A Comprehensive Survey on Federated Learning Techniques for Healthcare Informatics

Healthcare is predominantly regarded as a crucial consideration in promoting the general physical and mental health and well-being of people around the world. The amount of data generated by healthcare systems is enormous, making it challenging to manage. Many machine learning (ML) approaches were implemented to develop dependable and robust solutions to handle the data. ML cannot fully utilize data due to privacy concerns. This primarily happens in the case of medical data. Due to a lack of precise clinical data, the application of ML for the same is challenging and may not yield desired results. Federated learning (FL), which is a recent development in ML where the computation is offloaded to the source of data, appears to be a promising solution to this problem. In this study, we present a detailed survey of applications of FL for healthcare informatics. We initiate a discussion on the need for FL in the healthcare domain, followed by a review of recent review papers. We focus on the fundamentals of FL and the major motivations behind FL for healthcare applications. We then present the applications of FL along with recent state of the art in several verticals of healthcare. Then, lessons learned, open issues, and challenges that are yet to be solved are also highlighted. This is followed by future directions that give directions to the prospective researchers willing to do their research in this domain.


Introduction
Healthcare and related services help prevent illness, treat it when it occurs, and promote people's physical well-being. Healthcare providers are increasingly incorporating technology into patient registration, data monitoring, lab testing, and self-care applications. Consequently, people are able to plan ahead while still being in a position to make good decisions about their physical or mental disabilities.
In every organization, data collection is critical. Data can be utilized to forecast current trends and future events. Particularly in healthcare, a large amount of sensitive data is generated and it is very hard to manage and secure the private data. Data security is becoming an increasingly important concern for users. Tere is a need to develop innovative ways for handling and securing sensitive data in healthcare sectors [1][2][3]. In order to implement ML models with multiple privacy-preserving methods, it is difcult to construct frameworks and infrastructural facilities [4].
Clinics, moreover, work under stringent privacy standards and can face regulatory, logistical, or ethical restrictions requiring data to stay local. FL is a potential approach for such implementations since it can lighten the stress on a system and allow personal communication among diferent technologies/institutions [5].
Health records in the public sector are often scattered and confdential, making it difcult to attain reliable outcomes. For example, diferent clinics have electronic health records (EHRs) with various patient demographics, which are complicated to exchange among hospitals due to their delicate existence [6][7][8][9][10].
FL enables healthcare records that are located across diferent institutions to be connected without revealing personal information [11]. Figure 1 shows the result of a diferentially private analysis, which ensures that anyone viewing it will conclude the same thing (answers 1 and 2 are virtually the same). Te FL concept was frst proposed by Google in 2016, for Gboard, a virtual keyboard app for touchscreen mobile devices that supports more than 600 languages. FL efectively overcomes the limitations of classic ML methods by emphasizing data. In contrast to traditional centralized ML methods, which require datasets to reside on a single server, FL minimizes security and privacy concerns by keeping local data stores [12,13]. FL has distinct privacy advantages over ML models [14,15]. Tese special features of FL make it appealing for health research, where a large fraction of the population may want to contribute to novel health fndings but have restrictions about sharing their personal and private data. Figure 2 illustrates the applications of FL for personal healthcare.
Recently, several researchers have presented survey papers on FL. Dai et al. [16] presented a systematic strategy for replicated statistical analysis in which the Newton-Ralphson technique and an alternating direction method of multiplier (ADMM) framework are employed to conduct shared solutions. Yang et al. [17] describe foundations, infrastructures, and methods for FL, as well as privacypreserving techniques used for FL. Kumar et al. [18] initiated a unique learning technique that uses blockchain technology to detect COVID-19 in respiratory computed tomography. Xu et al. [19] employed machine learning to diagnose COVID-19 from computed tomography gathered in several clinics across China. Tese models perform better with FL and also generalize well to models based on just one region. Also, the authors in [20][21][22][23] have discussed FL and its applications in various domains.
Similarly, some of the researchers have presented papers on FL for healthcare. Lee et al. [24] gave an example of how to develop patient similarity learning programs that are federated across institutions without protecting the confdentiality of their patients. Kim et al. [25]developed tensor factorization models from massive electronic health records for use in FL environments. Vepakomma et al. [26] designed multiple setups for a global deep learning (DL) algorithm named SplitNN [27] that allows medical groups to train DL algorithms together. Silva et al. [28] studied neural.
Topological correlations across diagnoses, as well as medical samples, illustrate their FL framework. Liu et al. [29] used a distributed strategy to do user presentation modeling as well as overweight-associated phenotyping and got impressive outcomes.
Pftzner et al. [30] focus on how FL can be applied in healthcare. It implies that real-time data usage is not feasible while preserving patient data. Also, it deals with the pseudonymization of a few felds, which are again retraceable. It discusses how to distribute the load of the training process onto the FL. It identifes some open challenges, like privacypreserving hyperparameter optimization, entity resolution for vertically split data, and efcient ways of using encryption.
Rieke et al. [31] discuss the current FL eforts for digital health and their impact on stakeholders, clinicians, patients, hospitals and practices, researchers and Artifcial Intelligence (AI) developers, healthcare providers, and manufacturers. It highlights FedAvg and FedProx algorithms. It also specifes that though FL has been a challenging solution in neatly addressing issues related to sensitive medical data, which may open novel research and business avenues and can improve patient care globally, there are also so many open technical questions that have not been answered yet.
Xu et al. [11] focus on the current state of FL, including but not limited to the healthcare sector. It discusses the use of FL in healthcare as well as some of the challenges associated with the combination of FL and healthcare, such as incorporating expert knowledge, personalizing health, and obtaining model precision.
Nguyen et al. [32] discuss advanced FL designs that would be useful for federated smart healthcare, as well as the important applications of FL in smart healthcare, such as federated EHRs management, federated remote health monitoring, federated medical imaging, and federated COVID-19 detection.
Antunes et al. [33] examined the systematic literature review of current research about FL in the context of EHR data for healthcare applications. Te assessment of these articles reveals a variety of eforts to attain and provide best practices for protecting training data privacy.
A summary of the key fndings from the above discussion can be found in Table 1.
Te potential of FL in healthcare, such as efcient handling of sensitive medical data, improving the data quality and model precision, and proper electronic health records handling has not been focused on in the open literature [11,[30][31][32]. Moreover, a comprehensive discussion of the impact of FL on disease diagnosis and applications in medical imaging, the Internet of things (IoT), and the   Computational Intelligence and Neuroscience 3 COVID outbreak is still missing. Particularly, we provide a state-of-the-art survey of the applications of FL in various key healthcare services such as medical imaging, IoT, COVID outbreak, managing electronic health records (EHRs), and health cooperation. In the last part of this review, there are some potential directions for FL in healthcare.
For the purpose of better comprehension, the defnitions of the abbreviations used in the paper are summarized in Table 2.

State of the Arts and Contributions
2.1. ML for Healthcare. AI methods are increasingly used to support experts in the medical feld due to their efectiveness in detecting and classifying diseases. A few ML approaches like prediction, categorization, grouping, and learning techniques are used in various felds, like image analysis, speech recognition, and healthcare [34,35]. Te state of the art of ML techniques used for computer-aided diagnosis to detect breast cancer from various imaging techniques has been investigated [36].
Multiple sectors were being suggested to utilize COVID-19 ML approaches, including clinical practice, healthcare quality forecasting, and monitoring and identifying diseases [37][38][39][40][41]. Te use of ML and DL has become a fundamental method of knowledge discovery in several industries. Large, diverse datasets are essential for the success of data-driven applications. However, it is difcult to get medical datasets. FL solves these problems by allowing group work by centralizing information, and it is already being utilized in digital applications across the health feld [31].

FL for
Healthcare. FL brings together all the healthcare institutions, allowing them to share experiences and ensuring their privacy. In this scenario, the enormous medical dataset will help improve the quality of the ML model. In healthcare, FL is often used to study a range of tasks, like client identity, client knowledge, diagnostics, identifying possible hospitalizations, detecting death rates, intensive care unit admission periods, etc. ML algorithms require vast, comprehensive datasets, and FL provides controlled, indirect access while protecting patient privacy. FL's promise is simple: by permitting ML from pseudo data, it will overcome personal information quality issues. FL environments apply their respective policy implementations with security measures to each provider, and they also monitor Internet connectivity. Figure 3 shows how FL may afect the way AI models are taught while also benefting the overall wider healthcare ecosystem.
FL ofers great assurances on analytics of health data. Tere are uses for both providers (such as developing a prediction strategy for such risk of chronic illnesses using electronic health records (EHR) [42]) and consumers (patients) (for example, cardiac medical exam by wearable device with electrocardiograms [43]). Te federated model learning process is designed to allow academic researchers to easily maintain sensitive patient data.
Federated Transfer Learning (FTL) is another architecture for FL. FTL [44] is used to train models using data from a diferent source. FTL recently attracted signifcant attention in diferent industries, particularly healthcare [45]. FTL employs cryptography and estimation to verify that confdentiality is successfully preserved to prevent potentially exposed client data. FL is an efective solution for healthcare, as it connects data from multiple silos without requiring a transfer of all patient data. Te future of healthcare is dependent on the development of innovative technologies that respect data privacy.

Impact of FL on Disease
Diagnosis. Te process of developing an understanding of a situation by using clinical reasoning and utilizing information acquired based on observations is called disease diagnosis. Tere are many steps involved in diagnosing a disease because this is an essential part of medical science. A source of uncertainty exists at every step of the diagnostic process. A diagnosis begins with acquiring knowledge frst from the client's diagnostic test, as well as knowledge provided from laboratory tests and other medical diagnosis methods.
A diagnosis has an enormous impact on patients, in terms of both care and research. Tere have been multiple views on the nature of diagnoses, such as a process and classifcation scheme, or a previously planned set of categories. Clinical decision-making takes place according to the proper understanding of the health problem of the patient; therefore, accurate and timely diagnosis is key to receiving the best possible health outcome.
Te FL method for early diagnosis of Type-2 diabetes uses feature selection algorithms and federated multilayer perceptron models. Furthermore, a comparison between a centralized ML model and a decentralized FL model is made

Paper Organization.
Te majority of this article is based on the following: Section 2 describes FL's background and motivations. Section 3 describes applications of FL in verticals of healthcare. Sections 4 and 5 include recommendations regarding future directions of research based on lessons learned from previous reviews. Section 6 discusses the paper's conclusion.

Background and Motivations
Tis section introduces an overview and status of disease diagnosis using FL, fundamentals of FL, and motivations of using FL for healthcare and discussed the main contributions of the survey.

Overview and Status of Disease Diagnosis Using FL.
FL is a paradigm that aims to collaborate on data management and privacy issues by using evaluation metrics without transferring samples [47][48][49]. Initially, this methodology was introduced in a separate discipline, and it has lately been adopted in the medical industry since it overcomes the challenges which are typically encountered while intending to collect patient data. In the context of electronic medicine, it suggests FL allows fndings to be obtained collectively between entities without sharing patient data, such as in the form of a universal model. Te strength of FL substitutes to prevent sensitive training data movement beyond its frewalls. Healthcare has been transformed by wearable devices in numerous ways, including improvements in patient care, rehabilitation, and disease management. Tese devices generate data that can be used to detect early signs of cognitive illnesses like diabetes and dementia. However, the personalization and privacy issues they generate have raised concerns among users. An FL framework called FedHealth is being developed by Chinese researchers in order to minimize the risk of security breaches. FedHealth aims to address the vulnerabilities of today's healthcare industry. FedHealth software creates capable ML models based on FL and homomorphic encryption to protect users' privacy.
It is based on four primary components, according to Jindong. In the frst place, the server's cloud model is trained based on publicly available data. It is then disseminated to all users, who may subsequently train the model using their own data. Once you have achieved this, a new cloud model can be created through model aggregation. FedHealth uses transfer learning techniques to ensure that each organization has a uniquely tailored model after improving its cloud model. Figure 4 depicts the team's framework.
FL allows data isolation to be solved by combining user data to build ML models. FedHealth can update the cloud model and user model simultaneously as soon as new data is available. In this way, the more a wearable device or application is used, the more customized it becomes and how the model is viewed. FedHealth is not currently being implemented, but it does ofer new avenues for wearable medical applications. According to Jindong, this framework can perform diagnoses of other diseases and illnesses, including coronavirus disease, using radiology test images.  Computational Intelligence and Neuroscience makes it possible for several workers to learn a strong and generic ML model by providing information so that key issues like user privacy, information security, authentication rights, and exposure to large datasets are addressed. Tese technologies cover a variety of sectors, like military, mobile communications, IoT, and healthcare [50]. FL is designed to build a teaching practice with many individual variables containing network elements by actively transferring raw data, for future reference, into deeper neural networks. Te main principle consists of training individual designs on available information and generating variables (for example, weights and distortions of a classifer) for the generation of a modeling framework maintained across all connections between neighboring controllers. Figure 5 shows how local model parameters are transferred to a primary controller to generate a large reference model. Once the federated model has been created, it is then returned to the clients for initial training and iterative improvement. Participants can also have their own computing resources. FL will be used in practice by large organizations, and it will be important in promoting security systems whose training data is being spread around the world.

Motivations of Using FL for
Healthcare. In healthcare organizations, the primary motivation for using FL is as follows: (i) FL reduces data security and privacy concerns by maintaining local data stores, as opposed to centralized ML techniques, which require datasets to reside on one server [51]. (ii) Trough FL, large datasets from multiple hospitals can be readily accessed by an individual hospital without centralizing the data into one place. As a result of this practice, critical issues such as data access rights and access to heterogeneous data are addressed. (iii) Te construction of good models without biases using small datasets is very difcult as it takes a lot of time, efort, and cost to collect, curate, and maintain high-quality data from a diverse population. Te experts in healthcare and life sciences can use FL to solve the unique problem of data governance by training algorithms behind the hospital's frewall and only sharing models so that data remain secure [31]. (iv) FL captures a wide range of data variables and analyzes patients based on any of their demographic characteristics. For example, with access to electronic health records, FL can help to fnd clinically similar patients and predict hospitalizations due to cardiac events, mortality and ICU stay time [52]. (v) FL can have an enormous infuence on a variety of stakeholders like clinicians, patients, hospitals, medical researchers, and healthcare providers. (vi) FL is a potential concept for safe, reliable, and impartial models of data. FL makes it possible for several parties to work together without exchanging or centralizing datasets. (vii) FL provides AI developers with access to bigger and diversifed data packages, which better portray current patients. As a result, AI-based healthcare solutions will be able to scale globally at an unprecedented level.

Main Contributions of the Survey.
Tere are numerous existing works that discuss the enabling technologies, protocols, applications, and challenges of FL. FL facilitates communication between various parties without the need to  communicate or centralize data, thereby resolving issues with critical health information. In addition, this could lead to new business opportunities and improved healthcare worldwide. Te contributions of our work can be summarized as follows: (i) Several terms and defnitions for FL for healthcare are gathered from the available literature, resulting in a more comprehensive understanding of FL for healthcare from various perspectives.
(ii) FL is next developing and enabling applications in healthcare such as medical image processing, IoTbased smart healthcare applications, and outbreak prediction. (iii) Despite several research and development activities, many challenges and issues are imposed in FL for healthcare. We present these difculties in terms of confdentiality and safety, data heterogeneity, traceability and accountability, and system architecture. We also highlight promising research directions toward the realization of FL for healthcare. (iv) Finally, we discuss future directions of FL for healthcare applications like healthcare 5.0, including digital twins (DT) in healthcare, FL and blockchain for healthcare, collaborative robots in healthcare, FL and Explainable AI for healthcare, and FL for integration with 6G in healthcare.

Applications of FL in Verticals of Healthcare
Tis section discusses some of the potential applications of FL for healthcare.

FL Applications for Medical Image Processing.
Medical imaging has transformed the healthcare sector, enabling practitioners and scientists to discover more about the human body than ever before. Medical image processing provides techniques for enhancing and analyzing raw medical image data for selective visualization and analysis according to a given problem. Medical imaging can also help with the treatment and long-term management of a problem. As medical technology advances, doctors can detect problems that are more difcult to detect through simple external examinations. Medical imaging is important for determining the severity of an ongoing illness. Medical imaging refers to a variety of technologies that are used to see the human body in order to diagnose, monitor, or treat medical disorders. Each technology provides unique information regarding the part of the body being examined or treated, whether it is related to disease, injury, or the effectiveness of medical treatment. Fast and accurate reconstruction of magnetic resonance images (MRIs) from training data is essential in many clinical applications. In recent years, DL-based techniques have been discovered to perform better in MRI reconstruction. However, these methods necessarily involve massive amounts of data, which are difcult to generate and distribute due to high acquisition costs and medical data privacy regulations. To address this issue, we ofer an FLbased system that makes use of MRI data from several institutions while protecting patients' privacy [53].
Te FL algorithm is used to describe and identify physician-related persons and to estimate their hospitalization, deaths, and survival rates based on electronic health records (EHR) [54]. A complete brain part of an MRI [55] is helpful, and for brain tumor segmentation [56], the usage Computational Intelligence and Neuroscience and advantages of FL were further demonstrated. Recently, the fMRI [57] classifcation technique was used in fnding a suitable risk factor for several diseases. FL will help in detecting various diseases as shown in Figure 6. FL is proposed as a promising approach to COVID-19 detection. Qayyum et al. [58] implement the surging concept of clustered federated learning (CFL) for an automatic diagnosis of COVID-19 using X-ray and Ultrasound datasets.
FL allowed medical professionals to assess the intensity of skin disease. Hashmani et al. [59] proposed an architecture that can diagnose the type of skin disease by conducting experiments using dermoscopy images to test and validate the model's classifcation accuracy and adaptability.
FL is more potential in analyzing the medical images and also in protecting the patient's privacy. Lee et al. [60] used ultrasound image analysis and FL to determine whether thyroid nodules were normal or dangerous.
FL enables efcient and accurate heart disease diagnosis. Linardos et al. [61] present the frst FL study for the diagnosis of hypertrophic cardiomyopathy (HCM) using the subsets derived from M and M and ACDC datasets.
FL is a privacy-preserving AI model to identify brain tumors. Li et al. [56] focus on practical FL systems for brain tumor segmentation by using the BraTS dataset.
A summary of the key fndings from the above discussion can be found in Table 3.

FL in IoT-Based Smart Healthcare
Applications. FL is a concept developed to construct smart and confdential IoT devices. FL has developed a system that makes it possible for doctors to use data from multiple medical institutions without sharing patient data across them [62]. Hospitals can train their own AI models by uploading the data to a global aggregator. FL brings together multiple hospitals to develop collaborative healthcare settings that speed up diagnostic testing of victims while maintaining their personal rights. Te healthcare sector is one of the most attractive applications of IoT, as depicted in Figure 7.
Te use of AI-based approaches to learn health data has been adopted widely in smart healthcare, such as the use of intelligent imaging to detect disease [63]. Medical information requires a degree of insensitivity and is regulated by medical rules like the Health Insurance Portability and Accountability Act (HIPAA) [64]. Ensuring data protection from public sources shared with its server or storage system is an issue [65]. Traditional AI systems relying on some kind of master database software are not really suitable for contemporary health.
FL's utility in the smart healthcare sector with sophisticated features has been proved in recent research. Te role of electronic health records in the healthcare industry is growing rapidly. Here, we focus on two business scenarios, namely, EHR maintenance and involvement in health coverage.

Managing Electronic Health Records Using FL.
Te application of FL in healthcare operations has really been studied in order to provide adaptable as well as confdential EHR maintenance. An FL-based student-centered architecture is responsible for an EHR system with diferent clinics as well as a database server [66]. Every company uses its own electronic health records and a cloud server to manage a neural network (NN). A new method of data destruction has been developed by Google. Te concept is intended to protect the learning experience in data storage from cyberattacks. It uses a lightweight data destruction technique to disrupt training data and to ensure confdentiality of the FL model parameters that can protect them. Although an attacker can get troubled EHR information, the source data is difcult to gather or retrieve. Te AlexNet NN simulation was performed with the CIFAR-10 standard dataset to achieve accurate and safe results for the study of EHR.
Liu et al. [67] built a distributed NN activity plan which enables any clinic to know and understand a part of the model from its EHR source. Te integration of autonomous position monitoring and conventional nonconvex enhancement principles [68] builds a diferent innovative method for FL. Te aim is to save network resources by interacting with the remote EHR server. FL allows the sharing of EHRs without their sharing due to its dynamic approach to education. New features to the model sufer anonymity, which leads to communication path inference attacks. To tackle this problem, diferential privacy strategies can help enhance the protection of personal data for the learning of FL-based EHRs [69].
EHR information is used for detecting FL's adverse drug reactions (ADRs). It can be used to detect rare ADRs at a    Computational Intelligence and Neuroscience EHR user, to increase the quality and precision of integration. FL-based healthcare imaging architectures protect patient privacy as a key feature [72]. In this context, clinics and healthcare physicians cooperate together over the development of a safe and multiparty FL system for service users with medical algorithms.

Health Cooperation FL.
In order to contribute, FL can ensure proper healthcare coordination for better provision of medical facilities through its centralized and secure nature. Yuan et al. [73] present a cooperative framework for healthcare that leverages FL to enable medical IoT devices. Te next generation of FL solution is presented in [74] cloud edge-based healthcare. Tis solution could be used to address issues that directly afect the FL process, such as device, data, and model. Personalization training on edge computers was decided to be carried out, in this case, to reduce heterogeneity and achieve high-quality individual models. For the purpose of stimulating the federation of portable devices using FL technology, the FedHealth framework [75] was developed. Information collected by a number of healthcare institutions using diferent IoT devices may be collected using Fed-Health's surgical instruments to enable the development of a powerful AI method that can be used to identify people's behavior and to protect ciphertext data [76]. An innovative method [77] is the use of chaindirected synchronous stochastic gradient descent to minimize human latency between FL clients and servers in personal mobile sensing applications. Recent studies reveal that FL is helpful in preventing infections like COVID-19 [78].
A summary of the key fndings from the above discussion can be found in Table 4.

FL for COVID Outbreak Prediction
. FL utilizes multiple devices or servers that store local samples of data, without aggregating clinical data to create a statistical model, which is undesirable for several reasons, including patient privacy concerns. COVID-19 infection has been confrmed in China since December 2019. After that, the outbreak began spreading to China and several other countries across the globe [79]. A pandemic with rapid growth (and thousands of infections and hundreds of deaths) presents considerable obstacles to control the virus.
Tere are currently a number of diferent ways to detect COVID-19, whereas computed tomography and X-rays remain the major surgical techniques [80][81][82]. A number of clinical symptoms are observed in COVID-19 patients hospitalized with Acute Kidney Injury (AKI) [83][84][85][86]. According to studies, the incidence of AKI ranges between 46 percent and 71 percent; however, most deaths occur within the AKI sample [87,88]. Diagnostic methods that judiciously help individuals at greater risk of contracting COVID-19 can be benefcial during an outbreak where facilities can be restricted [89].
Despite FL's promise, it is still a relatively new idea for physicians, patients, payers, researchers, and hospitals. FL is strongly encouraged to examine whether and how it can provide valuable support during and after the unprecedented COVID-19 pandemic to control it.
A summary of the key fndings from the above discussion can be found in Table 5.
In COVID-19, FL demonstrates its value in diagnosis, treatment, and prognosis prediction. Xu et al. [19] demonstrated that with the help of chest computed tomography (CT) scan data from several institutions in Wuhan, China, they were able to overcome data scarcity, isolation, and heterogeneity and achieve improved detection sensitivity. Based on their fndings, an FL-based architecture would allow global participants to beneft from a globally distributed and real-time CT-COVID-19 diagnostic tool. A study by Vaid et al. [92] published recently found that FLbased models tended to perform better than locally trained predictive models when using data from medical centers to predict 7-day mortality in hospitalized patients. Te intent of this study is to evaluate FL in predicting a meaningful outcome for hospitalized COVID-19 patients.
A confrmatory fnding of FL's superiority is likely to spark signifcant interest, especially in its potential to improve outcomes for COVID-19 patients. FL has been able to harness all of the full learning power of existing data to ofer data-driven insights and personalized recommendations due to the uncertainty of long-term complications of COVID-19, the efectiveness of medical treatments, the safety of vaccines, and immunity protection.
COVID-19 care is only a small part of FL's value as a health provider. Tis pandemic has severely disrupted non-COVID-19 multicenter clinical trials. It has become increasingly challenging to utilize data generated from each participating institution. In recent years, it has been said that decentralizing clinical research would beneft traditionally underrepresented subgroups and underserved areas in particular. Tis kind of decentralized research could beneft greatly from integrating FL into the study design and data analysis to evaluate the quality of care and outcomes, for example, predicting mortality, complications, hospitalizations, and adverse drug reactions. Another area for FL to explore is digital health. Healthcare, precision medicine, wearable technology, and clinical decision support have all seen a rise in popularity as a result of the epidemic. Te suitable use of FL produces generalizable models that will help the achievement of equitable, efective, and patientcentered care.
FL can efectively address the issue of data unavailability and get a shared model without obtaining local data. Boyi Liu et al. [90] proposed an experiment in which four popular federated ML models (Mobile Net, ResNet18, MobileNet-v2, and COVID-Net) were applied to CXR images of patient's chests to compare their performance. Te authors developed these models to detect COVID-19 pneumonia, using the same parameters for all models.
FL proved to be more efcient compared to traditional ML models in COVID detection. Pang et al. [91] created a federated model based on a digital city twin concept to predict outcomes of diferent COVID-19 prevention plans over time and assess the survival rates of multiple cities over the study period. Furthermore, using the digital city twin 10 Computational Intelligence and Neuroscience platform, they were able to track the efectiveness of each prevention plan and create local models which were sent to the federated sites for safety.

Lessons Learned
In this section, we summarize the key lessons learned from this survey, which thus provide an overall picture of the current research of FL applications for medical image processing, FL toward privacy and security in healthcare applications, FL in IoT-based smart healthcare applications, and FL for outbreak prediction.

FL Applications for Medical Image
Processing. Based on the observations from several states of the art, the FL algorithm analyzes electronic health records to describe and estimate hospitalization, survival, and death rates for people associated with physicians. In addition to identifying disease risk factors using the fMRI classifcation technique, a promising COVID-19 approach was presented as well.

FL in IoT-Based Smart Healthcare
Applications. FL has developed as a distributed collaborative AI approach that has the ability to enable a wide range of intelligent IoT applications by allowing AI training at distributed IoT devices without the need for data exchange [93]. Based on the observations from several states of the art, FL developed a system that allows doctors to access patient data across multiple hospitals without constantly transferring patient information. FL provides smart healthcare solutions and reshapes current healthcare systems by improving privacy for users and reducing latency among healthcare providers and users alike. A lesson to be learned from this is that the FL can enable healthcare operations to manage EHRs centrally and with privacy preserved, by collaborating with multiple medical institutions to build intelligent EHR systems. Further, few fully decentralized FL approaches are able to provide decentralized optimization and stochastic gradient tracking by combining the cooperation of hospitals with a decentralized stochastic gradient algorithm to improve convergence rates.

FL for COVID Outbreak Prediction.
Based on the observations from several states of the art, in an outbreak when facilities may be restricted, it is important to develop diagnostic methods that will help individuals at greater risk of contracting COVID-19 judiciously. We also fnd that using FL in decentralized research would beneft researchers by improving study design and analyzing data on outcomes and quality of care, including predicting mortality, complications, and adverse drug reactions. Goldfarb et al. [89] consolidate existing case studies and identify the future challenges in defning FL's benefts and impact on healthcare applications, as well as the obstructions and concerns surrounding its adoption in healthcare.

Challenges and Future Directions
FL does have some advantages, but it also poses challenges that must be considered when setting up federated training eforts.

Confdentiality and Safety.
Tere is a lot of important data in the healthcare industry which needs to be defended. FL poses many privacy-preserving challenges in terms of alternatives, techniques, and consequences.
6.1.1. Privacy versus Performance. FL is primarily designed to safeguard confdentiality by discussing method updates instead of information, although it does not solve many security issues but, like any ML algorithm, always carries some risk. FL ofers a level of privacy preservation that is superior to the current commercially available ML models in terms of protection [94]. However, these methods have a tradeof when it comes to performance, which may cause the fnal model to be inaccurate [48]. Furthermore, future techniques and additional information may compromise a previously low-risk model.

Level of Trust. FL collaboration can be classifed into two broad types:
Trusted: FL consortia, whose partnership is binding and regarded as trustworthy, typically prevent a few serious reasons, like a plot to steal crucial data or damage their system. To reduce the necessity of specialized countermeasures, researchers can reexamine the fundamentals of regular cooperation studies [95]. Nontrusted: FL systems that are used in a wide range of settings could prove difcult to develop a collaborative partnership that guarantees timely, appropriate collaboration between all parties. In some cases, users may attempt to misuse the system, undermine its performance, or obtain information from others intentionally. In order to reduce these threats, security strategies including cryptography of product proposals, efective authorization for participants, tracking of operations, It is challenging to share data securely (without compromising the privacy of users) and to train global models for -detecting positive cases [20] UCADI framework A decentralized model, the unifed CT-COVID AI diagnostic initiative, distributes and performs the AI model at each participating institution independently without sharing personal data (i) Data defciency (ii) Data isolation (iii) Data heterogeneity [38] AI and big data Te coronavirus disease COVID-19 is being controlled with the use of AI and big data Privacy and security issues due to insufcient standard datasets [ Te lack of unifed databases is a concern for protecting the privacy and the security of blockchain [90] AI-related technologies A comparison of FL to training without an FL framework was conducted using four diferent models: (i) MobileNet FL presents a number of statistical and system challenges when distributed device networks are used to train machine models (ii) ResNet18 (iii) MoblieNet, and (iv) COVID-net [91] A novel collaborative city DT framework FL combined with city DTs alleviates the data sparsity challenge, facilitates collaboration, and provides privacy protection by design Collaborative training problems, such as: (i) Disaster surveillance and (ii) Prediction

12
Computational Intelligence and Neuroscience asymmetric security, conducting testing, performance reliability, confdentiality of design, and protection from opponents are necessary [96].

Data Breach.
Patients' medical information is usually not shared across FL devices and hospitals. Nevertheless, the observations can indeed efectively leak personal information required for regional development, such as reversal [97] model updates, gradients [98], and opponents. Traditional training difers from FL in that multiple parties are exposed to the training process. Terefore, monitoring alterations in a model over time, monitoring exact model upgrades, or manipulating a model can occur as a consequence of a dynamic analysis breach.

Data
Heterogeneity. In addition, caused by things like data gathering methods, company biomedical applications, and area demographics, medical data difers in types, sizes, and attributes. In light of this, algorithms and strategies apply to FL as an important factor: many traditional techniques distribute relevant information independently and in the same way to all users (IID). Given the fact that there can be no comprehensive information distribution among institutions [56,62], FL can be trained using non-IID data. In such conditions, techniques such as FedAvg [47] frequently fail, which violates the fundamental aim of active learning. Researchers have been studying this issue by utilizing FedProx [99] and part-sharing strategies [100]. Furthermore, heterogeneity of data may lead to a situation in which local solutions are not optimal for global solutions. It is therefore crucial that all participants agree on the defnition of model training optimality before the training begins.
6.3. Traceability and Accountability. FL in healthcare is a system that needs to have reproducibility as it is a safetycritical application. Running multiparty computations in complex hardware, software, and networking environments is diferent from training with centralized data. In order to fulfll the traceability requirement, it must be possible to trace network connections, connectivity records, and confguration modifcations, including hyperparameter tuning, throughout the training process. Additionally, traceability can be logged information about a model's training history, particularly to ensure training and test datasets are not overlapping. Execution integrity is fundamental to traceability and accountability in nontrusted federations [31].
Using FL, researchers cannot view pictures of designs that are generated. Te individual sites still have their unique original information available, but organizations can choose to give any central viewing facility to meet necessity or indeed make it possible to explain and interpret the optimization method.

System Architecture.
Computing resources and networks are typically better suited to large-scale FL in healthcare institutions than in consumer devices [31]. Tese methods make it possible to teach a greater number of products, and also the data collected with these methods can be shared more widely. As a result, FL brings both opportunities and challenges to healthcare, such as the following: (i) It is crucial to examine when discussing whether safety prevention is achievable (e.g., by providing unneeded packages). (ii) How should encryption methods be designed to utilize computational resources efectively? (iii) Which methods can be used to reduce idle time and to take advantage of independent systems in order to develop suitable network confgurations?
6.5. Client Management. Client management is an essential issue in FL, in contrast to the centralized ML architecture. Some clients may be looking for benefts without contributing; it is the responsibility of the server to decide which customers should participate in the learning process. Client management involves helping a patient or client develop a plan that coordinates and integrates essential support services for the most optimal results and outcomes [101]. Tere are many components of client management. Hudon et al. summarize several descriptions of case (client) management, including those of the Case Management Society of America and the Canadian National Case Management Network, and describe six core elements that include patient identifcation and eligibility determination, assessment, care planning along with goal setting, plan implementation, plan monitoring, and transition and discharge [102].
6.6. Health Dataset Issues. Te healthcare sector has the capability to deal with a wide range of data types and content, including text, images, audio, and time series, as well as blood types, heart rates, facial images, and body temperatures. Te majority of FL approaches are typically examined on a single dataset with a limited number of features. However, despite the fact that both are proposed for privacy-preserving FL-based healthcare services, the work in [103] is tested on a dataset related to diabetic retinopathy, and the work in [69] is evaluated on an EHR dataset. Te central server can navigate heterogeneity by using private ensemble learning [104] to navigate heterogeneous FL approaches involving multiple parties having diferent models. An inference strategy is presented to enable participants to use an ensemble of heterogeneous models without needing to explicitly join the data in a single place.
A summary of the key fndings from the above discussion can be found in Table 6.
6.6.1. Future Directions. In the future, healthcare tasks may be performed better using collaborative strategies across multiple institutions, as opposed to very large fles limited to a specifc health center. FL can integrate information acquired and maintained by diferent institutions in order to capture broader data variability and analyze patients from Computational Intelligence and Neuroscience 13 diverse demographics. Furthermore, FL allows the addition of multiexpert annotation and data from multiple sites acquired with diferent instruments and techniques. In order to achieve this collaborative goal, diferent agreements must be put in place, including details about scope, purpose, and technology, which may not be known at present as the feld is a new one.
6.6.2. FL for Healthcare 5.0. In the traditional ML model, the data used in training the models raised issues regarding security and privacy. In this context, FL has been found to enhance scalability, improve accuracy, reduce training time, improve performance, increase privacy and enhance safety, among others [105]. FL has some advantages over the traditional methods, which are as follows: (i) An FL network makes the entire network scalable by enabling diferent devices to learn from one another. (ii) Te creation of local models reduces latency and lowers the power consumption when compared to training a single central model. (iii) As a result of their use of many local models and simultaneous perspective approaches, FL models are more innovative than centrally trained models. (iv) FL provides predicted values in real time because you can access the datasets without a centralized server. It reduces the lag time in data access and enables you to access data without connecting to the central server. Data can be transmitted and received directly through a local server.
(v) Compared with standard central ML algorithms requiring datasets on a single server, FL reduces security and confdentiality concerns by managing local database objects.

FL for Digital Twin in Healthcare.
In the future, healthcare will become automated, citizens will adopt a more proactive approach to their health (empowerment, prevention), and clinical decision support will be integrated throughout practice. Te combination of these measures will allow the healthcare system to be sustainable in the future, even when the population grows older. Several applications of digital twin technology are found in the healthcare sector. In healthcare, technology is enabling advances that, once thought impossible, are becoming possible thanks to its growth and development. Te health sector uses many applications that do not directly beneft the patients but have a benefcial infuence on how they are treated. Terefore, these systems play a vital role in improving patient care.
Healthcare is even more dependent on simulation and actuations in real time as it can mean the diference between life and death. In addition to predictive maintenance and ongoing equipment repair, the digital twin can also assist with the diagnosis and prevention of problems. Medical digital twins can make life-saving decisions on the basis of real-time and historical data with the help of AI [106,107].
Regulations are one step to ensuring personal data is protected, but FL is another way to build decentralized training models. Te privacy and security issues associated with data analysis within a digital twin are addressed by An inference strategy is presented to enable participants to use an ensemble of heterogeneous models without needing to explicitly join the data in a single place keeping users' data localized within an FL model, which provides the desired outcomes in terms of implementing data analytics [108].

FL and Blockchain for
Healthcare. Te FL-based technique would be able to help the healthcare providers because it can improve the accuracy and robustness of the AI model and that helps to make the model more generalizable so that it could be used in the real-time environment [109]. At the same time, this approach would be able to save time and cost. Te most important part is that we would be able to get all the benefts from FL without leakage of information.
Te software and hardware manufacturers would get a lot of benefts from the FL because collaborative learning is possible between hardware devices and software applications without information leakage and the other beneft will come in terms of validation of the AI-based model since it is combined with the hardware devices that helps in continuous improvement of AI-based models [110].
Researchers and developers get a lot of benefts from the FL-based services because a huge amount of real data would be available to them [111]. Tis will help the AI researchers and developers to think about diferent algorithm strategies that would help them to make robust models.

FL for Collaborative Robotics in Healthcare.
Collaborative robotic (Cobot) technology has been widely adopted both by healthcare professionals and by those involved in the medical device industry as a tool to improve workforce efciency, facilitate workfow improvements, and streamline safety procedures [112].
Healthcare collaborative robots are automated systems deployed in the medical industry to perform various tasks, starting from administrative tasks, lab testing, patient care, and surgical aids. Tese cobots require an hour to fll the gap between the ongoing medical industry burden and staf shortage. Te rising requirement to facilitate automation in the healthcare industry to reduce infection exposure to front-line workers along with increased technological advancement in inpatient care services has infuenced the healthcare collaborative robots market growth.
Te cobots are proved to be extremely efcient while performing the lab testing tasks. High precision, fast turnaround time, and reduced dependency on manual tasks will positively infuence the lab testing applications. Te other key promising federated application is patient care which includes medicine dispensing, taking swab samples, checking temperature and blood pressure, and conducting various sample tests which have made it easier for health workers to reduce their burden and utilize more time on urgent matters. 6.6.6. FL and Explainable AI for Healthcare. IoT has transformed the healthcare domain by introducing the Internet of Medical Tings (IoMT); however, choosing analysis for distributed IoMT environment, analysis of the enormous amount of data generated by IoMT devices in a distributed environment, and achieving security of IoMT devices is a challenging task. However, recent researches on AI-enabled remote health monitoring systems were able to monitor and prevent cyberattacks. Explainable AI is a promising modern technology for identifying the compromised data during cyberattacks in IoMT-based patient monitoring systems by enabling caregivers to fx the problems. Also, the FL-based Wearable, explainable AI frameworks will enable the user to have better communication using knowledge-based methods and also improve user acceptance and task performance [113].
6.6.7. FL for Integration with 6G in Healthcare. Te enormous usage of IoMT devices in our daily activities results in an explosive growth of data trafc, ML, and data-driven approaches. Moreover, the surge in the development of communication technology led to the in-novation of 6G networks, by transforming wireless communication from "connected things" to "connected intelligence," expecting to embody advanced AI various applications promising graterlevel of security and stronger privacy protections in the healthcare domain. A large number of IoMT devices with massive data in the 6G era will force individuals to deploy efcient ML and AI-based algorithms to provide highquality services. However, implementing FL-based Edge intelligence in 6G will bring in improved performance, ultralow latency service, and enhanced privacy of the system [114].

Conclusion
FL approach is a promising method for achieving strong, precise, safe, robust, and unbiased modeling results. FL facilitates communication between various parties without the need to communicate or centralize data, thereby solving difculties associated with critical health information. In addition, this could lead to new business opportunities and improved healthcare worldwide. In this paper, the authors addressed the health sector's opportunities and drawbacks for FL. Te FL investigation is expected to continue for a further decade since not every fundamental barrier has been removed recently. Despite this, we think precision medicine will have a great deal of impact on medical care in the future.

Data Availability
No data were used to support the fndings of the study.

Conflicts of Interest
Te authors declare that they have no conficts of interest to disclose.