Random Caputo-Fabrizio fractional differential inclusions

1.   Introduction

A new era in healthcare is beginning, one in which extensive medical data will be more and more important. For instance, accuracy in healthcare looks at a range of patient information, in addition to variances within molecular features, Electronic Health Records (EHRs) and surroundings, as well as ways of living. This guarantees that the proper drug is provided to the appropriate patient at the correct point in duration [1], [2]. As a result of increased access to healthcare data, healthcare research has both opportunities and challenges. Discovering the connections between all the data in huge medical datasets is a significant challenge for developing a reliable healthcare system based on data-driven methods, deep learning (DL) and machine learning (ML). In the past, academics have tried to merge data from various sources to create collaborative knowledge sets that may be used for prediction and discovery [3]. DL and ML, technique-based prediction tools, have not been widely deployed in the healthcare industry despite the fact that existing models have significant promise [4], [5]. Biological data's high complexity, variety, dependency on time, sparsity, and unpredictable nature make it difficult to effectively use [6].

Our cities are becoming more digital as a result of vigorous devices equipped with sensors for data collecting, environmental inspection, digital carriage, health betterment, simple entrance to facilities and services, and general assistance requirements for one and all across the urban areas. The term “smart city” is frequently used to describe a digital metropolis. A smart city contains many elements, all connected by increased technology and the requirement to offer its residents high-quality services [7]–[9]. Smart health is one of these elements. The application of cutting-edge technology to improve and provide high-quality healthcare is known as smart health. Therefore, it necessitates the use of astute gadgets, electronic health surveillance equipment, and web services that are entirely associated with a data core in order to draw conclusions about a person's or a community's health. Since many people use smart health, its creation is a great development. This has increased demand for smart health apps in recent years, along with the continued advancement of technology instruments. For instance, with a mobile smartphone and the inbuilt sensors and applications available on these devices, a person may check their blood pressure [10]. Additionally, the effects of weather on health may be explored by the weather or climatic data of a specific location inside a city to learn what sort of regions of the city to escape [11].

It is standard procedure in biomedical research to have a specialist choose the phenotype that will be employed. Nevertheless, supervised feature space definition loses opportunities to find novel patterns and scales poorly. The graphical representations required for the forecasting can instead be automatically found from the raw data using learning-based methodologies. The nonlinear modules known as layers of representation are also used by DL and ML methods to transmute the properties at a specific level (starting over the initial contribution of the data) toward some properties at a more metaphorical level. The DL and ML present a fascinating as well as a contemporary paradigm which is potential for biomedical informatics considering their demonstrated effectiveness in a variety of fields (such as recognition of voices, machine apparition, spontaneous language processing chores, etc.) along with the prompt speed of methodological improvements [12]. The utilization of DL and ML aptitudes in healthcare is currently proposed or is already in use. For example, International Business Machines (IBM) Watson Health uses this technology to interact with patients and healthcare providers and focuses on providing affordable healthcare to a large population; Google DeepMind has put up a strategy to utilize its specialized expertise to enhance healthcare [13], for the purpose of identifying health concerns on computed tomography (CT), scans, Enlitic employs deep neural networks and X-rays [14], and these are just a few examples.

The extensive application of computer-assisted decision-making and outcome evaluation in the provision of healthcare highlights the technical value of modeling expertise and knowledge [15]. On the other hand, conventional rule-based models primarily rely on feature representation of issue domains and are unable to accurately mimic the complexity of human brains. So, we attempt to address this issue by using a deep model. The deep model creates a single model that can imitate human thought processes by fusing feature representation with learning. However, a vast number of medical conditions that could profit from DL and ML approaches haven't been fully investigated. The various advantages of DL and ML, such as improved performance, a lateral learning model with incorporated feature learning, and the aptitude to deal with intricate, diverse information, among others, could be advantageous to the healthcare industry. Additionally, the sparse, noisy, heterogeneous, and time-dependent nature of healthcare data presents a number of challenges for the DL and ML research community [16]. Additionally, better methods and tools are required to integrate DL into clinical decision-support functioning.

In view of this study, we address existing and planned DL and ML applications in medicine, drawing attention to the prominent factors that appreciably influence health care. We don't wish to furnish a comprehensive context on technological particulars (for example, [17], [18]) or common uses of DL and ML [19]. Instead, we emphasize on biological information obtained through clinical imaging, EHRs, and wearable technology. These industries have not yet made considerable use of DL and ML, despite the fact that other data sources, such as the metabolome, antibody one, and other omics data, are projected to be helpful for health monitoring. We consider these aspects in the context of the characteristics and powers of various DL and ML methods, as well as their practical uses in problem-solving [20]. Another key goal of this study is to identify important research opportunities and issues, such as those pertaining to new algorithm design, data-driven hyperparameter learning, model optimization, integrating domain knowledge, and adapting to resource-constrained devices, as well as effective data representation. One possible product of this work is “Future Generation ML-DL Modeling”. Therefore, the purpose of this work is to serve as a reference for scholars and industry professionals who wish to learn about and create data-driven intelligent systems and enhance intelligent healthcare through the use of DL and ML techniques. We have reviewed a significant number of publications with an emphasis on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) technique, which is a frequently used checklist and guideline for reporting systematic reviews and meta-analyses of studies that evaluate healthcare interventions or other healthcare-related concerns. The following checklist items—justification, eligibility requirements, search method, research selection, data extraction, and synthesis—were covered in this study that coordinated with this technique. The concrete contribution of this paper is summarized as follows:

Organization of this survey: This article is arranged as follows: Table 1 displays the abbreviations used in this paper. Section two describes the survey organizing technique PRISMA, which is maintained throughout the research. Section three provides a review of ML, DL and healthcare. After that, section four presents integration among them with a couple of tremendous taxonomies. Applications of ML, DL, and healthcare in diverse areas have been highlighted in Section five. Also, section six considers potential challenges as well as future opportunities in the area of our desire. At last, we conclude this study in Section seven. In summary, this study presents a logical road map in Figure 1.

2.   Searching methodology and implementation of PRISMA approach

We have implemented PRISMA in this survey paper. Although there are so many terms and techniques, some of them are considered here that are appropriate for this scenario. The following are the important terminology and principles for this approach:

We have filtered our papers using the above terminology and criteria. In addition, Figure 3 described the step-by-step process used in this study to adopt PRISMA technology. For the identification process, we have searched papers with predefined keywords along with some added specific terms. In the screening part, the duplication of some papers has been removed, and in the eligibility section, the eligible studies have been considered. After finding the required papers, the inclusion part has been completed.

3.   ML, DL, and smart healthcare: State of the art

The utilization of ML and DL together in the healthcare industry is relatively new and hasn't been fully investigated before. The medical healthcare industry is a promising subject according to current research trends, and we will discuss some of the most important recent literature in the following sections regarding the methods, contributions, and applications of ML-DL to the many fields of this industry. A discussion of the surveys (i.e., 49 papers) about DL and ML technologies, together with the integration of smart healthcare, is presented in Table 2. Each of the groups shown in this table has had its approach and contributions investigated.

3.1.   Key features of ML

Classification: Categorization is a common way that supervised learning issues are represented. The objective is to develop a technique that, given a limited number of already-known subgroups, determines which group a new result belongs to. The training set is fed instances with known subcategories. Each observation may receive more than one label during categorization as opposed to each observation just receiving one. However, the comprehensive taxonomy of the ML approach applied to healthcare is illustrated in Figure 5.

Clustering: Clustering, which seeks to organize data in a way that they are more similar to one another than to observations in other groups, is a standard unsupervised learning technique. In contrast to classification, the clustering is determined by the training data rather than being known beforehand. The three primary types of clustering algorithms are as follows– centroid-based based, hierarchical, and density-based [70]. The first two methods are sometimes merged into a single partitional, which is a class. Many of the solutions mentioned in this article use K-means, one of the more well-known algorithms [71].

Association Rules Mining (ARM): ARM is the method that is used to unveil latent relationships between different entities using important metrics. It is one of the most common techniques mostly used in data mining and ML fields Group regulations unsupervised learning also includes mining. Support and confidence metrics are the major metrics used to assess association rules.

Reinforcement Learning (RL): Another use of ML that draws on trial-and-error learning is RL. Real-world learning is based on how people learn by observing their environment. It helps agents who aim to get the most out of each task overall. RL considers what to do next in order to categorize data based on the advantages the input provides in the short- and long-term. RL and decision-making are frequently combined; for example, this configuration permits human-level performance in a variety of games.

3.1.1.   Process of ML

Here, we cover every phase of creating a framework for learning in the medical field. Consider that the purpose of this section is to guide scholars toward creating a system of learning for the medical field. Five crucial factors need to be considered while generating a learning model for the healthcare sector: Problem description, information set, preparation of information, ML model production, and assessment. These phases are depicted in Figure 4. Each of these stages is thoroughly explained in the sections that follow.

Problem Specification: We must initially consider the following questions while developing an approach to learning for the medical industry: “What is the purpose of designing this learning model?” Identifying healthcare problems and obstacles is the first step in developing a relevant model. The availability of data is an important feature of the first stage. This indicates that data sources should be known to investigators since the information needed to develop and evaluate the learning model should be readily available. A lack of digital data may result in a lack of data in the healthcare sector, patient confidentiality, economic problems, or rare diseases.

Database: Datasets are used in industrial healthcare applications to train, validate and test learning models. The healthcare database contains demographic data, pictures, test results, genetic data and sensor data [72]. These data are produced or collected via various platforms, including personal computers, cellphones, mobile applications, network servers, e-health records, genetic data, and wearable technology [73]. Today's global relationships could be enhanced due to the internet and cloud computing [74], [75]. Data accessibility has improved as a result.

Data Preparation: When developing a model for the medical sector, data preprocessing is one of the most difficult obstacles to overcome since a ML prototype needs data with greater attributes to deliver better training results and more accurate results. Analyzing disturbance data, absent values, replicate data, and contradicting data is known as data preprocessing. This approach aims to improve the information's integrity before creating the framework for learning. To estimate missing values or filter anomalies, data preprocessing might turn out to be necessary. Furthermore, certain information minimization approaches, such as feature assortment or feature parentage, are perhaps applied if the data has high dimensionality [76], [77].

ML Model Development: We must consider the margin of the database, the variety of learning stratagem, and the length of the model inference process while establishing a learning model for the healthcare industry [78]. To prevent overfitting or underfitting, we measure the degree of complexity of a model to learn about the size of the information stored in the database. It's crucial to take a learning model's training duration into account. More parameterized learning models, however, can result in more precise outcomes. Reward-based learning, unsupervised learning, supervised learning, and semi-supervised learning are the four prominent categories of learning techniques [79].

Evaluation: Executing numerous procedures to find discrepancies between the system's current behavior and the expected behavior is what evaluation of a ML-based system entails [80]. We must reassess the learning model's performance after being deployed in real situations to assess how it will behave while interacting with actual users [81]. The following components are supposed to be taken on board when assessing the performance as regards to the final model:

Subsequent that we discuss several relevant degrees for evaluating a learning prototype. This scale is on the basis of the following criteria: TP, TN, FP, and FN:

Sensitivity: In order for a classifier to accurately predict a positive outcome whenever the associated ground truth is similarly positive, this scale is denoted as a probability. The true positive rate (TPR), which is another name for this scale, is determined as pursue:

Specificity: The probability that a classifier will accurately predict a negative outcome when the associated ground truth is likewise a negative result is specified as this scale. The true negative rate (TNR), another name for the specificity, is computed as follows:

Positive Predicted Value (PPV): This rating system is characterized as the likelihood that the classification will correctly predict the outcome of the test when the test result (the classifier's output) is positive. PPV is also known as precision, and it is computed as follows:

Negative Predicted Value (NPV): The probability that a classifier would correctly predict the outcome when the test result is opposite is how this grading system is defined. The computation of this scale appears as follows:

Accuracy: This scale has a lot of significance. Usually, this scale is used to evaluate classifiers. It is described as the percentage of samples that the classifier is properly categorized. That is determined as follows:

Matthews Correlation Coefficient (MCC): This method is stated as the correlation between the actual results and those projected. Its value ranges from +1 to 1. If MCC=+1, then the classifier correctly anticipates the outcome. If MCC=0, then the classifier cannot prophesy the outcome ameliorate than an arbitrary process. If MCC=1, therefore, there is a complete inconsistency between the expected outcome and the associated basis truth. The MCC margin is determined as follows:

False Discovery Rate (FDR): The proportion of the total number of favorable outliers to the percentage of positive samples that were mistakenly anticipated is represented by this scale. This is how the FDR scale is created:

Area Under the Receiver Operating Characteristic curve (AU-ROC): This scale is yet another important variable to consider while assessing classifiers. It is calculated using the area under the ROC curve. It should be noted that the ROC was calculated using TPR and FPR. This scale is derived as follows:

F1-Score: This scale combines the sensitivity and accuracy measures. It is described as the average of their scores, weighted. Recall is a measure of overall amount, whereas precision is a measure of quality. An algorithm with a greater precision will return more relevant results than irrelevant ones, while one with a high recall will return the majority of relevant results, regardless of whether irrelevant results are also returned. We can clarify as follows:

When F1-Score is one, it has reached its ideal value. F1-Score = 0 is said to be the overpower value in contrast. This scale is derived as follows:

3.1.2.   Role of ML in smart healthcare

Identification and Diagnosis of Disease: The primary applications of ML in healthcare are for diagnosing and predicting ailments and conditions that are traditionally seen as being difficult to study. This might include anything from early-stage cancers that are hard to find to other inherited illnesses. The best example of how to combine fast analysis with psychological registration is IBM.

Engineering and Medical Innovation: In the new era of medicine production, there are primary clinical learning programs for robots. This also involves research and development (R&D) abilities like next-generation sequencing and precision medicine, which may be used to identify different treatment options for complex illnesses.

Analysis of Medical Imaging: Computer visualization is a technical advancement made possible by DL and artificial intelligence (AI). The Microsoft-created Inner Eye activity, which targets picture diagnostic tools for picture analysis, has recognized this. As ML gets more open and as they approach its informative limit, this AI-driven symptomatic interaction turns out to be a component of the objective of grasping new information based on shifting clinical symbols.

Personalized Medicine: The combination of prognostic analysis and personal fitness may not truly produce individualized care. Currently, doctors can only do a limited number of studies or assess threats. In the next years, more gadgets and bio-sensors with advanced health assessment capabilities will hit the market, making information easier to obtain for cutting-edge ML-based medical care originalities.

Integrated Health Records: The process of maintaining a modern health record is labor-intensive and complicated. The information section measure has been assisted by technology developments, although this is still true for the majority of cycles. In the medical sector, the primary role of ML is to streamline procedures in order to save time, effort, and money. Using vector technologies and ML-based optical character recognition (OCR) recognition algorithms, archives are being organized in an ever-increasing variety of ways.

3.2.   Key features of DL

Fully connected neural network: In a fully connected neural network (FCNN), every neuron in a layer may communicate with every other neuron in the layer above it. For the same reason, these layers are referred to as thick layers. These layers are excessive in terms of computation since each neuron communicates with several other neurons. When there are fewer neurons in the layers, it is preferable to employ this calculation since it would take a lot of time and computing resources to finish the work otherwise. Due to its wide network, it could lead to overfitting [82].

Convolutional Neural Network (CNN): CNNs are a particular type of neural network designed to interpret visual input, such as images and sounds. As a result, it is employed in operations such as OCR, object localization, and other picture-processing procedures. CNNs may also be used to recognize text, voice, and video. The pixels that make up an image decide how much white there is in the image. Every pixel in a picture stands for a part that the neural network will look after [83].

Generative Adversarial Networks (GAN): Generative adversarial networks (GAN), a single learning process, find and integrate examples from the data as a result. It develops a discriminator- and sub-models-generator-divided model. The generator model aims to create fresh images from the data. With convolutional neural organization, the discriminator model often functions as a parallel classifier. The producer model aims to trick the discriminator model into recognizing the picture, whilst the discriminator model aims to identify fraudulent photos effectively. Both models make every attempt to improve their performance [84].

The Multilayer Perceptron (MLP): Between the input and output layers of a feed-forward neural network, known as an MLP, there are frequently one or more hidden layers. In this case, the perceptron can use any activation function and is not constrained to being a strictly binary classifier. As stacked nonlinear transformation layers for learning hierarchical feature representations, MLPs may be thought of in this way.

Autoencoder: Unsupervised models called AEs (AEs) attempt to recreate the input data at the output layer. The middle layer, commonly referred to as the bottleneck layer, frequently represents the main characteristics of the incoming data. The denoising AE, sparse AE, variational AE, marginalization AE, and contractive AE are only a few examples of the several types of AEs. However, Figure 8 depicts the comprehensive taxonomy of the DL approach applied to smart healthcare.

3.2.1.   DL process

In the case of the DL concept, the extensive aggregation of many neurons between the unseen layers communicates the subsequent propagation concerning the neural network from the activation layer to the end layer. This is made possible by the activation function's nonlinear adjustment. The following is the equation:

The activation function is denoted by the letter m in the equation above, where V^T represents a matrix of weights. DL neural network training aims to enhance the neural network's capacity for decision-making and data-suitability [85]. Additionally, the parameters must also be optimized in order to establish the best decision boundary. This requires comparing the loss function, often known as the dissimilarity between the sample's original value, k, and its foretold value, f(l). The equation below can be used to determine the loss function:

The symbol c in the equation above stands for loss. The reduction of the overall loss of all training data is the aim of neural network operation [86]. Feature diagrams can be created considering this layer. The p^th feature diagram is repeatedly expressed as m^p, and the parameter that makes up the convolution kernel V^p and s_p, then the equivalency below can be acquired:

The input feature map is represented by one in the previous equation, the activation function is portrayed by f(. ), and the hidden layer parameter is represented by V. Each layer region will be evaluated individually during the convolution computation, and the features will be added to the matrix element multiplication summation and overlaid deviation, as shown below:

In the depicted equations, s denotes the degree of deviation, while Z_h and Z_h+1 denote the convolutional input and output of the h + 1^th layer, which is also known as a feature mapping. The size of Z_h+1 is denoted by H_h+1, and the measurement of the length and width of the feature are supposed to be the same. The pixels of the related feature map are denoted by Z(i, j), and the amount of channels for the feature map is denoted by P. The convolutional layer parameters are f, b₀, and k, which describe the magnitude of the convolution kernel, convolution stride and padding count. The aforementioned equation can be used to compute the output feature map's sketch if the following conditions are met: The step dimension is B, the filter volume is C, the inlet feature map extent is W, and the amount of zero padding added to the border is K.

Similar to how an Artificial Neural Network (ANN) is connected, the completely connected layer is also connected. More variations have been proposed as a result of the CNN network's current rapid development, including the Institute for Global Communications Network (IGCNet) [87], gated CNN (GCNN) [88], GoogleNet [89], AlexNet [90], and the visual geometry group network (VGGNet) [91].

3.2.2.   Role of DL in smart healthcare

Deep Learning in MRI: Standard magnetic resonance imaging (MRI) is becoming more and more dependent on gadolinium-based contrast agents (GBCAs). Though GBCAs were believed to be benign, they were linked to nephrogenic basic fibrosis, a real, progressive and often severe illness. By using less gadolinium, this testimony may most likely be mitigated. Unexpectedly, low-partition contrast-enhanced MRI frequently produces unsatisfactory demonstration images.

Use of DL in CT Scan: Despite their widespread usage in medical procedures, CT methods pose a radioactive danger. When receiving a physical after more than seven years, for instance, many people would be subjected to 200 chest X-rays worth of radiation, which is the radioactive component of a CT scan. We can see in Figure 7, how the building blocks and layers together are incorporated together to classify a binary decision.

3.3.   Key features of smart healthcare

Even yet, there are some crucial procedures that must be done before real-time applications are introduced.

3.3.1.   Combining role of ML-DL in smart healthcare

Finding signs of brain bleeding: The probability that a patient will recover relies on how quickly a brain hemorrhage is discovered. Analysis shows that these bleeds are understandable anywhere between 12% and 51% of the time and that every year, diseases associated with brain bleeds result in the deaths of almost six million individuals. The level of patient care is greatly reduced by such a vast range of variance.

Robotic or Robotic-assisted surgery: In summary, the surgical advancement phase we are currently entering may be described by the integration of surgical robots with AI and data gathered from robotic systems. Motion pursuit data will reveal insights that might fundamentally alter how we now do medical procedures. When the next generation of surgical robots is driven by AI and large information analytics, three of the most promising AI systems—Alpha Go, IBM Watson, and ML algorithms—may be integrated into surgical robots [92].

Cancer diagnosis: Malignancies may often be found and identified using ultrasonography, X-rays, MRI, and CT. Unfortunately, certain cancers cannot be accurately recognized enough for these technologies to consistently save lives. Examining microarray sequencing profiles is an alternative technique, although doing so needs several processing hours unless the study is AI-enabled. At this point, it has even been demonstrated that AI-enabled diagnostic algorithmic criteria is as effective at identifying potential skin cancers.

4.   Integration on ML-DL, ML-healthcare, DL-healthcare, and ML-DL-healthcare: Advancement

4.1.   Overviews of ML-DL techniques

Whereas DL artificial neural networks can only learn after processing millions of data points, ML versus neural networks is better suited for business scenarios that can collect thousands of data points for the training datasets. Yet, ML maintenance calls for a group of specialists who can manually select features, categorize them, and modify algorithms when they stray [93]. DL involves substantially less human input and is capable of self-correcting when high forecast accuracy is at stake. So, the industry that requires DL has a vast array of unstructured data. In other situations, applying ML algorithms can help you save time [94]. Figure 6 depicts the basic block diagram of the ML-DL method in a healthcare system.

4.1.1.   Integration benefits

4.2.   Overviews of ML-healthcare techniques

ML techniques are utilized in healthcare to improve patient outcomes by utilizing the expanding amount of health data made available via the IoT. These techniques have a lot of potential applications but also substantial drawbacks. Medical imaging, the natural language processing of medical records, and genetic data are the three principal applications of ML. These fields frequently center on diagnosis, detection, and prediction. By utilizing data and minimizing human engagement, ML is utilized in healthcare to enhance the efficacy and overall character of care. Furthermore, as patient data becomes more available, ML technology will become even more important for healthcare providers and health systems to interpret medical data. In Table 3, the analysis of current works focused on the analysis of smart healthcare and the integration of ML in smart healthcare is shown.

4.2.1.   Integration benefits

ML has a spacious assortment of achievable applications in clinical custody, from enhancing patient data, diagnosis and treatment to cutting costs and improving patient safety. Only a few advantages that ML applications in healthcare can offer medical professionals are listed below:

4.3.   Overviews of DL-healthcare techniques

The majority of DL-related news items in the industry at this moment are about diminutive-scale pilots or research schemes that are at present in the pre-commercialization stage. DL, on the other hand, is quickly being applied in creative systems with significant therapeutic applications. Some of the most promising used cases incorporate cutting-edge patient-facing applications as well as a few unexpectedly advanced ways for boosting health information technology (IT) user experiences [97]. Table 4 presents the overview of existing works based on the integration with DL and healthcare with their methods, focused points, and highlighting the drawbacks or challenges of their works.

4.3.1.   Integration benefits

4.4.   Overviews of ML-DL-healthcare techniques

The current healthcare system places a high value on ML and DL to enhance the treatment provided to patients, doctors, and other healthcare personnel, as well as their overall health. It has been shown to be advantageous to use ML and DL for diagnosing acute illnesses, image scrutiny, drug breakthrough, drug shipment, and smart health surveillance. The effectiveness of previously unavailable treatment options will increase as ML-DL based applications are soon widely implemented and loaded with real-time patient data from a variety of healthcare systems in a variety of nations [99], [100]. Figure 9 illustrates how the ML-DL method has been integrated into the healthcare system.

4.4.1.   Integration benefits

DL and ML have made outstanding changes in many areas, including business, industry, schools, colleges, and healthcare systems. As a result of the availability of a wide range of online and offline facilities, we can say that additional changes are being discovered in the medical profession [101], [102]. For the automatic detection of cancer cells, DL is essential. While the DL uses ML to operate independently or autonomously, the ML can address a wide range of issues but calls for human participation. DL, as a contrast to ML, immediately solves the whole problem. Young children, coma patients, and elderly patients can all benefit greatly from DL when it comes to heart disease diagnosis [103].

5.   ML and DL in smart healthcare: Applications

5.1.   ML-based Applications in Healthcare

The majority of ML applications are employed in the healthcare sector to improve patient outcomes and the level of care, despite the fact that new applications are constantly being developed. You can specialize on ML, which has many applications in the healthcare industry [138]. Figure 10 shows how ML is used in the healthcare sector.

5.2.   DL-based applications in healthcare

Nowadays, a lot of DL news stories focus on preliminary research projects or small-scale trials. However, DL is gradually making its way into groundbreaking technology with highly beneficial medicinal applications. Moreover, some of the most intriguing use examples involve cutting-edge patient-facing applications and a few surprisingly well-established techniques for improving the user experience of health IT [139]. Figure 11 illustrates the DL applications in the healthcare sector.

5.3.   ML-DL-based applications in healthcare

Every problem may be admitted by dint of DL, and performance based on data is constantly superior. Traditional DL can create quick and effective ML for language acquisition while assuring accurate patient data since ML is nonlinear. Patient information or data can be displayed in a variety of forms, including text and images, much like how a patient's status is shown in real time [140]. Real-time results are crucial in IoT applications for healthcare and other industries; therefore, this has been a major problem [141], [142]. DL has the ability to enhance the engineering system for worth control. ML allows for the evaluation of a dynamic system that is capable of changing its environment and recognizing data in various formats. Clinical professionals can diagnose patients' diseases using DL and ML, which can identify normal and abnormal patient data. The basis of DL models is CNN learning, a sort of ANN used for image visualization, including ultrasound, MRI, CT scan and X-ray. In addition to the least, Figure 15 provides a summary of the taxonomy of the ML-DL learning approach used in healthcare applications.

However, ML-DL supports so many alternative techniques, and they are updating their performance from time to time. In Table 5, there is an overview of various techniques that are used in different case-solving purposes in the vast area of healthcare. Again, Table 6, discusses the datasets that have been utilized by the researchers for the task of solving critical issues in healthcare sectors. In Figure 12, there is a brief analysis of the varieties of techniques mostly used in this field. In addition, according to this figure, the X-axis describes the name of the methods, and the Y-axis explains the frequently used methods. The observation is collected from the Table 5 data along with some others studies from [51], [117][122], [143]–[154]. The CNN model is used for the maximum time. On the other hand, data analysis in ambiguous sectors is a sweltering concern. In Figure 13, a large-scale analysis of the accuracy rate generated from different methods and techniques incorporated with ML-DL in healthcare is presented. Studies from [37], [155]–[164] are plotted in the graph to make an analysis that different methods return different accuracy. The X-axis represents the study in the field of healthcare, and the Y-axis represents the accuracy level found from those studies.

6.   ML and DL in smart healthcare: Challenges and opportunities

6.1.   Open issues and challenges

To build a solid model and address some significant medical issues, ML or DL approaches for the healthcare system need real-time datasets. However, gathering data is a challenging endeavour, particularly for the medical system, as it may include patient parties or occasionally hospital authority private information. As a result, there is less interest in data sharing from both patients and hospital staff. The introduction of federated learning [220], [221], which eliminates the need for data transfers from the source, or the use of Blockchain-based security measures for secure and transparent data transfers from source to destination are two privacy measures that need to be focused on in order to deal with this issue [222], [223].

The weight of the model is yet another issue with DL deployment. DL produces a heavy model since it uses numerous layers. This substantial model integration becomes extremely challenging in other situations, such as those involving the IoT and wireless sensor networks [224]. To circumvent this difficulty, a lightweight DL model that can perform just as well as typical DL models but is less physically demanding must be used, such as Vision Transformer [225], [226]. Advanced ML or DL models, on the other hand, need a substantial amount of real-time data. However, gathering healthcare data is a difficult task. Thanks to cutting-edge AI techniques that offer data augmentation methods [227], [228]. This augmentation methodology, such as the GAN or time series GAN (ts-gan) [229] method, enables the generation of synthetic data from the already-existing real-time data. In this situation, CNN might be improved in some way and evaluated. Before using the models in the field, a cost analysis for them should be conducted using specific statistical techniques. Figure 14 displays an overview of the use-case sectors and challenges of the ML-DL method in the healthcare system and Table 7 shows existing research in the field of machine and DL technology and its associated research gaps and future directions in this research field.

6.2.   Overall discussion and limitations

6.2.1.   Overall discussion

The aforementioned study aids in understanding how DL and ML automate a number of crucial elements and procedures that medical infrastructure depends on and cannot function without. For example, we find that AI can find every possible way to predict the root cause of an incurable disease and find a related antigen to dissolve it, thereby mitigating its effect; that NLP can automate hospital management processes while saving a significant amount of time and resources; that DL could speed up the diagnosis process. Applications of DL and ML solutions span a range of healthcare domains and tackle contemporary problems in their respective fields of study. Diagnosing different frequent diseases and ailments is one such subfield that adds to the annual rise in the worldwide death rate. A significant portion of this issue can be resolved by utilizing AI-based solutions, particularly DL. Doctors and other workers would also find it challenging to do their work as efficiently at this high pace of population growth due to overwork and a lack of free time. Hospitals can therefore use ML and DL to assist them in overcoming this issue. These technologies can learn to predict every conceivable outcome at every stage with the least amount of error, allowing for the acceleration of further processes while saving crucial time and resources. Using a virtual assistant would be cost-effective, even for hospitals. Thus, it is possible to adopt DL and ML-based systems for diagnosis without a significant financial investment.

Though there are certain restrictions on the use of such cutting-edge technology that would be in charge of making crucial decisions pertaining to people's lives, there might be some difficulties in replacing all of the antiquated or conventionally used technical methodologies that are being used in the current healthcare infrastructure. These problems won't be resolved until every component of the current infrastructure—including the medical staff—supports enhancing the circumstances and the available resources. DL and machine learning-based technologies have the potential to solve many of the most pressing issues in healthcare, and their advancements open up a wide range of interesting and intriguing medical applications that could raise the standard and effectiveness of healthcare delivery. Nevertheless, much study in this area is still needed to fully realize the technology's promise. Moreover, one of the main challenges and obstacles to achieving it is the requirement for a sizable dataset.

6.2.2.   Limitations

We draw attention to the following limitations that we believe will be encouraging for DL in healthcare in the future.

6.3.   Future aspects

6.3.1.   ML with AI

Future developments in ML and AI are anticipated to have a substantial impact on a variety of industries, including healthcare, finance, transportation, and many more. It is anticipated that AI will advance even further, eventually taking on activities that were previously thought to be solely humanly feasible. The following are some major factors that will probably influence how AI and ML develop in the future:

Edge AI: The increasing need for real-time AI processing and decision-making at the network edge will stimulate the rise of edge AI. Since data processing happens locally on the edge device, there's less need to transmit sensitive information to the cloud. This can enhance privacy and security, particularly in applications where data confidentiality is a priority of healthcare with low latency. Devices like wearable health monitors or medical imaging equipment can benefit from local AI processing. Real-time patient monitoring, remote diagnostics, predictive analytics for patient outcomes, emergency response systems, are the optimized results brought up with the help of edge AI.

Explainable AI (XAI): XAI refers to a set of strategies and methods designed to make AI systems' decisions and processes comprehensible and interpretable by humans. The lack of transparency in some AI models, particularly DL and complicated machine learning algorithms, can be a substantial impediment to their adoption in essential applications requiring trust and responsibility. Explainable AI seeks to overcome this issue by revealing how AI systems make certain decisions or predictions. Some common strategies used in XAI are explanation through features, local interpretable model-agnostic explanations (LIME), shapley additive explanations (SHAP), decision trees, and rule-based models. As AI systems become more intricate and pervasive, there will be a greater demand for the openness and explicability of their decision-making processes.

DL on Graphs: This is a brand new field of study that concentrates on using graph-structured data for DL models.

RL: RL is anticipated to become more crucial in AI, particularly in fields like robotics and autonomous systems. Overall, ML and AI have a bright future and are expected to make significant advances in a wide range of industries and applications.

6.3.2.   XAI with healthcare

The healthcare sector may undergo a change thanks to XAI, which increases the openness, accountability, and reliability of AI systems. Future uses for XAI include:

Enhance diagnosis and treatment strategies: By explaining their predictions and suggestions, XAI algorithms can assist medical professionals in making judgments that are more accurate and well-informed.

Increase patient participation and training: By using XAI, patients can receive more individualized information about their health and available treatments.

Simplify clinical studies and research: XAI can aid in the discovery of trends and connections in immense quantities of clinical evidence, facilitating the discovery of novel therapeutic approaches and treatments. Moreover, XAI may boost trust in AI systems by creating them to be more accessible and understandable, which would encourage healthcare practitioners to use them with more assurance.

6.3.3.   DL with 6G in cyber-physical system

The efficiency and effectiveness of cyber-physical systems (CPS) could be significantly increased by integrating DL and 6G. Future applications of this pairing include:

Strengthen monitoring and efficiency: In order to improve governance and increase the effectiveness of CPSs, DL algorithms may be used to examine the enormous volumes of data that these systems produce.

Boost safety and dependability: It will be feasible to improve security and reliability by collecting and analyzing data in instantaneously utilizing 6G's high speed, low latency, and high throughput capabilities. This will enable quicker and more informed decisions [230].

Boost cybersecurity: DL may be used to recognize and stop potential safety risks, and 6G's encrypted communication features can assist in assuring the privacy and protection of sensitive files. Again, integrating DL and 6G in CPSs can greatly enhance their capabilities and bring new levels of efficiency and safety to various industries.

6.3.4.   Federated learning with 6G in healthcare

The usage and sharing of healthcare data has the potential to be completely changed by the merging of federated learning and 6G. Future applications of this mixture encompass:

Enhance customized drug: By analyzing patient data from many sources using federated learning algorithms, more individualized treatment plans based on a wider variety of data are possible.

Improved data privacy: This is possible because of 6G's strong encryption features, which may help make sure that private patient information is secured when it is transferred amongst healthcare professionals.

Streamline clinical trials and drug discovery: Federated learning can aid in the discovery of trends and connections in vast volumes of medical studies, making it simpler to locate potential new remedies and cures.

Facilitate telecare: The deployment of telehealth apps that can bring direct remote consultations between healthcare practitioners and patients can be supported by 6G's high bandwidth and low delay.

6.3.5.   XAI with 6G in healthcare applications

The openness, accountability, and trustworthiness of AI systems in the healthcare industry might be significantly increased with the integration of Explainable XAI with 6G. Future applications of this pair tend to involve:

Improve outcome: XAI programs can offer more clear and explicable suggestions and forecasts to healthcare professionals, assisting in increasing the precision and knowledge of their choice.

Enhance patient participation and skills training: 6G's high speed and low latency can support the deployment of telemedicine petitions that can enable real-time remote counseling sessions between healthcare providers and patients. XAI can be used to educate patients while also providing them with more personalized information regarding their treatment and care alternatives.

Aid to the invention or development of new Medicine: XAI and 6G may be used to evaluate enormous volumes of clinical data, making it simpler to uncover potential new treatments and cures while still guaranteeing the confidentiality and safety of critical material. This will streamline clinical trials and drug development.

6.3.6.   6G-based security in smart healthcare

The integration of 6G-based security in intelligent healthcare has the potential to significantly enhance patient data protection and confidentiality. Potential developments of 6G's security features usually involve:

Privacy of Data: Improved data privacy is possible because of 6G's secure communication and data storage capabilities, which may help guarantee that private medical information is safeguarded and kept private.

Boost networked medical equipment' confidentiality: 6G's encrypted communication capabilities can assist in preventing hacking and unauthorized access to medical devices, ensuring their safe and trustworthy functioning.

Reliable Telehealthcare: Facilitate the implementation of secure telemedicine apps that can enable real-time virtual consultations between medical physicians and patients while simultaneously protecting the privacy and security of sensitive data. 6G's rapid adoption and reduced latency can facilitate this.

Safeguard Clinical Trials: 6G's secure messaging capabilities may help ensure the confidentiality and preservation of sensitive trial information, making it simpler to carry out big clinical studies and create novel treatments and medications.

7.   Conclusion

In this article, we conducted three different fields, including ML, DL and healthcare. This study thoroughly mines a massive amount of information based on tremendous areas such as ML, DL, and smart healthcare. Most significantly, we go into great depth on how ML and DL methods are now being used in smart healthcare applications. We presented vital features of ML, DL, and healthcare in a state of the art manner and significantly incorporated among them as ML-DL, ML-Healthcare, DL-Healthcare, and ML-DL-Healthcare. Notably, we focused on applications, challenges, and future opportunities efficiently. Though we have provided both ML and DL processes for the smart healthcare industry and classified them into various solutions, healthcare is wider than just ML and DL areas. Still, numerous issues and challenges need to be handled adequately. In the future, we plan to include security, privacy, and relevant issues with special attention. In addition, we provide further prospective study directions in this area.

Use of AI tools declaration

The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.