Federated learning-based AI approaches in smart healthcare: concepts, taxonomies, challenges and open issues

Abstract

Federated Learning (FL), Artificial Intelligence (AI), and Explainable Artificial Intelligence (XAI) are the most trending and exciting technology in the intelligent healthcare field. Traditionally, the healthcare system works based on centralized agents sharing their raw data. Therefore, huge vulnerabilities and challenges are still existing in this system. However, integrating with AI, the system would be multiple agent collaborators who are capable of communicating with their desired host efficiently. Again, FL is another interesting feature, which works decentralized manner; it maintains the communication based on a model in the preferred system without transferring the raw data. The combination of FL, AI, and XAI techniques can be capable of minimizing several limitations and challenges in the healthcare system. This paper presents a complete analysis of FL using AI for smart healthcare applications. Initially, we discuss contemporary concepts of emerging technologies such as FL, AI, XAI, and the healthcare system. We integrate and classify the FL-AI with healthcare technologies in different domains. Further, we address the existing problems, including security, privacy, stability, and reliability in the healthcare field. In addition, we guide the readers to solving strategies of healthcare using FL and AI. Finally, we address extensive research areas as well as future potential prospects regarding FL-based AI research in the healthcare management system.

1 Introduction

The Internet-of-Medical-Things (IoMT) revolution brought comprehensive changes in the way medical facilities operate by uplifting the quality of services [1]. IoMT devices with the ability to sense and transmit health updates of an individual, are extensively used to gather healthcare data. This information is then processed using artificial intelligence [2] to materialize a variety of healthcare applications which includes distant monitoring of patient and prognosis of diseases. Deep Learning (DL) approaches, for example, have shown potential in biomedical image analysis for earlier identification of acute illnesses by managing enormous volumes of relevant data to improve healthcare efficiency [3]. Dong et al. [4] suggested a detailed analysis based on blockchain-based FL to discuss the issue of privacy concerns using datasets to identify the current state of this problem. The study suggested that blockchain-based FL has developed significantly over the last five years, and the technology is now being applied to the fields of IoT and smart healthcare applications, such as recording patient health data, medical image analysis, to study data related to analysis of cancer related illness, and relevant economic data. Smart healthcare systems have typically focused on centralized AI capabilities stored in the cloud for learning and data analytics [5,6,7]. Due to raw data transfer, this centralized approach is inefficient in terms of communication latency and cannot achieve high network scalability. More specifically in e-healthcare, personal data are subjected to regulations like the Health Insurance Portability and Accountability Act of the United States (HIPPA) [8]. Furthermore, a centralized AI system may be considered unacceptable in future healthcare systems, and instead must be disseminated throughout a large-scale IoMT network. As a result, switching to distributed AI approaches at the network edge for scalable and privacy-preserving intelligent healthcare applications is critical.

Federated Learning, a new distributed interactive AI concept, is especially promising for smart healthcare since it allows numerous clients (such as Hospitals) to participate on AI training whilst maintaining data privacy. As a result, the authors investigated the application of FL in smart healthcare extensively [9]. To begin, we will discuss current breakthroughs in FL, as well as the reasons and prerequisites for adopting FL in smart healthcare. The authors have presented a state-of-the-art overview of FL’s developing applications in major healthcare areas such as COVID-19 detection, medical data recording, distant monitoring of patients, and biomedical image analysis. In a recent research, FL was recommended for usage in a number of IoT actions, including e-healthcare and intelligent transportation system, among others. FL, for instance, has made e-health services more accessible by allowing machine learning (ML) modeling despite the absence of health data [10]. Health data owners, such as hospitals, can avoid exchanging healthcare information by employing FL. Rather, healthcare personnel can locally train the model and then share the parameters to the accumulator for data compilation. Federated Learning has presented itself as a viable method for implementing economic, innovative healthcare systems while ensuring privacy [11,12,13,14,15]. FL allows training of AI models through averaging of local updates from numerous healthcare facilities and smart devices, such as IoMT, despite the lack of local data.

On the other hand, AI has been used in a wide range of disciplines, which includes the IoT, machine vision, natural language processing, and robotics, as a result of fast spreading of AI technologies. More specifically, researchers have attempted to use AI to boost scientific analysis and analyse potential remedies, hence improving the overall efficacy of the healthcare industry [16, 17]. The advantages that AI can bring to medicine have been predicted for decades. The role of AI in biomedical engineering has even been reviewed [18]. AI and its applications in healthcare have recently made considerable strides [19]. Medical facilities with an attempt towards becoming further individualized, with a predictive instinct, preventative measures, and participatory behavior, can approach AI to assist in this endeavor. Determined from the achievements so far, we predict that AI will pursue further evolution and grow as a powerful device for future healthcare.

Many studies have been undertaken to study FL based AI related issues, including healthcare, as a result of recent advancements in the field. The publications in [9], for example, provide the basic FL idea and its supporting protocols, as well as the technical hurdles of FL design and implementation. [20] covers the security and privacy challenges in FL systems, as well as potential methods for assessing harmful threats in FL networks. The authors of [21] investigates issues in FL base AI implementation like as communication expenses, resource distribution, and security concerns. In [22] researchers looks at the junction of FL-AI and the IoT by providing a review of technical challenges in FL schemes which includes sparsification, security, and extensibility, as well as a concise analysis of FL-based AI technologies in IoT [23]. Furthermore, in [24] the authors provided a review of FL applications in industrial IoT, with an emphasis on the features and basics of FL, with little mention of FL use in healthcare. Another research [25] focuses on the technological challenges and needs of adopting FL-based AI approaches in the future of digital health. The most recent advancements in FL, for instance FL aware of resources, safe and reliable FL, the potential of privacy-boosted FL, stimulus-aware FL, and tailored FL has yet to be extensively investigated.

Table 1 List of common abbreviations with description

Table 2 Related surveys/works regarding FL, AI and Healthcare. The works are grouped based on the related technology and reported in chronological order within each group

Despite these attempts, as far as we know, no established research provides complete overview of FL-based AI uses in the sphere of smart healthcare. Furthermore, the available literature lacks a comprehensive taxonomy of FL’s use in new healthcare applications. These disadvantages motivated the authors to carry out an in-depth examination regarding the usage of FL-based AI techniques in the healthcare realm. We investigate the most updated FL-AI concepts in smart healthcare. Following that, we give a splitting analysis regarding the novel FL-based AI applications in smart healthcare, including Electronic Healthcare Records (EHR) management, distant monitoring of health, biomedical image analytics, and identification of COVID-19 traits [89]. Poll results are also shared so that readers may have a better grasp of how FL can be employed in smart healthcare. Our research intends to give a bibliometric analysis of characteristics such as authors, nations, citations, and keywords to help academics and practitioners plan future research. Finally, future prospects and research problems in FL-smart healthcare are discussed. Also, we will suggest and analyze a dominating framework of variables in this topic.

1.1 Major contributions and organization of the survey

This work discusses different leading technologies–FL, AI, XAI, and intelligent healthcare. In detail, we share a state-of-the-art analysis about their concepts, taxonomies & motivations, addressing issues and solutions and further applications. We also describe some integration’s of these technologies to better support applications in various areas. To this end, the survey’s main contributions are outlined here:

• We provide a cutting edge outline on FL-based AI in the area of healthcare, initially with basic concepts regarding FL, AI key concepts, XAI features, and a detailed discussion of smart healthcare efficiency.

• We present the lately advanced FL-AI taxonomies and emerging AI-FL integration’s and motivations applicable to intelligent healthcare applications.

• We address some technical challenges in the existing system and then directly solve the issues using vast technologies–including FL, AI, and XAI in the healthcare applications.

• Finally, we analyse the issues of different applications and direct further research towards FL-AI in innovative healthcare areas.

This study is the only potential and informative survey that integrates all the considered technologies to the best of our knowledge. Some of the notations are listed in Table 1. The rest of this survey is organized as follows: Section 2 represents a conceptual overview regarding FL, AI, and e-healthcare individually. Sections 3 offers the motivations behind the FL-AI-healthcare, FL-healthcare, and AI-healthcare integration’s, also providing taxonomies of these components. Further, the addressing different issues and solutions of these issues have been presented in Sect. 4. After that, Sect. 5 focuses on the current research challenges and future directions of the study. Finally, we conclude the survey in Sect. 6.

2 FL, AI, and healthcare: state-of-the-art

This segment presents advanced concepts regarding Federated Learning, artificial intelligence, and e-healthcare properly. In FL segments include various aspects of developments, the necessity of FL, different features of FL in 2.1. Further, we introduce AI technology and cover the parameters of various characteristics and usefulness in 2.2. We discuss diverse concepts and issues of healthcare technology in 2.3. We also present the related works regarding Federated Learning, Artificial Intelligence, Explainable Artificial Intelligence (XAI), and Healthcare in the Table 2.

2.1 Fundamentals of federated learning

Federated Learning is essentially a machine learning (ML) algorithm which allows for collective learning of a distributed model while preserving the data composed on their devices. FL that has been shown in Fig. 1, it allows for equipping ML models with decentralized data and at the same time protects user confidential data by design. FL as a shared ML model is gaining increased popularity in situations where user’s data privacy is vital [90]. In FL, local data is used to train clients using a distributed approach. The scenario involves training decentralized clients or nodes using local data and sharing system parameters. Accumulation of system or model parameters creates a global model using a server [72]. In practice, the user does not need to share their private data using FL. Rather, users drill the network locally and share the model to a central body known as the server. However, it is essential in FL for each user to effectually and feasibly transmit it’s learned model through the server. The server in FL aggregates the local data into a global network through an iterative process [91]. Fig. 1 represents the training process of the federated machine learning. Following are the steps used in this figure.

• Every device will have a local copy of our centralized machine learning application, which users will be able to access whenever they need it.

• The model will now gradually learn and train itself based on the data provided by the user, allowing it to become smarter over time.

• The devices then send the training results from the local copy of the machine learning program back to the central server.

• The same thing happens when you run the program on many devices, each of which has a local copy of it. The results will be compiled in a centralized server, but no user data will be included this time.

• The centralized cloud server now uses the pooled training data to update its centralized machine learning model, which is considerably superior to the prior version.

• Users update the program with the better model produced from their data while the development team upgrades the model to a newer version.

The FL is a distributed processing system built on AI, Blockchain, and machine learning algorithms at its core. The terms distributed machine learning and federated machine learning are not interchangeable. In Federated Learning, the information supplied to the server by each participant resembles no original data, rather a trained sub-model. Simultaneously, Federated Learning permits asynchronous transmission, which reduces requirements for communication purposes [92]. On this foundation, the federated machine learning formula may be revised as follows:

$$\begin{aligned} \underset{w}{\arg \min } L(x, y, w)=\sum _{k} p_{k} L_{k}(x, y, w) \end{aligned}$$

(1)

where, k denotes number of clients, $p_{k}$ is the weight value of the $k^{th}$ client, and the decentralized multiuser $F_{1}$,$F_{2}$,....,$F_{k}$) scenario is used for Federated Learning. The current user’s dataset ($D_{1}$,$D_{1}$,....,$D_{k}$) is available to each client user.

2.1.1 FL: Key design aspects

For constructing intelligent and privacy-enhanced IoT systems, the notion of Federated Learning has recently been presented. The major steps in the FL-smart healthcare process are as follows:

• Initialization of system and client selection: The aggregation mainframe chooses an analytic task, such as automated biomedical imaging or detection of any motion, as well as further needs like objective categorization, and learning parameters and rates. Furthermore, the endpoint chooses a number of users to engage in FL method.

• Local Training and Updates: As soon as the dataset of learning users has been identified, the network delivers an preliminary model to the clients, together with an initial global gradient, to start the distributed training. Each user teaches a local model adopting its personal data and estimates its model update. The scenario is such as, after configuring the training, the server creates a novel model, i.e., ${\mathbf {w}}_{G}^{0}$, and sends it to the users to begin distributed training. Employing its own data $D_{k}$, each client k drills a local model and assesses an update $\mathrm {w}_{k}$ by mitigating a loss function ${\mathcal {F}}\left( {\mathbf {w}}_{k}\right)$:

  $$\begin{aligned} \mathrm {w}_{k}^{*}=\arg \min {\mathcal {F}}\left( \mathrm {w}_{k}\right) , k \in {\mathcal {K}} \end{aligned}$$

  (2)

  For various FL algorithms, the loss function might be different [93]. The loss function ${\mathcal {F}}$ of a linear regression Federated Learning model may be described as: ${\mathcal {F}}\left( {\mathbf {w}}_{k}\right) =$ $\frac{1}{2}\left( x_{i}^{T} \mathrm {w}_{k}-y_{i}\right) ^{2}$ with a set of input output pairings $\left\{ x_{i}, y_{i}\right\} _{i=1}^{K}$. Then, for aggregate, each client k transmits its estimated update $w_{k}$ to the server.

• Aggregation and Download of Models: Following the gathering of all updates from the specified clients, the server uses an aggregation mechanism to combine them. For example, in Google’s Federated Averaging (FedAvg) technique [94], where the gradient parameters of local models are averaged element-wise with weights proportionate to the sizes of the client datasets, we may employ the model averaging strategy. The server then generates a new version of the global model and broadcasts it to all clients as the foundation for future local model updates in the next learning round. A real scenario is, the server combines all model changes from local clients and calculates a new version of the global model as follows:

  $$\begin{aligned} {\mathbf {w}}_{G}=\frac{1}{\sum _{k \in {\mathcal {K}}}\left| D_{k}\right| } \sum _{k=1}^{K}\left| D_{k}\right| {\mathbf {w}}_{k} \end{aligned}$$

  (3)

  by figuring out how to solve the following optimization problem:

  $$\begin{aligned} (P 1): \min _{w \in K} \frac{1}{K} \sum _{k=1}^{K} {\mathcal {F}}\left( w_{k}\right) \end{aligned}$$

  (4)

  subject to    (C1): $w_{1}$= $w_{2}$=...=$w_{K}$=$w_{G}$ The accuracy of the FL method, for example, the accuracy of an FL-based object classification task [35], is reflected by the loss function ${\mathcal {F}}$. Furthermore, the constraint (C1) assures that after each training round, all users and central server have identical training model for Federated Learning problem. Following the model’s derivation, the server transmits the new global update $w_{G}$ to all users, allowing them to optimize their trained models in the subsequent training round. The iteration of the process proceeds until the loss function congregates or the requisite precision is attained.

Federated Learning method can be deployed to deliver numerous appealing benefits to develop smart healthcare based on the revolutionary operating idea, as mentioned below:

• Enhancement of Data Privacy: In FL, local data are not required for training. This is used to train other machine learning algorithms by combining numerous local datasets and not transferring data. Local Machine Learning (ML) models are trained on local heterogeneous datasets during training. The model parameters are shared between these local data centers on a regular basis. Many models encrypt these parameters before sending them. Data samples from the local area are not shared. This increases data security and protection. In addition, a shared global model is created. As a result, risk of personal data being leaked to an external source is reduced, and privacy is ensured. FL offers security features which creates an excellent alternative for designing smart and secured IoT devices, especially in light of increasingly severe data privacy protection regulations such as the General Data Protection Regulation (GDPR) [95, 96].

• A Reasonable Balance of Accuracy and Utility: FL is capable of providing a suitable balance between precision and usefulness, as well as privacy enhancement, as compared to traditional centralized learning. Furthermore, FL training keeps the model generalizability while sacrificing nominal accuracy. As a result of FL’s distributed learning characteristic, the smart healthcare system’s scalability may be improved.

• Health Data Training at a Low Cost: FL can assist to mitigate communication expenses, such as data lag and power transmission, associated with transfer of raw data by avoiding the offloading of large data volumes to the server [97]. Because model gradients are typically much smaller than their actual datasets, FL can be deployed to mitigate communication expenses, such as delay and power dispatch associated with communication of local data. As a result, FL saves a significant amount of network bandwidth and reduces the risk of network congestion in large healthcare networks.

2.1.2 Why needed FL?

The ever growing number of IoT resources together with relevant applications means the requirement of processing large amount of data [98, 99]. The availability of big data analytics and computation methods such as machine learning and deep learning has enabled users to achieve effective data management. Artificial Intelligence applications is being successfully deployed to counter the issues related to optimized resource management, efficient selection of antenna in wireless systems and several other areas of communication networks. The traditional AI models usually necessitates the users to share individual information to a master network for learning purposes. A key concern with such techniques is the privacy of users sensitive information. FL is highly effective in areas where decision making is based on substantial data scattered over a wide range of training nodes at the same time addressing privacy and security concern [100]. Machine learning models are developed using data collected from several sources to allow for prediction. Regardless, as a result of bandwidth issues, security and storage facility, raw data transmission to a centralized location becomes unrealistic. FL acts a distributed learning model to ensure optimal learning, efficient utilization of collected raw data and transmission to a centralized place. FL also plays a significant role in advancement of smart cities as detailed by authors in [101]. Policy makers in urban smart cities can utilize FL to transmit the sensitive information gathered from IoT devices for effective management of priority assets. The framework in FL allows the users to access data without gaining personal information regarding other clients. Eventually, the updated global model constructed by the server is distributed to all clients [21]. The clients then download the new updated global model and utilize cloud distribution to understand interference on their individual devices.

2.1.3 Different features of FL

Federated Learning approach is a concept introduced by google to allow on-device learning and security of sensitive data [93]. FL has the benefit of distributed data processing and improved privacy factors. FL is suggested as means of training a ML model in [91] which does not require accumulation of all user’s data sample. In [26], authors have investigated the deployment of FL to detect malware in IoT devices. Researchers in [34] detailed the advantages of coupling vector quantization with FL to generate an efficient decentralized training model. Federated Learning also offers a platform to allow distributed analysis in the industrial IoT sector [102]. The issues of executing FL over a wireless network is studied by authors in [35]. Another key feature of FL is the ability of the server to organize the training model and analyze the contribution of every participant [103]. FL is also being applied to the Industrial Internet of Things (IIoT) by researchers to develop efficient and robust models while preserving sensitive data. Several other features of FL include its application in healthcare and medicine sectors as discussed in [12, 83]. Electronic healthcare records from multiple data sources (several hospitals) require joint data access and uploading of data to a single database.

However, due to reasons such as privacy concerns of each institution, it is quite challenging to materialize such expectations without guaranteeing security issues. FL is a highly effective method to gather EHR data from hospitals to provide a knowledge-sharing platform without the need of sharing personal data, all while preserving the privacy [104, 105]. FL in healthcare is used to detect similarity in the patient in [106] and is used for predictive modeling as detailed in [107, 108]. Authors in [109] converted electronic health records data into readable phenotypes by using a tensor factorization model to analyze data in an FL framework. A Federated Learning framework to analyze brain structural relationships for diseases is studied in [110]. Researchers in [111] used FL based predictive modeling on EHR data to analyze the advantages of early treatment of patients. In [112], FL was deployed towards a predictive analysis to estimate prolonged stay period of patients and in-hospital mortality across a number of medical facilities. Two-stage federated Natural Language Processing (NLP) to detect patients and phenotyping from EHR data for obesity and comorbidities from several medical facilities is studied in [113]. Novel Federated Learning framework for smart wearables for activity recognition and data aggregation is studied in [114, 115]. Model-centric, cross-device, horizontal, cross-silo, vertical, data-centric, reinforcement, and other types of Federated Learning exist. The three primary forms of Federated Learning are–horizontal FL, vertical FL, and federated transfer learning. Horizontal FL uses data with the similar trait space across all devices, suggesting that Client A and Client B are using the same features. To train a global model, vertical Federated Learning incorporates multiple datasets from distinct feature areas. Finally, federated transfer learning is vertical Federated Learning that uses a pre-trained model that has been learnt on a comparable dataset to tackle a different challenge. Assume the global model is $M_{FED}$ after an assignment is completed, and associated learning model is $M_{SUM}$ after data aggregation. In general, the global model $M_{FED}$ remains operational owing to parameter exchange and aggregation operations. Throughout the whole learning process of the global model’s output, there will be a loss of accuracy. The performance of $M_{FED}$ falls short of that of the aggregate model $M_{SUM}$. After stating the output of the global model $M_{FED}$ on test set as $V_{FED}$, and the output of the aggregate model $M_{SUM}$ on the test set as $V_{SUM}$, to quantify this difference. The model’s loss accuracy [116] is defined as:

$$\begin{aligned} \left| V_{\mathrm {FED}}-V_{\mathrm {SUM}}\right| <\delta \end{aligned}$$

(5)

where delta is a positive integer. However, because the primary criterion of FL is privacy protection, the aggregation model $M_{SUM}$ cannot be realized in practice.

Fig. 1

Federated Learning and its Training Process

2.2 Artificial intelligence concepts

2.2.1 AI–features

Smart devices with measuring sensors are on the rise [117]. All these devices mean an exponential amount of generated datasets with an intense opportunity to learn from intelligent systems. With the increased level of data, a system has the potential to increase its accuracy. Artificial Intelligence can be termed as the simulative approach by computers in a way such that it can be considered as intelligent by humans. There are several features of AI, including knowledge-gathering, problem-solving abilities, predicting, learning, and implementing with the capability to reason, manipulate and move. AI enables the ability to detect the fault and evaluate without the presence of an expert. An AI system has the ability to solve complex problems in an automated and accelerated manner. It has been generally used to optimize conventional technologies involving data-driven strategies. AI has subclasses such as machine learning, neural networks, and deep learning. ML is essentially an operation of artificial intelligence that permits a machine to train and grow without being preprogrammed based on experience.

• Supervised Learning: This type of machine learning employs labeled samples and applies what it has learned in the past to a fresh dataset. This will necessitate knowing the algorithm’s outputs and having the information required to train the model labeled with the accurate reactions. The algorithm correlates its actual performance to the correct output depending upon these responses, and if erroneous, it trains from these response and improves its effectiveness [118].

• Unsupervised Learning: Such a method of learning is used with information that was not labeled previously. No correlations among input and output for the model. This particular type of learning is more complex and utilized less as compared to supervised learning. When opposed to supervised learning, unsupervised learning allows the user to accomplish more complex processing tasks. These algorithms include–clustering, anomaly detection, neural networks, etc [118].

• Semi-supervised Learning: This type of learning lies in between supervised and unsupervised learning. This method of training is deployed in the event of a combined issue which requires both supervised and unsupervised learning. Data with labels are used in supervised learning, while data without labels are used in unsupervised learning, but data with and without labels is used in semi-supervised learning. The model will learn from tagged data and then apply its knowledge and patterns to unlabeled data [118].

• Reinforcement Learning: In order to teach the algorithm, this method of learning deployed a performance - based compensation technique. The trained model will be compensated for accurate output and condemned for erroneous performance in this method of learning, so it will be trained to optimise result and at the same time minimize consequence. The primary goal of health-related AI applications is to investigate correlations between treatment or preventative strategies and patient outcomes [119]. Diagnoses of disease, to create protocol for treatment, innovation in developing medicine and potential cure, and monitoring of patient are fields in healthcare where AI is used. Even though AI’s usage in healthcare is steadily rising, it is now focused on a few ailments, like cancer categorization, nervous system disease, and cardiovascular disease.

2.2.2 How does AI works?

AI can gather information from several sources and perform a task, solve critical problems or make a decision without any human instructions [120]. AI can potentially replace humans by working in wide-scale activities in the realm of industry, retail, finance, and healthcare with considerable impacts on system performance and efficiency [121] that depicted in Fig. 5. The most common AI applications in healthcare may be split into 4 groups. Among these groups, the first three classes are meant to sort large data effectively and enable rapid passage towards data to deal with challenges in healthcare facilities. These applications includes contents like assistance for the elderly and physically disabled, natural language processing methods, and basic research.

• Artificial Intelligence for Living Assistance: In order to assist aged and physically disabled individual, automated systems coupled with intelligent robots can significantly elevate standard of life. Recently, [122] released an outline of intelligent home with capabilities for patients with loss of autonomy, as well as smart models developed on sensory devices over wireless networks, data collection, and AI. NNs may be trained to detect human facial expressions as commands using certain image-processing processes. Furthermore, facial expression analysis-based human-machine interfaces (HMIs) enable persons with impairments to operate wheelchairs and robot support devices without the need to use a console or relevant sensors [123]. In, smart communication architecture (SCA), solutions for Ambient Aided Living (AAL) were created to permit AI to interpret data from various communication networks. Sensors capture information about the surroundings and individual activity in this scenario, which is subsequently evaluated via cloud computing or edge intelligence [124].

• AI in biomedical information processing: Natural language processing (NLP) used for medical applications has progressed significantly. Biological question answering (BioQA) aims to discover swift and correct responses to queries by users from a massive database. As such, NLP systems can be predicted to seek replies that are resourceful [125]. Initially, biological queries must be divided into several categories to achieve relevant information from reactions. For accuracy over 90%, ML can be categorize biological queries into 4 main kinds [126]. Then, using a smart biomedical document gathering process, parts of the details containing biomedical queries may be efficiently retrieved. AI has the ability to perform these work as accurately as an expert assessor to enhance efficiency and accuracy.

• AI in bio-medicine: AI offers untapped potential as a strong tool in biological research [127], in addition to acting as a eDoctor for illness diagnosis, management, and prognosis. AI has the potential to boost reviewing and indexing of research papers in biological research and development works throughout the world. Moreover, through a semantic graph-based AI technique, researchers may effectively accomplish the complex work of outlining the literature on a specific topic [128]. Furthermore, for a significantly large research article, AI can equip biomedical researchers with searching and ranking of literature. This allows researchers to enhance scientific conception, which are critical aspects in the area of biomedical research.

2.2.3 Explainable AI (XAI) concepts and features

XAI refers to approaches and techniques for using artificial intelligence in such a way that the solution’s outcomes are understandable by humans. It contrasts with the “black box” concept in machine learning, when even the designers are unable to explain why an AI made a particular decision. XAI could be an example of the social right to explanation in action. Even if there is no legal right or regulatory obligation, XAI can improve a product or service’s user experience by assisting end-users in trusting that the AI is making good decisions.

The goal of XAI in this manner is to explain what has been done, what is being done now, and what will be done next, as well as to reveal the knowledge on which the actions are based. These features allow you to (i) evaluate what you already know, (ii) dispute what you already know, and (iii) produce new assumptions.

Properties of XAI: Transparency, Interoperability, and Explainability all work as an interface between human and AI systems. They consist of AI systems that are both accurate and understandable by humans [129]. Although the semantic implications of these phrases are identical, they confer distinct levels of AI that are acceptable to humans. For further information, the ontology and taxonomy of XAI can be seen here at a high level:

• Transparent Model: K-nearest neighbors (kNN), decision trees, rule-based learning, Bayesian networks, and are examples of typical transparent models. Although transparency as a feature does not guarantee thet a model will be easily explainable, the decisions made by these models are frequently transparent. [130]

• Opaque Model: Random forest, neural networks, Support Vector Machine (SVMs), and other opaque models are common. These models are not transparent, despite the fact that they frequently reach excellent accuracy. [131]

• Model-agnostic: XAI techniques that are model-agnostic are created with the goal of being widely applicable. As a result, they must be flexible enough to operate purely on the basis of matching a model’s input to its outputs rather than relying on the model’s intrinsic architecture. [132]

• Model specific: Model-specific XAI techniques frequently take advantage of prior knowledge of a particular model and try to bring transparency to a specific type of one or more models.

• Explanation by simplification: We can identify alternatives to the original models that explain the prediction we’re interested in by simplifying a model via approximation. For example, we can build a linear model or a decision tree around a model’s predictions, then use the resulting model to explain the more sophisticated one.

• Explanation by feature relevance: This is analogous to the concept of simplicity. After all possible combinations have been explored, this form of XAI technique seeks to evaluate a feature based on its average expected marginal contribution to the model’s choice.

• Visual explanation: This type of XAI strategy is built around visualization. As a result, the data visualization approaches can be used to interpret the prediction or judgment made based on the input data.

• Local explanation: Local explanations replicate the model in a small area, usually around a single instance of interest, and provide information about how the model works when it encounters inputs comparable to the one we’re interested in describing.

Advancement of XAI: Despite the benefits of intelligent systems, the XAI research initiative raises concerns about giving them too much power without the ability to explain the decision-making process that lies beneath such complex systems to domain experts (e.g., doctors, lawyers, financial experts, etc.) in terms and in a format that they can understand. This not only aids in comprehending specific decisions made by such systems but also pushes researchers to develop more human-like (anthropomorphic) solutions, as well as encouraging further research and understanding of the brain as a natural information processing phenomena.

Because of the growing relevance of this topic, NIST published Four Rules of XAI in August 2020, which describe the following key principles [133] that an AI must follow in order to be designated a XAI:

• Explanation: According to this principle, an AI system must provide proof, support, or explanation for each decision it makes.

• Meaningful: According to this principle, the AI system’s explanation must be clear and meaningful to its consumers. Because diverse groups of users may have varied wants and experiences, the AI system’s explanation must be fine-tuned to fit each group’s unique traits and needs.

• Accuracy: this principle states that the AI system’s explanation must correctly reflect the system’s processes.

• Knowledge limits: according to this theory, AI systems must recognize circumstances in which they were not designed to operate and, as a result, their answers may not be credible.

2.3 e-Healthcare system

e-healthcare may be characterized as providing constant healthcare assistance to patients through technical tools and techniques such as information and communication technology, cellular technology, medical support, and wireless facilities [134]. Electronic health records, scientific journals, and similar archive on the internet, visual, and audio consultation with doctor, or internet-based process to connect with medical personal, give feedback to doctors, transfer test results, and so on are examples of e-health systems that can revolutionize the healthcare services they provide. The e-healthcare system may give general support on a larger scale, such as management assistance, healthcare service delivery, and so on, as well as specialized help, such as citizen health data [135]. The provision of health care via the internet and technology is carried out by combining existing tools and assuring the quality of the services supplied [136]. The spread of the contagious coronavirus is an issue that needs immediate attention. With the advent of new variant of coronavirus, medical personnel are facing an enormous challenge to address the health issues of the infected individual. The complexities of COVID-19 coupled with scarce resources and rising healthcare cost has encouraged many to access remote health management services using computer or similar devices [137,138,139,140]. IoT plays a fundamental act in achieving automated distant health care services. IoT and smart devices are used to transfer the health data of a patient to the cloud server to access health information of an individual [141]. The IoT and its application for healthcare management services can be divided into three steps [142]: i) gain access to the location of a patient or healthcare staff ii) to identify and authenticate an individual, and iii) to automate data collection using sensory devices. Electronic health records (EHR), decision based support systems, online consulting, and telemedicine are a wide range of components related to eHealth. Computers play a vital role to deliver therapeutic session to patients. EHR data of patients serves as the base for prescriptions originated using computers, which offer several benefits, such as linkages to software that emphasizes medication. Two key features in eHealth have ethical implications for the handling of EHRs. The first is professional, whereas the second is technical. The technical issue is promptly resolved. Assuming that there is a fiduciary physician-patient connection in eHealth, the ethical obligations for the treatment of patient data in general and EHRs in particular that arise from the fiduciary relationship in standard health care also apply in eHealth [143]. The novel coronavirus develops pneumonia like symptoms with an individual suffering from respiratory diseases. The prognosis of a patient infected with COVID-19 related illness is limited with high mortality [144, 145]. The implementation of AI in healthcare sectors [47, 146] have created an urge amongst researchers to pursue innovative methods to develop e-healthcare systems. Authors in [47], have developed an AI based system to differentiate between pneumonia of a coronavirus infected patient from that of common pneumonia traits. AI based algorithm is used in [147] for earlier detection of kidney related illness amongst diabetic individual. Researchers in [148] utilized AI to provide specific diagnosis and treatment for patients.

FL-AI is possible to deliver numerous appealing benefits to promote smart healthcare, as mentioned below, based on the innovative operational concept:

• Improvements to Data Privacy: Only local updates are needed by the central network for AI training in the FL-AI-based smart healthcare system, while raw data is stored at medical facilities or equipment. This approach limits the danger of personal user data being leaked to an external source, ensuring better level of client security [149].

• A Reasonable Balance of Accuracy and Utility: FL-AI is capable of providing a suitable balance among accuracy and usefulness, as well as privacy enhancement, as compared to traditional centralized learning. Furthermore, FL training keeps the model generalizability while sacrificing nominal accuracy. As a result of FL’s distributed learning characteristic, the smart healthcare system’s scalability may be improved.

• Health Data Training at a Low Cost: FL-AI may assist decrease communication costs, like latency and transmit power, associated with raw data transfer by evading the dumping of large data to the central network because model gradients are often considerably less in size than their real datasets [97]. As a result, FL saves a significant amount of network bandwidth and reduces the risk of network congestion in large healthcare networks.

e-Health is a broad word that encompasses the integration of healthcare and technology to serve individuals better and save healthcare expenditures. Some industrialized nations that have been adopting health-related activities have consistently prioritized eHealth [150]. However, implementing eHealth has the potential to enhance the whole health care system in both developed and developing nations. eHealth is referred to as an end-to-end method that may be used in any situation.

3 Taxonomies and motivations of integrating FL with AI in healthcare system

In this section, we make a group of unique taxonomy architectural views based on FL-AI and FL-AI-healthcare in 3.1. On the other hand, we provide the integration benefits among the FL, AI, and intelligent healthcare technologies in 3.2.

Table 3 Overview of recent works analysis on FL in Healthcare

Table 4 Overview of recent works analysis on AI in Healthcare

3.1 Taxonomies of FL with AI in healthcare

Federated Learning, a distributed collaborative AI model, is specifically appealing for intelligent healthcare since it allows different clients (for example, hospitals) to collaborate on AI training without the need to share local data. As a result, we have put together a detailed analysis on FL’s application in advanced healthcare. To begin, we introduce the current breakthroughs in FL, as well as the reasons and prerequisites for adopting FL in smart healthcare [31, 157]. Following that, the latest FL ideas for intelligent healthcare are reviewed, which includes FL for resource management, FL aware of security concerns, incentive FL, and tailored FL. Following that, we present a cutting edge overview of FL’s developing applications in major healthcare areas such as data management, distant health monitoring, biomedical image analytics, and so on [9, 158, 159].

Fig. 2

Taxonomies of Federated Learning with AI

The proposed taxonomy for Federated Learning (FL) with Artificial Intelligence (AI) is shown in Fig. 2, and the taxonomy for FL with AI techniques in healthcare is shown in Fig. 3. We showed numerous AI and machine learning algorithms based on learning, reasoning, and discovering in these taxonomies. We divided the privacy level as well as the communication architecture. Different FL-based algorithms were also examined. Furthermore, the development of FL-AI techniques in healthcare, the applications of Smart technologies in healthcare have been systematically analyzed, and the security and privacy concerns, as well as medical issues have been portrayed. In the suggested taxonomy, we thoroughly examined the healthcare sector with the interconnection of FL-AI [4, 83, 160].

Fig. 3

FL-AI with Healthcare Overview

3.2 Integrating of FL with AI in healthcare

3.2.1 FL-healthcare integration

The current world is focused on data analytics, particularly in the medical sector. Healthcare data, including prescription information and supplies, patient data, healthcare professional information, and affiliated organizations responsible for insurance or other financial-related transactions, has become increasingly important for data analysis. However, the data on the healthcare sector is dispersed. They are not in the same format, and the data is susceptible in nature, as it includes information from the medical industry. The most sensitive of them are insurance sector data, which cannot be transmitted from one sector to another. Without data transfer, processing with traditional learning and training methodologies is impossible, resulting in a significant gap in data analysis.

Fig. 4

Federated Learning with Healthcare

As a result of the integration of Federated Learning with healthcare data informatics, the analysis of extremely sensitive data has become significantly more efficient, this integration scenario shown in Fig. 4. In this situation, data is not transported from source to destination in order to merge or generate a dataset for analysis; rather, the local client trains their model with their data and connects with other clients via the server bypassing only the results. As a result, the data remain secure at their origin, and the network, gains access to the server’s training results. This integration will benefit both suppliers, such as the organization in charge of producing and supplying the necessary equipment to a hospital, and users of the application, such as the general public. As the linking sectors are vast in this era, and data production is also vast, it is difficult to identify malicious users among the thousands of connected clients; therefore, by leveraging the benefits of FL, the clients’ data is preserved safely with their own model, and data leakage is prevented to avoid any malicious data modification in the system. Moreover, we present the overview of current analysis based on Federated Learning in healthcare in Table 3 There are various applications in the healthcare sector where FL integration will increase the overall effectiveness of the process analysis:

• FL can assist to improve the security and privacy of medical industry data analysis and prediction by incorporating a data protection system [25, 161, 162] .

• FL has also allowed predictive modelling depending on several sources, from which physicians can access extra information about the potential threats and advantages of providing earlier treatment to patients [163].

• Among the many applications of FL-based healthcare models for prediction, similarity among different patients is one [164].

• Medicine resistance or therapy against various diseases, as well as survival rates or explanations, can be examined with strong privacy towards sensitive data [165].

• Hospital mortality prediction, duration of stay, or rather the rate of admission can also be examined using FL-based models for the proper management of some critical care units in the hospital while protecting the privacy of in-hospital data [161, 166].

3.2.2 AI-healthcare integration

As discussed in the preceding section, data from many sectors have risen considerably in the current context. Traditional or statistically based prediction for the medical sector is problematic with the expanding amount of data because it involves human data collection. There is a significant demand for automation throughout the data gathering, processing, and result creation processes. Again, a system that can receive information from multiple sources and perform a task, solve crucial problems, or make a decision without human intervention would be extremely valuable to the healthcare business. In the world of healthcare, the word artificial intelligence plays a very important and critical function in interpreting the human thinking process through the computer employing many efficient AI-based algorithms to make the difficult examination easy for healthcare professionals [167]. That is, making sensible use of machines that draw key conclusions based on data fed to them as input. The objective of integrating AI in the healthcare sector is to blend numerous aspects that may influence a specific decision and produce a proper evaluation of the result offered by the AI model using ML or DL. In general, AI models are useful for making key decisions such as ICU surveillance, mortality prediction, proper usage of hospital resources, drug inspection and reaction, and so on. By using the AI method and the vast amount of data created by the healthcare industry, such as the EHR, the remedy and conclusion may be formed efficiently for future reference of the doctors or clinicians [168]. In addition, we present the overview of recent work analysis based on Artificial Intelligence in healthcare in Table 4.

Recently, explainable AI (XAI) has been gaining popularity in cyber-physical systems such as smart healthcare. Cancer identification from MRI pictures is a critical problem in the medical field. However, with an effective cancer detection process, the chance of survival may increase because the cure is achievable at an early stage of cancer. With sophisticated XAI, the reason for the diagnosis of a certain cancer is apparent and visible to the system’s user [169]. The XAI intends to reveal the internal mathematical calculations used to solve the black box known as the AI. Thus, by using this method during the diagnosis phase, clinicians can benefit from the use of XAI technologies [170]. There are several sectors where AI is integrated with healthcare like–

• The agent-based AI can interact with humans and provide comfort based on the needs of the patients. This can be useful when the patient is elderly and requires close monitoring and care. Patients with chronic diseases, in particular, require frequent monitoring of certain tests such as heart rate, blood sugar levels, and so on. All of this is simply obtained with the AI-based agent [171].

• Another significant responsibility in the medical field is assessing medication effects and forecasting how a drug will affect a human being. One essential method in this scenario is the combination of AI with computer-based medication creation and analysis [172, 173].

The benefits of AI integration in the healthcare sector are enormous which is shown in Fig. 5. Though it has various concerns, such as control over AI-based devices, it must be handled with extreme caution or it can be destructive to a nation. In this regard, the security of these systems must be powerful enough to withstand any form of intruder attack or to take appropriate steps to raise the alarm when any inappropriate data tampering occurs in the system.

Fig. 5

Use case of Artificial Intelligence in Healthcare System

3.2.3 FL-AI integration

The system with autonomous control is incorporating more and more real-time data aggregation as new technologies evolve. Intelligent systems were elevated to a new level owing to the ML and DL models. The data produced by each IoT sensor has increased exponentially with the IoT-based system, and the cloud storage system has made data storage easier in many ways. However, collecting data and feeding it to a model with advanced DL to develop an AI-based model is a very important and delicate operation. If data is altered from source to destination by an attacker, the forecast will be inaccurate, and because this method deals with highly sensitive data, the security of the data will be compromised. In the above scenario, there are a variety of methods for ensuring data security, one of which is FL integration. It produces a model and transmits a replica of the model to each local client instead of transferring data from the local machine to the central server. With their own local data, the client trains the model and transfers the new parameters or weight to the central server. During the whole process, no data must be exchanged; only the result must be transferred from clients to the central network. At the server, the local training outcomes are averaged, and the revised weights are then relayed back to all clients connected to the server. There is no need to transfer data because the raw data is kept in the local area, and thus data security may be assured. The integration benefits of FL and AI covers a huge area. For example;

• FL-based mechanism aggregation at the edge may be beneficial because the volumes of information gained from nodes are quite high in rate using such approaches [174].

• Device heterogeneity can be preserved in a safe data training and evaluation environment [175].

• The FL with AI is more powerful thanks to its raw data protection method and local and global privacy levels [176].

• Integration of an FL-based AI model can result in a more energy-efficient solution [177, 178].

4 Addressing and defending challenges on FL with AI in healthcare system

This section is the most strong side compared to the other survey. However, firstly, we discuss the existing problems and issues–security, privacy, reliability, confidentiality, scalability, etc. in 4.1. In addition, we present the different solutions strategies that have been shown in Fig. 6 based on the proposed techniques–FL, AI, healthcare in 4.2. Furthermore, we summarize the overview of Federated Learning with AI in Healthcare as well as the key applications area, challenges, and solution in Table 5.

Table 5 Overview of Federated Learning with AI in Healthcare Research covered with key applications area, challenges, and solution

4.1 Existing challenges in healthcare system

4.1.1 Security and privacy concern

The current healthcare system relies on the internet to transfer data and other sensitive information. The Internet of Medical Things, or IoMT, is this term. It has drawn significant attention in the twenty-first century since it can carry information from one portion to another in a concise period of time and at a very low cost. However, along with the benefits, there are also drawbacks. One of the most serious topics is security, and privacy [197]. This type of healthcare solution includes data transfer, mobile application usage, telemedicine, and accessing the information of the healthcare remotely. In the IoT-based healthcare system, several attacks can hamper the total loss of the system. Preference, side-channel attack [198], data duplication using the tag, physical hampering of the sensor devices, or sensor tracking are some of the attacks that may cause issues in the perception layer [199]. Then, some attacks from the intruder can access the information sent from the hardware system by directly attacking at the decryption point or by weakening the authentication procedure such as Eavesdropping attack, reply attack, MITM attack, rouge attack, DoS attack, DDoS attack and these type of attack causes to hamper the network layer of the IoMT [200].

Fig. 6

Challenges and Solution Scenarios of Federated Learning

4.1.2 Reliability problem

Following the aforementioned security concerns, the current healthcare system’s reliability has also become a source of concern. By embracing session hijacking in the existing smart healthcare system, an attacker can take control of the session as well as the data being transferred in that session, potentially causing a major problem for the entire system. The intruder can then gain access to the IoT network using a side script. When a faulty SQL code is injected into the database, it can jeopardize the system’s vital information [201]. Packet interception can also impact the loss of information which is in the form of identity while a user tries to authenticate or log in. An intruder can encrypt the personal data of a healthcare system user via a ransomware assault. As a result, these attack possibilities raise a question of trustworthiness. For instance, when intruders are capable of encrypting compassionate information from a system, the question of whether the system is reliable or safe to transfer data through remains open [202].

4.1.3 Confidentiality issues

Existing healthcare solutions necessitate the relocation and merging of data from multiple nodes to provide a long-term solution. However, because the data being moved is very sensitive, this data migration creates a confidentiality concern [203]. Another issue in IoT-based systems is heterogeneity. Because the sensors are from diverse vendors, the system’s anonymity isn’t always guaranteed. Furthermore, securing data security while transferring susceptible data from one node to another remains a difficult task [204]. However, in order to produce information, the data must travel from one end to the other. The transit of data or sensitive information causes security breaches, which, in turn, causes a confidentiality problem in an IoMT-based system [205].

4.1.4 Scalability problem

In today’s healthcare systems, a huge number of inbound and outbound networks must be managed. As a result, there is a problem with supporting a large number of high-speed wireless systems without causing system speed degradation [206]. However, with the current growing number of networked devices, ensuring speed, quality, and cheap cost at the same time is becoming increasingly difficult. With the growth in data traffic, the network’s bandwidth is once again challenged [207]. Furthermore, the network’s architecture has scalability challenges due to its hybrid nature, including several providers’ devices. Another challenging task is to manage several devices remotely and without any loss in time or management. Because in this type of framework, there is a need to integrate devices along with their data and information coming from different aspects [208].

4.2 Solving the existing challenges using FL with AI in the healthcare system

4.2.1 Role of FL

To execute learning tasks in something like a decentralized environment, where one central server controls the global learning procedure and numerous devices train local models with their own data [21] is the process known as the Federated Learning process. FL can be used in a variety of settings, including healthcare and education. Models based on machine learning logic necessitate a huge quantity of information to train the model from various parts of the system, such as hospital data in the healthcare sector. The healthcare industry handles a lot of sensitive data that, if tampered with, could bring the entire system down. In this case, FL devised a privacy-preserving approach that does not require the transfer of sensitive data from the source, instead of training local models with their own data and then sharing results with the global network [209]. As a result, this strategy can protect the privacy of patient information by using an iterative process in which model parameters are constantly exchanged and modified, but raw data is not disclosed. As a result, to preserve the security and privacy of the hospital system, FL provides a privacy-preserving method while also providing the benefits of machine learning models to solve numerous difficulties related to healthcare systems [193, 210].

4.2.2 Role of FL with AI

The new breed of AI is called Federated Learning, and it is based on a decentralized training and learning mechanism. A new AI model architecture that can be spread over a large number of devices without knowing personal user data. This technique assures user privacy and data security because it does not access raw data [21]. The demand for user privacy has evolved in tandem with the advent of artificial intelligence. The FL, which sits at the intersection of AI-cloud and AI-user equipment, plays a key role in bringing the UE and cloud together to create a significantly more secure and robust system than earlier architectures. Again, due to FL’s properties, imbalanced and non-IID information can be handled appropriately, ensuring that the AI system’s performance is not limited and that a large device in multiple contexts may be trained with their local data and local model without difficulty [211].

Fig. 7

Results analysis of different research in terms of Accuracy (%)

Fig. 8

Results analysis of different research in terms of AUC (%)

4.2.3 Role of FL with XAI in healthcare

Although AI in a federated context can address the concerns described previously, deep learning has an explainability difficulty. Because deep learning algorithms are typically black-box models, there is no acceptable explanation for a given prediction. Deep learning in healthcare is limited due to this ambiguity. Explainability is vital in healthcare since we need to explain why a given prediction for sample input is made to persuade a clinical healthcare practitioner and a patient. Humans can understand and describe how an AI system reached a decision using Explainable Artificial Intelligence (XAI) [212]. Artificial intelligence approaches such as deep learning with the integration of FL have recently played a revolutionary role in healthcare, particularly in terms of diagnosis and surgery. In these fields, these strategies have been proven to be beneficial. Some deep-learning-based diagnosis jobs are even more accurate than human medical specialists. The deep learning model’s black-box character, on the other hand, hinders its explainability and practical application in medicine.

Amoroso et al. provided an XAI framework for breast cancer therapeutics in their paper [213]. The experiment findings showed that the framework could outline an essential clinical feature for the patient and design oncological medicines using the clustering and dimension reduction method. Dindorf et al. proposed an explainable disease in a spinal posture-dependent classifier [214]. EI-Sappagh et al. suggested an RF model for detecting Alzheimer’s disease (AD) progression and diagnosis in [215]. The fuzzy rule-based system could also generate natural language forms to assist patients and physicians in comprehending the AI model. Peng et al. established an XAI framework to help doctors predict the prognosis of hepatitis patients in [216]. In this work, the authors compare intrinsic XAI approaches like logistic regression (LR), decision tree (DT), and kNN to complicated models like SVM, XGBoost, and RF. The authors also used the SHAP, LIME, and partial dependence plots (PDPs) post-hoc approaches. Sarp et al. proposed a CNN-based model for chronic wound categorization in [170], after which they used the XAI approach LIME to explain the CNN-based model. Rucco et al. introduced an XAI program to diagnose glioblastoma in [217], which combined topological and textural characteristics. Table 6 presents overviews of different datasets used in FL, AI, XAI, and Healthcare.

Table 6 Overviews of Different Datasets used in FL, AI, XAI, and Healthcare

4.2.4 Analysis of performance metrics

To give a manual describing different binary classification assessment metrics, we use examples of metrics and evaluations collected from a number of published papers, including top AI algorithms used in FL. Note that, we do not comment on the quality of these studies; rather, we use them to demonstrate how different measures provide diverse interpretations of an AI model’s quality. A mixture model can be used to explain binary classification problems, with the overall data distribution modeled as:

$$\begin{aligned} p(X, \alpha )=\alpha p_{P}(X)+(1-\alpha ) p_{N}(X) \end{aligned}$$

(6)

where X represents data samples, $p_{P / N}$ denotes the positive/negative class distributions, and $\alpha$ the parameter of the positive class, calculated as $\alpha =\frac{N_{P}}{N_{P}+N_{N}}$, with $N_{P / N}$ the total number of positive/negative class data samples. The four entries in the confusion matrix are the key parameters for determining the metrics for a binary classifier which are discussed below.

$$\begin{aligned} {\mathbf {M}}=\left( \begin{array}{ll} \mathrm {TP} &{} \mathrm {FN} \\ \mathrm {FP} &{} \mathrm {TN} \end{array}\right) \end{aligned}$$

(7)

• True Positive (TP): The number of accurately categorized positive samples is referred to as the true positive.

• True Negative (TN): The number of correctly categorized negative samples is known as the true negative.

• False Positive (FP): The number of samples incorrectly categorized as positive is known as the false positive rate.

• False Negative (FN): The number of samples incorrectly categorized as negative is referred to as false negatives.

The following measures are used to assess the performance of leading AI algorithms or models used in FL:

• Accuracy (ACC): The ratio of correctly identified samples to the total number of samples in the evaluation dataset is the accuracy. This metric is one of the most often utilized in medical applications of machine learning. The accuracy ranges from [0, 1], with 1 representing properly predicting all positive and negative samples and 0 representing successfully predicting none of the positive or negative samples.

  $$\begin{aligned} \text {ACC}=\frac{\text {TP}+\text {TN}}{\text {TP}+\text {FP}+\text {TN}+\text {FN}} \end{aligned}$$

  (8)

• Area Under the ROC Curve (AUC): AUC is a metric that combines effectiveness throughout all categorization criteria. The AUC indicates the percentage of correctly categorized positive samples. Rather of measuring absolute values, it assesses how well predictions are ordered. It assesses the accuracy of the model’s predictions regardless of the categorization level used.

• Specificity (SPEC): The specificity indicates the percentage of correctly categorized negative samples. It’s measured as the proportion of correctly categorized negative samples to all negative samples. The specificity is limited to [0, 1], where 1 denotes flawless negative class prediction and 0 denotes inaccurate negative class sample prediction.

  $$\begin{aligned} \text {SPEC}=\frac{\text {TN}}{\text {TN}+\text {FP}} \end{aligned}$$

  (9)

• Sensitivity/Recall (REC): The recall, also known as the sensitivity or True Positive Rate (TPR), is the ratio between correctly classified positive samples and all samples assigned to the positive class, and it is calculated as the ratio between correctly classified positive samples and all samples assigned to the positive class. The recall is limited to [0, 1], with 1 representing perfect positive class prediction and 0 representing inaccurate positive class sample prediction. This statistic is also considered one of the most significant in healthcare, as it is desirable to overlook as few good cases as possible, resulting in a high recall.

  $$\begin{aligned} \text {REC}=\frac{\text {TP}}{\text {TP}+\text {FN}} \end{aligned}$$

  (10)

• Precision (PREC): Precision is calculated as the ratio between correctly identified samples and all samples assigned to that class, and it indicates the proportion of recovered samples that are relevant. The precision is limited to [0, 1], with 1 representing all accurately predicted samples in the class and 0 representing no valid predictions in the class.

  $$\begin{aligned} \text {PREC}=\frac{\text {TC}}{\text {TC}+\text {FC}} \end{aligned}$$

  (11)

  where C stands for “class” and can be either positive (P) or negative (N) in binary classification (N).

However in Table 7 the results of several research which have been discussed throughout this article in terms of accuracy and AUC have been shown and also depicted graphically in Fig. 7 and Fig. 8, respectively. In the graph the X axis represents the studies that achieved the scores and the Y axis represents the performance score in percentage. Inside every bar the data that have been considered in the work is indicated and on the top of every data bar the method used have been shown.

Table 7 The Results of Several Research Discussed in this Article in terms of Accuracy and AUC

5 Open issues and future direction

We cover this section on the additional issues as open discussion and guide some matters for further discussion in 5.1 and 5.2, respectively. We also present the existing work analysis of FL-AI with Healthcare based on central idea, applications, approaches, open issues and further opportunities in the Table 8.

Table 8 Existing work analysis of FL-AI with Healthcare based on Central Idea, Applications & Approaches, and Open Issues and Further Opportunities

5.1 Open issues

Federated Learning is a continuous process where a model keeps updating over time. That’s why different challenges and problems were brought out with further developments. The issues are open to the researchers so that they could be solved efficiently within a short time. Since the whole procedure runs over the internet, privacy and security are always a major concern in this field, especially when it comes to competent health care systems. The health-related data is susceptible compared to the others. In this regard, people are utilizing various technologies into the health care systems such as SDN, Blockchain, etc [234, 235]. However, it is an advanced machine learning method that needs large datasets [233, 236]. A suitable dataset is essential for such problems. It is not so easy to manage such datasets for medical cases, which is a demerit for this field. Also, the datasets are not present at the user end, where the distributed training would be performed as a minor update of the global model. Apart from this, the performance obtained from the health data is not satisfactory because the domain of the information is not the same and also contains outliers into the data points. Processing the health data and training machine learning models is expensive and time-consuming. Efficient methodologies are still critical to find out.

However, to upgrade the performances of such systems in the field of medication, researchers are trying to apply novel methods and technologies. Increasing the security of the healthcare systems remains the vital concern of the personnel. The patient’s personal and sensitive data needs to be processed to develop smart healthcare systems. To make them more reasonable and available everywhere, the Internet of Medical Things is being used. So, the security could be given utilizing SDN, Blockchain, NFV, etc [237, 238]. In this era of artificial intelligence, we can use data augmentation to imitate the existing datasets to prepare more data since Federated Learning requires a vast amount of data. More data analysis needs to be performed to find out an optimal training time and maximum performance. The generalization power of the current machine learning models is not satisfactory when it comes to the medical sector. Some melioration of CNN could be designed and tested in this context. The cost analysis for the models should be performed using some statistical methods before applying them to the field.

5.2 Future directions

5.2.1 Role of FL in data science

Data generation in all industries is extensive with the increasing usage of IoT-based methods in the current system. This is the data scientist’s current challenge. One of the most important is to ensure data security, and the sensors that are associated with the IoT-based mechanism have less memory. They do, however, generate a massive amount of data for every nanosecond. As a result, storing or transferring these data in a short period of time is both difficult and critical [239]. The benefit of utilizing FL in this scenario is that the data storing mechanism is avoided and distributed among multiple clients, and because the data is trained locally on each client, the security of individual data is ensured [240].

5.2.2 AI with blockchain in security

Artificial Intelligence is the most potent agent for any applications areas. Sometimes AI acts as a communicator bridge of real-life applications. On the other Blockchain is a provider of distributed ledger for any vulnerable components [241]. BC also offers a temporary ledger for performing security and confidentiality parameters efficiently. AI applications need a secured platform to continue its activities in the applications environments. BC easily collaborates with AI and shows their better performance for the desired system. In the BC environment, AI agents are solely attached during communications. Combined, they produce a highly protected environment that can continue their services to the current and future users incredibly [242, 243].

5.2.3 Controlling robots using FL-AI

In today’s modern technological environment, robots are used in numerous sectors to relieve humans of many dangerous tasks. Again, AI-based agents have a very efficient processing capability and can handle heterogeneous data transactions. Furthermore, with the modern cloud storage system, effective data and information storage with very high memory storage in the cloud system is conceivable. Continuous learning and training are now possible because of the ability to retrieve data from numerous sources in a short period of time. In today’s Industry 4.0 era, the implications of autonomous agents have grown. Robots are used in a variety of industries, including hospitals, educational institutions, the military, and even several hotels. However, because training an agent necessitates the collection of large amounts of data from many sources, it is vital to provide secure data collection for training. In this regard, FL is a well-suited solution since the key demanding task, data transfer, is forbidden in this scenario, ensuring data privacy while maintaining the autonomy system’s efficiency [244].

5.2.4 FL with big data analytic

The primary idea behind Big Data is that must-have data volume, velocity, and variety. The basic job in extensive data analysis is to collect data from various sources and generate a dataset with a diversity of data from various sources. With the development of wireless networking resources such as 5G or even 6G, as well as IoT and Industry 4.0, this has become easier. All of these developments aid in data collection and in a real-time environment. The industrial sector collects and generates big data from a variety of sources, including heterogeneous sources. This technique is difficult because when data flows from source to destination, it can be manipulated by an intruder or hacked in the middle, resulting in a violation of critical user data privacy. FL is a new breed of AI technology in which the ML models are locally trained using the data of the local system. Continuous training is ensured by using the source data, as are the significant data criteria. As a result, data privacy is also ensured, and there is no need to transmit sensitive user data from one location to another [245].

5.2.5 FL with blockchain in industry 5.0 applications

The future IR 5.0 is designed with the primary goal of connecting smart machines and the general public. The usage of IoT devices allows for the collection of data even in remote regions, and with the gathered information, communication may be established without any additional effort on the approach is the data collected from the IoT devices and transferred to the owner. However, the acquired data is raw and may not be in the suitable format for the computer to understand the nature of the data. This is when the ML algorithm’s AI is put to use. The essential computations are carried out using AI algorithms, and the resultant analysis is promptly processed using edge computing. But, there is a challenge of protecting sensitive data from the attack of the intruder. As the transmission of data from device to device, there is a possibility of attack at the nodes of the data hub be attacked by intruders. In this case, FL and blockchain can be a possible solution where the data of each side is collected, trained, and processed on each local place, and the result is passed to the server. The initial parameters then passed through the blocks of the blockchain technology for secure transmission [246, 247].

5.2.6 Applying FL-blockchain for accessing medical data

An essential issue is examining medical disease or medical sector data from the healthcare system. However, medical data is extremely sensitive, and any changes could result in a massive loss for the entire system. In this case, Federate learning combined with the blockchain mechanism can produce a secure yet efficient model for the smart healthcare system that benefits from sophisticated learning and training algorithms while without transferring sensitive data to the network. For the secure aggregation, the initial parameters, number of rounds, and the keys, which is the public one of all clients, are collected in the secured blockchain. While an alternate block is generated after the verification procedure is completed, the parameters are obtained by the clients or the local models. Thus the medical data through the blockchain and Federated Learning security is ensured, and an efficient model can be assured [232].

5.2.7 FL with healthcare in smartindustrial IoT

With the increasing use of IoT devices, it is being employed by various industries, including education, healthcare, data science, data analysis, and so on. In the era of smart IIoT, the necessary data for analysis, such as MRI images, test samples of a specific disease, information of a patient staying in the hospital, or even data of hospital staff or hospital management, is easily gathered and sent to a central server for further processing in the healthcare system with the help of IoT [248]. However, as these technologies have advanced, collecting data has become the main draw for attackers. As a result, while these data are being sent from source to destination, an attacker can simply gain access to them and take over the system. In this regard, FL may be a viable solution because it keeps data safe on the local machine while processing tasks are performed at each client. The output is then sent back to the central server for additional examination using a secure scenario.

5.2.8 AI-FL in smart healthcare using big data

There is a high demand for standardization throughout the data collection, processing, and result generation processes. Again, a platform that can gather information from numerous sources and perform a task, solve critical problems, or make a decision without the need for human participation would be tremendously beneficial to the healthcare industry. In healthcare, AI performs a vital and crucial role in understanding the human thinking process through the computer using various effective AI-based algorithms to make complex examinations easier for healthcare professionals [83]. Data aggregation using FL and smart wearables with AI benefits to create an intelligent and secured system, predicting mortality and length of hospital stay period using electronic health records where the FL model assists the secure data collection and building a brilliant AI model and predicting mortality and length of hospital stay period using electronic health records where the FL model helps the secure data collection and building an intelligent AI model.

6 Conclusion

The recent progress in healthcare has designed a keen interest in the research community as well as FL to integrate AI in the desired networks. This survey comprehensively mines the immense information regarding emerging topics–FL, AI, XAI, and e-Healthcare. Moreover, we discuss in details the cutting edge developments of FL and AI in intelligent healthcare applications. The FL-AI, FL-Healthcare, and AI-Healthcare have been incorporated with each other significantly. Also, we address different issues–security, privacy, reliability, scalability, and confidentiality using our mentioned terms. Although we have studied both FL and AI-XAI techniques for the healthcare system and categorized them into different types of solutions, however, healthcare is not limited to only AI and FL techniques. Still, various problems and challenges that need to be addressed. The following study analyses the progress of security, state-of-the-art discussion, benefits of integration, taxonomies, and open issues have been presented. Additionally, we have offered several future research guidance in this regard.