandaa and Sharmab in [50]. They considered the error images obtained using FTR pair to improve the performance of multimodal medical image fusion algorithm. To validate the proposed algorithm, different datasets were used, and the result was compared with other multimodal medical image fusion algorithms. The proposed algorithm showed a significant improved in edge strength, standard deviation, and feature mutual information.
Survey Papers on IoMT and Medical Signals: Joyia et al. [51] presented the contributions of IoT in the medical field and their major challenges in the IoMT. Numerous applications and research in IoMT were discussed in terms of how they solved issues faced by the global health care industry.
Irfan and Ahmad [52] reviewed current architectural models and produced a new one for the IoMT. They pinpointed the motivations that would lead medical practitioners to decide to adopt the IoMT and further demonstrated privacy and security problems in the IoMT.
Authors in [53] presented a comprehensive review of the current architecture for IoMT devices and discussed different aspects of the IoMT, including communication modules and major sensing technologies. The paper further discussed the challenges and opportunities related to using the IoMT in the health care industry. Communication gateways, data acquisition, and cloud servers were the main components of the IoMT framework.
In [54], the author presented a comprehensive overview of multimodal fusion of brain imaging data. This survey addressed the merits of multimodal data fusion in depth and summarized different methods of multivariate voxel-wise data fusion. A number of multimodal medical data fusion studies, particularly related to psychosis, have been reviewed. The author summarized this analysis by highlighting the importance of multimodal convergence in minimizing misdirection and perhaps discovering links between the brain and mental illness.
Table 2 presents a summary of the papers described above regarding IoMT and medical signals fusion.
C. Edge-Intelligent and Cloud-Based Smart Health Care
An edge- or cloud-based privacy-preserved automatic emotion recognition system utilizing a CNN was proposed in [55]. In [56], the authors suggested an appropriate training system for a deep neural network named ETS-DNN in an edge-computing environment. In order to change DNN parameters, ETS-DNN was combined with a hybrid algorithm for hybrid modified water wave optimization (HMWWO) In order to minimize data traffic and latency, data preprocessing and classification was carried out at the edge of computation. The results showed that ETS-DNN was superior to the compared approaches.
Han et al., in [57] provided effective communication by developing a clustering model for medical applications (CMMA)) for cluster head selection. The proposed CCMA aimed to enhance lifetime of communication, improve reliability, and offering energy efficiency in medical application. When choosing a cluster head, some criteria should be taken into consideration such as remaining energy, distance from the base station, capacity, delay, and queue of the IoMT devices. An improvement in terms of energy-efficient communication was shown in the proposed method compared with other existing methods.
Authors in [58] presented a cognitive IoT (CIoT) cloud-based smart health care framework with an EEG seizure detection method using DL. Authors in [59] proposed a voice pathology monitoring system integrating IoT and cloud technology.
In [60], Olokodana et al. used the ordinary kriging method to present a real-time seizure detection model in an edge computing paradigm. Fractal dimension features were extracted from EEG signals, and an ordinary kriging model was then used for classification. Computational time complexity is one of the limitations of kriging. In the proposed model, a previously trained ordinary kriging model was moved to an edge device for real-time seizure detection. The empirical study achieved a training accuracy of 99.4% and a mean seizure detection latency of 0.85 seconds.
In [61], an energy-efficient smart-health system based on fuzzy classification was proposed for seizure detection. The raw EEG data was processed at the edge before being transmitted to the mobile–health cloud (MHC). The proposed system minimized energy consumption by reducing the amount of transmitted data and provided high classification accuracy. The result showed an extension in battery life of 60% and a classification accuracy above 98%.
A new network paradigm, CIoT, has been proposed based on the application of cognitive computing technologies [62]. In [63], Chen et al. combined the advantages of edge computing and cognitive computing to create an edge-cognitive-computing–based (ECC-based) smart health care system which allocated maximum edge computing resources to higher-risk patients. The empirical experiments showed that the proposed system was capable of improving energy efficiency and user quality of experience (QoE).
Authors in [64] presented an edge-IoMT computing architecture which minimized latency and improved bandwidth efficiency. It consisted of two components: edge computing unit modules which compressed and filtered real-time video data, and cloud infrastructure modules which securely transmitted medical information to the physician.
Akmandor et al. [65] discussed different edge-side computing options which were designed to address challenges in smart health care systems. They demonstrated an edge-side reference model comprised of three levels: sensor node, communication, and base station. The compatibility between sensors and edge-side requirements enabled smart edge-side decision-making.
DL was utilized on a mobile health care platform to investigate a speech pathology detection method in [66] and an EEG-based remote pathology detection system in [67].
In [68], an automated voice disorder recognition system was used to monitor people of all age groups and professional backgrounds. By identifying the source signal from the speech using linear prediction analysis, the proposed system could determine the voice disorder.
In [67]–[69], the authors developed a voice disorder detection and monitoring system. In [69], they collected voice samples sent to the edge, which offers low latency and reduces delays in data traffic flow. After processing data using edge computing, data were transferred to the cloud for more processing and assessment. The medical information was then sent to a specialist, who prescribed suitable treatments for patients. The authors tested voice disorder classification and detection and compared the results with two related systems. The study found that the proposed technique improved performance in terms of detection and classification with 98.5% accuracy.
Oueida et al. [70] provided a resource preservation net (RPN) framework which integrated a custom cloud, edge computing, and Petri net. The framework improved reliability and efficiency and reduced both resources and time consumption. The proposed system was suitable for emergency departments and other types of queuing systems.
In [71], Kharel et al. used Long Range (LoRa) wireless communication and FC to produce an architecture for smart remote health monitoring. LoRa radio provides long-range communication and energy consumption for IoT devices and is used in the proposed system to link the edge user’s device with health centers. FC preserves network bandwidth and reduces latency by minimizing data exchange with the cloud. Tests showed that LoRa and FC had promising performance in remote health care monitoring.
In [72], the author utilized several wearable sensors and a DL method (namely a recurrent neural network [RNN]) to introduce a human activity prediction system. Data, features, and activity prediction were processed on fast edge devices like personal computers. To predict human activities from a public dataset, the RNN was trained based on the features, achieving 99.69% mean prediction performance.
Authors in [73] produced a task scheduling approach called HealthEdge that assigned priority to each task based on its emergency level in order to decide whether to process the given task remotely (i.e., in the cloud) or locally. They also provided a priority-based task queuing method which allowed emergency tasks to be processed earlier. The results showed that increasing the local edge workstation reduced processing time.
In [74], Vasconcelos et al. proposed a new method called adaptive brain tissue density analysis (adaptive ABTD) to improve the detection and classification of strokes. Edge computing devices provided low computation and cost and reduced time consumption in detection and diagnosis. The integration of the adaptive ABTD with edge devices and the IoT introduced speedy and efficient stroke diagnosis.
Authors in [75] presented a model for cloud-IoT–based health service applications in an integrated Industry 4.0 environment by enhancing the selection of virtual machines (VMs). They implemented their cloud-IoT model using three optimizers: particle swarm (PSO), genetic algorithm (GA), and parallel particle swarm (PPSO). The proposed architecture consists of stakeholders who use IoT devices to send tasks through cloud computing in order to receive services such as telemedicine and disease diagnosis. The cloud broker works in the middle to send and receive tasks over the cloud.
Authors in [76] proposed a tree-based deep model for efficient load distribution to edge devices. The input image was divided into volume groups and a tree structure passed through each volume. The tree structure had several branches and levels, each of which was defined by a convolutional layer.
In [77], Chung and Yoo increased the effectiveness of analyzing Big Data by proposing an edge-based health model using peer-to-peer DNNs. An edge-based health model and a server model were established separately to tackle the issue of response time delay. The results showed that combining DNN techniques and parallel processing models minimized response time delay.
Limaye and Adegbija [78] provided a comprehensive review of medical applications and algorithms in IoMT architectures and their integration with edge computing. IoMT workloads were compared using MiBench, an existing open source embedded system benchmark suite. The comparison showed that the IoMT applications differed from MiBench, indicating the need for a new benchmark sufficient for the IoMT microarchitecture. A cloud-based healthcare framework was proposed in [111]. In the framework, several aspects of data transmission and latency were discussed. An edge-enabled DNN-based method was proposed in [110].
Table 3 presents a summary of the papers described above on edge- and cloud-based smart health care.
D. Security and Privacy in IoMT-Based Health Care
The security and privacy of medical data are very important in smart health care frameworks. A patient’s data should be handled privately. If privacy is breached, the patient may be harassed in public, which can lead patients to become traumatized and depressed. If medical sports data are leaked, rival sports team members might use these data to solicit illegal advantages. Therefore, medical data should be dealt with privately and securely transmitted over communication channels [123]. This important issue has been addressed in a great deal of prior research.
Alsubaei et al. [79] presented a taxonomy of security and privacy in the IoMT. They categorized IoT layers (perception, network, middleware, application, business); intruder types (individual, organized group, state-sponsored group); impact (life risk, brand value loss, data disclosure); and attack method (social engineering, implementation layer, software or hardware bugs, malware). The perception layer includes wearable devices such as fitness trackers, BP sensors, and respiratory sensors; implantable devices such as capsule cameras; ambient devices such as door sensors and daylight sensors; and stationary devices such as CT scanners and X-rays. While there are many ways to fuse data from these devices, the authors did not discuss them in the paper.
In [80], the authors identified the potential security threats that can affect IoMT-based health care systems and recommended a series of security measures to tackle these threats. Some of the security issues mentioned in this paper include overlooking the aspects of built-in security, stakeholders’ unfamiliarity with security solutions and focus on marketing and financial gain, and a lack of consensus between stakeholders for overlapping solutions. Based on these threats, the authors proposed some ontology-inspired, stakeholder-centric, and scenario-based recommendations in line with available guidelines.
Ivanov et al. [81] introduced OpenICE-lite, a middleware for medical device interoperability designed to provide security for IoMT devices. Several applications were investigated for this middleware, including a critical pulmonary shunt predictor and a remote pulmonary monitoring system.
Lu and Cheng [82] proposed a secure data-sharing scheme for IoMT devices. First, the system guarantees the protection of and permitted access to mutual information. Second, the system conducts effective integrity tests until the customer opens mutual data to prevent an erroneous application or calculation performance. Ultimately, the system provides a lightweight procedure for both consumer and customer. The scheme removes the burden of generating encryption and decryption keys solely on end devices.
Mohan [83] presented some cyber threats to IoMT devices and provided some solutions to these threats. As IoMT devices are limited by their battery life, they have only limited encryption capability and are thus at risk in terms of integrity, confidentiality, and privacy. Sensitive patient data can be leaked, and denial of service attacks can be made by draining the battery. As solutions, IoMT devices must be installed during deployment and software details transferred to the cloud-based system provider. IoMT devices encrypt all patient data using lightweight cryptographic methods and store patient data on the cloud-based system. Only approved entities who send their verifiable attribute-based certificate to the cloud provider may access this data.
Nkomo and Brown [84] proposed a cybersecurity framework for IoMT devices in smart health care systems that had five attributes: identify, protect, detect, respond, and recover. First, asset management and risk assessment should be identified. Second, access control, data security, and protective technology should be developed. Third, anomalies and events should be detected. Fourth, response planning should be designed through analysis and mitigation. Fifth, a recovery path should be planned.
Rathnayake et al. [85] realized a security mechanism for a smart healthcare system using the IoMT. First, data from different IoMT devices were encrypted using asymmetric cipher and advanced encryption standard (AES) keys. The keys were protected using a ciphertext attribute-based encryption (ABE) protocol. The encrypted data were transmitted through an insecure network. At the receiver end, AES keys were decrypted using the ABE protocol. Data were then decrypted using the ASE keys. This mechanism maintained the privacy and the security of patients’ data.
Seliem and Elgazzar [86] proposed a blockchain for IoMT (BIoMT) to preserve security and privacy in a smart health care framework. The BIoMT had four layers. The first layer was a device layer, which contained IoMT and user interface devices. The second layer was a facility layer, which had a bolster to look after IoMT devices. The facility layer provided the basic blockchain modules for attribute number selection, security generation, and identity issuance. The third layer was a cloud layer that provided the computational power and storage, and the fourth layer was a cluster layer which contained medical facilities and the service provider.
Wang et al. [91] designed a fog-based access control (AC) method for the IoMT. The authors developed a method that installed an extra layer of control on fog servers to improve protection for local mobile devices. A register in the AC server was important for compliance with devices. Data access requests were submitted to the AC server, where the status of the application could be reviewed. The registry needed to ensure that the incoming function had been recorded in the past. The comparison should be performed as the work form was recorded to ensure that the privacy setting was changed. The architecture was situated in the fog layer, where functional-oriented servers could provide the required AC service to each device.
Dilawar et al. [92] introduced cryptography as a solution for the safe exchange of patient safety records using blockchain technologies to protect medical data. A unified blockchain-based technique would solve many of the difficulties related to a centralized cloud solution. Authors in [93] introduced an access management model that safeguarded patients’ medical data from internal information security attacks. It enabled only legally permitted people to connect despite physical limitations. The suggested model incorporated authorization consistent with permits and responsibilities, rather than positions for medical personnel only. It eliminated the contradictions of current AC models.
Omotosho et al. [94] identified and incorporated some of the main characteristics of a patient’s health report that should be published and made accessible at all times as well as qualities that should be disclosed only during emergency conditions or pre-hospital treatment. Creating medical features from patient health information that may be retrieved in critical cases is a proactive step that allows technicians to obtain access to required details in pre-hospital services while protecting patients’ dignity and confidentiality.
Farahat et al. [95] introduced a data encryption scheme that involved first encoding data, then encrypting those data with a rotated key until they were sent across the network. Doctors can recover the protected data using their login keys and credentials. The scheme was implemented using low-cost equipment and reliable applications to ensure safety in the delivery of medical information.
Guan et al. [96] proposed a differential private data clustering scheme to allow privacy-preserving IoMT using the MapReduce system. For large-scale data sets, MapReduce is a parallel programming system that abstracts parallel computing procedures into two functions: Map and Reduce. In this scheme, the authors refined the distribution of privacy budgets and the collection of initial centroids to boost the performance of the
Hamidi [97] proposed a modern paradigm for the application of biometric technologies to the advancement of smart health care using the IoMT, which, in addition to being simple to use, requires broad-scope data access. While card IDs and passwords control entry, these systems can be quickly broken and are known to often be inefficient. A biometric trait has four main features: universality, distinctiveness, permanence, and collectability. The author anticipated four levels of security strategies: IoMT device, communication, analytical, and management.
Alsubaei et al. [80] outlined a web-based IoMT security assessment framework focused on an ontological scenario-driven methodology to propose security steps in the IoMT and to evaluate safety and deterrents in IoMT solutions. The framework encouraged the development of a strategy that fits stakeholders’ protection goals and facilitates decision-making.
Elhoseny et al. [98] proposed a hybrid optimization of asymmetric encryption for IoMT security. An ideal private and transparent key-based authentication was used in IoT therapeutic images. Various approaches were considered to achieve optimal hybrid optimization, from which the researchers differentiated and analyzed the critical open-ended difficulties in enhancing IoT in healthcare.
Shakeel et al. [99] introduced learning-based Deep Q-Networks to reduce ransomware attacks when handling health records using IoMT devices. The approach analyzed the medical knowledge in various layers per the Q-learning principle, which allowed transitional attacks to be eliminated with less difficulty. Efficiency was measured in terms of energy, lifetime, throughput, accuracy, and malware error detection rate. Yi and Nie [100] proposed a multivariate quadratic equation–based cryptographic security system for IoMT devices. A physical analysis model of the cryptographic system was designed by analyzing fault tolerance and differential power on a cloud platform.
Survey Papers on Security and Privacy in IoMT-Based Health Care: A survey on security and privacy in the IoMT was presented in [101]. The authors identified four requirements for security and privacy: data integrity, data usability, data auditing, and patient information privacy. Existing solutions to these requirements were discussed and included data encryption, access control, trusted third-party auditing, data search, and data anonymization. For example, some encryption methods for access control include attribute-based encryption and symmetric and asymmetric key encryption. The paper ended by noting some future challenges, such as how to deal with insecure networks, develop lightweight protocols for devices, and share patients’ private data.
Hatzivasilis et al. [102] reviewed security and privacy in the IoMT. In an IoMT-based health care system, there are three main application settings: hospitals, homes, and body sensors. Three security aspects—confidentiality, integrity, and availability—should be enforced in device, connectivity, and cloud security. The survey analyzed different types of security components. Various types of protection mechanisms, identification and anonymity techniques, and data destruction for device reuse were also discussed.
Sun et al. [103] provided an outline of the latest problems, requirements, and possible risks to the protection and confidentiality of IoMT-based health care systems. To design an IoMT networks, one must address postural body movements, rises in temperature, energy efficiency, transmission range, quality of service, and heterogeneous environments. The security and confidentiality requirements have different attribute levels. At the data level, care must be taken regarding confidentiality, integrity, and availability. At the sensor level, the design must address tamper-proof hardware, localization, self-healing, over-the-air programming, and forward and backward compatibility. At the personal server level, device authentication and user authentication should be considered, while at the medical server level, important requirements include access control, key management, trust management, and resistance to denial of service attacks.
Li et al. [104] provided a survey of secured IoMT with friendly-jamming schemes. The authors reviewed the IoMT’s existing protection systems and defined key security issues in the IoMT. They recommended friendly-jamming schemes to protect patients’ sensitive diagnostic data obtained from medical sensors. They concluded that, when properly planned, friendly-jamming approaches could substantially reduce the probability of effectiveness of eavesdropping activity while having no substantial impact on legal transmission.
Ghoneim et al. [105] introduced a new medical image forgery detection method to verify that health care images had not been changed or altered. The method generates an image noise map, realizes a multi-resolution regression filter to the noise map, and feeds the output to SVM-based and ELM-based classifiers. Another copy-move image forgery detection method was proposed in [112]; the method could be used in medical image forgery detection.
Lin et al. [106] reviewed the security and privacy issues, challenges, and future directions in the IoMT field. There are four major categories of medical sensors: disposable health sensors, connected health sensors, IoT-supported sensors, and IoT market cap sensors. The authors provided a systematic review of these sensors in terms of their security and privacy, followed by the challenges they present. Some of these challenges included the integration of multiple sensors with proper protocols, data bursts, and social acceptance. In a related survey, Masud et al. [117] outlined some limitations and issues related to the security of IoMT devices and provided some recommendations. They listed risks such as the disclosure of personal information, data falsification, lack of training, and reasonable accuracy.
Challenges and Future Research Directions
The major challenges of IoT and AI-based smart healthcare include sensors’ interoperability, device communication, security and privacy, device management, information management barrier, and efficient use of AI. In some health care environments, the bulk of IoMT devices can be used to identify and diagnose an illness, and the data collected from heterogeneous sensors contains a variety of issues, such as hardware glitches, drained batteries, or connectivity problems [106]. There are certain basic problems that are normal and unregulated. In particular, there are sometimes unexplained errors in the usage of popular medical sensors, such as mobile phones and smartwatches. There are also regular complexities, such as battery power, distinctions between particular physical characteristics, and variations in the environment.
The above problems indicate that several difficulties exist in smart health care, though multimodal signals and several IoMT devices are being used. A simplified and easier fusion solution should be discussed to facilitate the general adoption of such smart health care [115], [119], [121], [123]. Below, we discuss some of these problems and potential solutions.
The healthcare system can get inconsistent data from the multiple sensors because of the unawareness from the researchers. Incomplete data may get thieved or faked by other people. Radio frequencies of IoTs might have an effect on reading areas, and readers might give false readings. Tag collisions and tag detuning should be corrected, along with metal/liquid effects and tag misalignment. The system can get redundant data which need to be refined.
Wearable sensors are equipped with batteries, Bluetooth, and other materials and were designed to be attached to human skin. For human safety, it is important to consider toxicity, flammable materials, and other factors when designing wearable sensors. Wearable sensors that constrain body movement, such as a belt worn at the waist or ankle, are uncomfortable, especially for the elderly and children. One challenge is to develop sensors that continuously monitor human vital signs using suitable materials and without reducing user comfort.
There is an increase of the number of connected sensors, devices, and IoTs in any smart system. A massive healthcare network will work only if it has sensing capabilities plus the capacity to produce important information. In the healthcare system, many millions of sensors and IoTs are linked that provide massive amounts of data to be studied. In the IoT, the entities should have compatible data model and knowledge representation model.
There is a need to recognize interoperability of IoTs or partnership between nations when it comes to the development of digital health infrastructure. This disadvantage, along with lack of IT infrastructure, is attributed to both a lack of IT skills and the need for international collaboration in the sharing of confidential medical data, which will promote remote telemedicine and the provision of high-quality medical care. Shibboleth is a distributed identification key, which allows individuals to be authenticated inside and through organizational systems. The conventional Shibboleth mechanism requires a user to confirm to an ID provider and then directs a demand for a site to be hosted by a service provider. With this distributed approach, Shibboleth allows digital health organizations to have a single sign-on capability, as in the case of digital health.
Automatic health care programs depend on self-sensing, self-adjustment, and self-tuning [108], [113]. As background such as sensor noise and recording environment, varies, fusion of sensors and IoTs can deal with the modifications, since they can have a direct impact on system properties such as precision. Information transfer methods for transfer learning should be used to permit the system to adjust to particular circumstances by collecting and transferring acquaintance from one situation to another.
Unauthorized access to IoT devices may contribute to extreme health and private information threats to patients. Linked computers, including the compilation, aggregation, retrieval and transmission of patient knowledge to the cloud. Cloning, spoofing, RF jamming and cloud polling is prone to system type. In the cloud survey, traffic is diverted such that commands can be injected directly to a computer by an individual in the center.
Attacks with denials of service (DoS) can impact health organizations and the security of patients. Although replication (use of several devices on the network) is a standard protection of DoS, it might not often be feasible to replicate resources in a healthcare setting since some of the devices are essential systems implant. Owing to the amount and sophistication of new device and hardware bugs, the quick identification of possible security hazards remains a problem. This problem is escalating as the Internet links more and more users. Standard security is also widespread today and unsecure user interface access raises the threat surface more.
Many wireless networking devices have also recently been used in the health care industry, including Wi-Fi, BLE and ZigBee, for linking various medical equipment and sensor forms to each other. Defense from eavesdropping, sybil assault, plunger hole attacks and sleep loss attacks must be applied with these wireless sensor and sensor technologies. In order to preserve protection and privacy, core data sets of personal details, family histories and electronic medical documents can also be guarding against hackers and malicious devices.
The misuse of access privileges by allowed insiders is a big concern. This kind of information sharing occurs when health facilities disclose sensitive medical information to unauthorized people, either due to irresponsibility, for individual or criminal purposes, or in return for illegal benefits. Celebrities’ health reports and the lawmakers’ information also leaks to the public from a centralized healthcare system. This could cause a breach of the regulation by the insiders and the documents that they would not have access to. For example, medical personnel who are not taking care of real patient and former staff who are not yet restricted from data query. A disgruntled party will cause problems to each other by accessing the protected details of each other. Intruders are trying to pretend to be healers in order to infiltrate. Cybercrime as a virus of today’s Internet sector is a big issue and a menace to health. There are high costs for unsafe medical practice such as negative impact on their reputation, penalties, legal liability, and many more.
Traditional AI-based healthcare systems may not gain acceptability to the doctors. Therefore, explainable AI-based system can be deployed, where the doctors can visualize the detection or classification of diseases. The optimization of edge resources can be efficiently done edge-intelligent algorithms [105], [107], [109], [114].
The practical usefulness IoMT activated healthcare systems is rarely addressed in literature. The main concern is that the most relevant data is owned by companies and is not accessible to the public. The efficient deployment and utilization of data fusion in practice will allow for more reliable measurement and evaluation of day-to-day physical activity utilizing low-cost monitors that can lead to easier and better preventive care for chronic diseases. We assume that hosting medical data in a public archive with appropriate protection precautions and exploring current data fusion strategies using such public data will be a crucial potential direction for future research.
The advancement of next-generation wireless networks poses a great prospect in smart healthcare [118], [120], [122]. With the help of 5G and beyond 5G networks, now the healthcare system can be reached anywhere in the world faster than before. In addition, federated DL and edge-based computing become easier and powerful [2], [104], [116].
Conclusion
Smart healthcare is a well-researched area. In the smart health care domain, there is a breadth of literature covering IoT, IoMT, medical signals, AI, edge and cloud computing at various rates and utilizing varied tactics. However, to the best of our knowledge, there was a lack of a thorough and systematic analysis of state-of-the-art IoT, IoMT, AI, medical signals use and fusion, edge and cloud computing, privacy and security in the smart health care domain. The purpose of this survey was thus to offer a formal classification and specific comparative context for IoT, IoMT, AI, edge and cloud computing, privacy and security in smart health care. The survey included the use of IoT, IoMT, and medical signals, the fusion of sensors, and the use of edge and cloud computing in smart healthcare. It further provided a survey of security and privacy issues involving IoMT devices. Finally, some research challenges and future research directions were discussed.