Federated Learning for Healthcare- Systematic Review and Architecture Proposal

Federated Learning for Healthcare: Systematic Review and
Architecture Proposal
RODOLFO STOFFEL ANTUNES and CRISTIANO ANDRÉ DA COSTA, Universidade do Vale

do Rio dos Sinos, São Leopoldo, Brazil
ARNE KÜDERLE, IMRANA ABDULLAHI YARI, and BJÖRN ESKOFIER,
Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany
The use of machine learning (ML) with electronic health records (EHR) is growing in popularity as
a means to extract knowledge that can improve the decision-making process in healthcare. Such methods
require training of high-quality learning models based on diverse and comprehensive datasets, which are hard 54
to obtain due to the sensitive nature of medical data from patients. In this context, federated learning (FL) is
a methodology that enables the distributed training of machine learning models with remotely hosted datasets
without the need to accumulate data and, therefore, compromise it. FL is a promising solution to improve
ML-based systems, better aligning them to regulatory requirements, improving trustworthiness and data
sovereignty. However, many open questions must be addressed before the use of FL becomes widespread. This
article aims at presenting a systematic literature review on current research about FL in the context of EHR
data for healthcare applications. Our analysis highlights the main research topics, proposed solutions, case
studies, and respective ML methods. Furthermore, the article discusses a general architecture for FL applied to
healthcare data based on the main insights obtained from the literature review. The collected literature corpus
indicates that there is extensive research on the privacy and confidentiality aspects of training data and model
sharing, which is expected given the sensitive nature of medical data. Studies also explore improvements to
the aggregation mechanisms required to generate the learning model from distributed contributions and case
studies with different types of medical data.
CCS Concepts: • Applied computing → Health informatics; • Computer systems organization →

Distributed architectures; • Computing methodologies → Machine learning approaches;
Additional Key Words and Phrases: Electronic health records, federated learning, systematic review
ACM Reference format:

Rodolfo Stoffel Antunes, Cristiano André da Costa, Arne Küderle, Imrana Abdullahi Yari, and Björn Eskofier.
2022. Federated Learning for Healthcare: Systematic Review and Architecture Proposal. ACM Trans. Intell.
Syst. Technol. 13, 4, Article 54 (May 2022), 23 pages.
https://doi.org/10.1145/3501813
This work is supported by: Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) financial code 0001;
Fundação de Amparo à Pesquisa do Estado do Rio Grande do Sul (FAPERGS) ARD 04/2019 grant 19/2551-0001340-0; and
Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) grant 309537/2020-7.
Authors’ addresses: R. S. Antunes and C. A. da Costa, Universidade do Vale do Rio dos Sinos, Av. Unisinos 950,
CEP 93022-750, São Leopoldo, RS, Brazil; emails: {rsantunes, cac}@unisinos.br; A. Küderle, I. A. Yari, and B. Eskofier,
Friedrich-Alexander-Universität, Carl-Thiersch-Straße 2b, 91052 Erlangen, Bayern, Germany; emails: {arne.kuederle,
imrana.yari.abdullahi, bjoern.eskofier}@fau.de.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires
prior specific permission and/or a fee. Request permissions from [email protected].
© 2022 Association for Computing Machinery.
2157-6904/2022/05-ART54 $15.00
https://doi.org/10.1145/3501813
ACM Transactions on Intelligent Systems and Technology, Vol. 13, No. 4, Article 54. Publication date: May 2022.
54:2 R. S. Antunes et al.
1 INTRODUCTION
The pervasiveness of computational systems in healthcare environments led to the widespread
availability of electronic health data [1]. These datasets enable the use of advanced data analysis
techniques based on machine learning (ML) to extract vast amounts of knowledge that can
improve various aspects of healthcare [2]. Ultimately, the analysis of electronic health data using
ML techniques can result in treatments and procedures with lower risks and better outcomes for
patients, thus increasing the quality of care [3].
Most ML methods applied to healthcare focus on the analysis of Electronic Health Records
(EHR) data, which contains the individual medical information from a person, possibly including
appointments, diagnoses, treatments, and exams such as medical imaging [4]. EHR emerged from
the natural adoption of information technology by medical institutions and became a relevant
source of data for many analytical methods aimed at extracting knowledge to aid medical practices
[5]. In particular, recent work shows the potential of deep learning methods applied to EHR data
to extract knowledge that can directly improve the outcomes and quality of care [6].
ML-based methods require a learning model trained to identify patterns and extract the intended
knowledge from raw data. This training step is crucial to the accuracy of the resulting analysis
method. It must be conducted with a large dataset with a diverse number of samples to guarantee
the ML model quality [7]. The training’s results directly depend on the number of data items, the
diversity of the samples, and the quality of dataset annotation regarding the expected classifica-
tion. Acquiring or producing such a dataset often is a time-consuming and expensive task. This
procedure involves several parties collecting the data, transferring it to a central data repository,
and fusing it to build a model. In turn, data owners may be unaware of these procedures and the
future use cases of the ML model. For that reason, in the particular case of healthcare, the tradi-
tional ML processes may violate laws such as the General Data Protection Regulation (GDPR)
of the European Union, the California Consumer Privacy Act (CCPA), and Health Insurance
Portability and Accountability Act (HIPAA) [8]. Such difficulties and restrictions on medical
data sharing are the major hindering factor in the development of advanced ML techniques for
healthcare [9].
Recent work proposes the concept of federated learning (FL) [10] to tackle these issues. In
summary, FL enables the training of ML models locally (at the location of the data) and only shares
the resulting model, which is not reverse-engineerable, with the requesting party. Therefore, FL
avoids the need to share the private datasets and sensitive data to others, preventing exposition
to entities conducting studies and enabling data usage for broader purposes [11]. A central entity
manages the learning process and distributes the training algorithm to each participating data
holder. Each participant generates a local model trained with their private data and shares the
resulting parameters with the central entity. Finally, the central entity employs an aggregation
algorithm to combine the parameters of all local models into a single global model.
Applications described in the literature successfully employ the methodology described above.
For example, Hard et al. [12] train the prediction model of a mobile keyboard using private data
stored in smartphones. Similar studies demonstrated the viability and proposed solutions enabling
FL in a wide variety of contexts. In particular, FL provides a solution to train high-quality models
for ML applied to healthcare [13]. As a result, a group of studies aims to evaluate the state-of-the-
art on this subject. Xu et al. [14] briefly discuss the applications of FL on healthcare applications.
Rieke et al. [15] explore various technical and social aspects related to FL in healthcare applied to
healthcare. Zerka et al. [16] survey the FL methods with the specific focus of their advantages and
drawbacks to the privacy and security of medical data. Li et al. [17] present a general overview
of FL applications but does not provide an in-depth analysis of challenges related to medical data.
Federated Learning for Healthcare 54:3
Mothukuri et al. [18] focus on the security aspects of FL architectures and strategies to mitigate
threats. Finally, Zhang et al. [19] survey FL research from the perspective of architectural chal-
lenges and future directions. These studies provide a general overview of FL, its application to
healthcare, and the implications regarding technical and social perspectives. However, none pro-
vide an in-depth exploration of current literature in light of FL applications for healthcare, classi-
fying it according to the employed methodologies and technical challenges. Furthermore, none of
the studies leverages the researched challenges and solutions to propose a general architecture to
enable FL for studies that require datasets owned by healthcare institutions.
This article aims at providing a comprehensive review of current studies related to the applica-
tion of FL to the analysis of EHR data with ML-enabled methods. The study follows a systematic
literature review methodology [20] with a well-defined search methodology to select a corpus of
44 recent articles for analysis. In summary, our review provides the following contributions:
• A comprehensive literature review of current work on FL applied to EHR data, including
main research issues and proposed solutions, employed ML techniques, and case studies.
• A discussion about a general architecture to facilitate the use of FL with ML-enabled appli-
cations for healthcare, including the concept of federated data analysis.
The remainder of this paper is organized as follows: Section 2 describes the methodology used
to search and select the articles that comprise the literature corpus. Section 3 presents the results
of our literature analysis, focusing on answering a set of well-defined research questions. Section 4
overviews an architecture to facilitate the use of FL and analysis with distributed EHR data. Finally,
Section 5 provides some final remarks and future perspectives related to our work.
2 METHODOLOGY
The methodology adopted in this survey follows the principles of a systematic review [20]. It
comprises a set of well-defined and documented steps that favor the reproducibility of obtained
results. In summary, these steps are: (i) definition of research questions; (ii) selection of keyword
and literature databases; (iii) article filtering and selection; and (iv) result analysis and discussion.
The remainder of this section describes in more detail the adopted methodology.
The first step is to define the research questions to answer with the literature review. This work
comprises one main question (MQ) and four specific questions (SQ), as listed below:
MQ What are the current applications of FL in the context of EHR analysis?
SQ1 What are the main research questions and respective solutions discussed in the literature?
SQ2 How do we classify the current work on FL applied to EHR?
SQ3 What are issues that remain as open research questions on FL applied to EHR?
SQ4 What is the design of an FL architecture that leverages EHR data?
Using the research questions, we can define the keywords to extract the raw corpus from lit-
erature databases. The keywords comprise the most relevant terms in the main research ques-
tion, including common synonyms. Regarding the literature databases, we employ those most
relevant to the fields of computer science and medicine. Table 1 lists the databases used in this re-
search and the employed keywords. These parameters resulted in an initial raw literature corpus of
200 articles.
The next phase of the methodology is to analyze and filter the raw corpus, aiming to obtain the
final corpus for the proposed literature review. The filtering process comprises a series of steps
that aim to eliminate results that do not add to the quality of the resulting literature review. It
comprises the following criteria:
Table 1. Databases and Keywords Used to Extract the

Raw Literature Corpus
Literature Databases
Name Address
ACM Digital Library https://dl.acm.org
IEEE Xplore https://plore.ieee.org
SpringerLink https://link.springer.com
ScienceDirect https://www.sciencedirect.com
PubMed https://www.ncbi.nlm.nih.gov/pubmed
arXiv https://arxiv.org
JMIR https://jmirpublications.com
Keywords
“federated learning” ∧ (“electronic health records” ∨
“electronic medical records” ∨ “EHR” ∨ “EMR”)
Fig. 1. Distribution of articles by publication year.
(1) Publication date: the corpus should include articles published in the last five years (2016
to 2021) to limit the review to the most recent work in the area.
(2) Title and abstract review: provides an initial screening to remove those not directly related
to FL applied to EHR.
(3) Duplicate removal: combines articles from different databases and remove any existing
duplicates.
(4) Quality assessment: remove articles that do not contain all elements expected in a com-
plete research work.
The application of the filtering criteria results in a set of 67 articles. These comprise the final
literature corpus used in this review. An initial analysis of the articles shows that FL applied to
EHR is a recent research area. Figure 1 illustrates this aspect with the distribution of articles by
publication year. In particular, we observe that the final corpus contains 47% of articles published
in 2020 and 12% in the first quarter of 2021.
The final step of the systematic review methodology comprises the analysis and discussion of
the literary corpus. The next section presents the results from this step, considering the research
questions proposed for this review.
3 LITERATURE ANALYSIS
This section presents the analysis of the literature corpus collected with the described method-
ology. First, Section 3.1 explores the evolution of studies on federated learning applied to health-
care, including the main research issues and proposed solutions. Next, Section 3.2 summarizes the
Fig. 2. Distribution by topics of articles in the literature corpus.
observed ML methods and case studies observed in the literature. Finally, Section 3.3 presents a
taxonomy to classify the observed literature and discusses the main insights obtained from our
analysis.
3.1 State-of-the-Art Evolution

The literature corpus shows that FL applied to EHR data analysis is a very recent research field,
with most publications from the last two years. Earlier publications focus on architecture defini-
tions and experimental case studies to evaluate the accuracy of learning models trained with FL
concepts. Some earlier studies also focus on evaluating the privacy guarantees of the FL process in
light of the stringent security requirements from medical data analysis. In turn, recent work added
discussions on specific challenges that hinder the accuracy and confidentiality of learning models
generated with FL. In particular, they explore the aggregation and cryptographic mechanisms em-
ployed in FL architectures.
Our analysis classifies the literature corpus in five general topics: (i) case studies of FL on
medical data; (ii) architectures for FL in medical applications; (iii) confidentiality guarantees
for training data; (iv) aggregation methodologies for global model generation; and (v) federated
data analysis. Figure 2 illustrates the distribution by topic of the articles in the literature corpus.
The remainder of this section explores the studies that focus on each of these topics.
3.1.1 Case Studies. This group encompasses some of the first work on FL applied to EHR data.
These studies aim to apply the concepts presented by Yang et al. [21] to explore the viability of
FL as a tool for training learning models over private medical data, including EHR. In this group
of early studies, Jochems et al. [22] employ distributed learning to train a Bayesian network and
evaluate the long-term evolution of lung cancer patients. Kim et al. [23] propose a federated tensor
factorization method that iteratively trains and aggregates a learning model to extract phenotypes
from EHR datasets. Sheller et al. [24] explore a distributed architecture to train a deep neural
network to segment MRI scans from patients with brain tumors. Liu et al. [25] develop a distributed
two-layer neural network for mortality prediction based on EHR data, in which the bottom layer
is private to each participant and the top layer is public and trained using the aggregation of
anonymous parameters from local models.
Regarding more recent studies, Liu et al. [26] employ a deep averaging network to extract
medical datasets from plain text health records. The result is used in a distributed architecture
based on support vector machines to train models to predict diagnoses based on EHR data. Silva
et al. [27] explore particular mathematical characteristics of learning models for neurological MRI
data analysis and propose a federated training architecture specific for this scenario. Their goal
is to train high-accuracy models for neurological disease diagnosis. Jiang et al. [28] propose an
FL architecture focused on models for topic analysis on distributed datasets. This type of analysis
focuses on identifying the general subjects from a given text and can be applied to analyze medi-
cal data in textual format. Sharma et al. [29] compares a learning model trained with a federated
architecture against one with a centralized method. Their results indicate that FL can achieve an
accuracy level similar to that of centralized methods, and the additional privacy is advantageous
to healthcare applications. Ribero et al. [30] employ k-means clustering to learn prototypes from
local dataset parameters that can be shared without compromising data confidentiality. These pro-
totypes can be aggregated into a global learning model that drives recommendation systems on
different applications, including medical data. Deist et al. [31] evaluate a case study of FL for gen-
erating regression models for two-year survival prediction of lung cancer patients. The case study
explores an architecture deployed at five institutions that holds a total of 20,000 patient records.
Szatmari et al. [32] propose a standard FL architecture to train learning models for the person-
alization of hearing aids. Sheller et al. [33] apply FL to train a model to identify cancer-affected
brain tissue using as a case study the BraTS 2017 dataset. Vaid et al. [34] evaluate the results from
traditional and FL training using as case study COVID-19 mortality prediction using EHR data
including demographics, medical history, and vital signs. Lee et al. [35] explore the case study of
ultrasound image analysis for tumor identification to compare FL and traditional training architec-
tures. Liu et al. [36] employ FL to predict the mental state of patients based on physiological and
motion data collected with medical IoT devices.
Another group of recent studies focuses on FL applied to data analysis from resource-
constrained Internet of health things (IoHT) [37] devices. Guo et al. [38] propose a solution
based on edge computing to enable FL in scenarios with resource-constrained devices. Chen et al.
[39] explore transfer learning to improve the accuracy of global models trained to analyze IoHT
data. Ju et al. [40] explore transfer learning to improve the accuracy of models to analyze EEG
(electroencephalogram) data to enable brain-computer interfaces. Zhang et al. [41] propose an
FL tailored to train models and analyze ECG (electrocardiogram) data to diagnose arrhythmia
using IoHT devices. Can and Ersoy [42] also explore FL for ECG data analysis but focuses on de-
tecting stress. Lim et al. [43] explore an incentive mechanism to guarantee that FL peers remain
active in long training sessions for learning models for IoHT data.
The described work focuses on employing FL mechanisms to different ML techniques for med-
ical data analysis. While deep neural networks stand out as a common technique for multiple
studies, there is interest in exploring other techniques, for example, Bayesian networks and k-
means clustering, which are well known for medical applications. Furthermore, these studies do
not present major proposals for specific FL challenges, such as model aggregation or data confiden-
tiality and integrity. Their main goal is to compare the precision of a model trained in a distributed
architecture against one that employs traditional, centralized training. Furthermore, some qualita-
tively explore the trade-off between the loss of precision that stems from distributed training and
the inherent advantages of FL (e.g., access to more varied datasets).
In general, these studies show that learning models generated with FL provide equal or better
accuracy than those trained with centralized methodologies. These results stem from the broader
information available from the datasets distributed in multiple institutions, which would not be
available in circumstances where patient anonymity cannot be guaranteed. Studies also indicate
that FL provides similar benefits to distributed ML, such as load distribution for model training.
However, studies also show that the characteristics of individual datasets can directly impact train-
ing results, mainly if there is a high deviation in class distribution among datasets. In those situa-
tions, the FL method also needs to account for such class disparities during training.
3.1.2 Architectures. The second group of studies turn their attention to the proposal of com-
plete architectures for FL applied to medical data. They have a greater focus on the combination of
techniques and the overall distributed architecture required for FL. General studies in this group
include the work from Brisimi et al. [44], which explores the design of a distributed architecture
to solve large-scale sparse support vector machine (SVM) models. The resulting architecture
enables the use of SVM in FL environments for medical data analysis. Lu et al. [45] propose a
completely distributed version of the stochastic gradient descent algorithm, which enables an FL
environment that does not require a central entity for model aggregation (peer-to-peer or server-
less FL framework). Their study focuses on simulated studies with mathematical models that verify
the correctness of their aggregation algorithm. Both studies explore a distributed version of a spe-
cific FL component and then describe a complete architecture around it. Also, these proposals aim
to process historical EHR data to predict the diagnosis of specific cardiac and mental conditions.
Some studies focus on adapting well-known distributed architecture designs and explore their
advantages in the area of FL. Roy et al. [46] employ the concept of BitTorrent swarms to distribute
local model updates to all participants on an FL architecture. Each client individually maintains
the global model according to the received local updates. Sanyal et al. [47] propose the use of a fog
infrastructure to enable FL on constrained IoT medical devices. A fog infrastructure encompasses
computational resources available in a hospital’s local network, which can be used to manage
complex tasks that otherwise would require cloud computing. Such an infrastructure enables the
resources of constrained devices to be dedicated to data collection and simple learning tasks. Xue
et al. [48] explore software-defined networks and mobile edge computing as architectures to de-
ploy an FL system for e-health treatment recommendation based on IoHT devices. Polap et al. [49]
propose a multi-agent infrastructure based on FL to train models for medical diagnosis using dis-
tributed databases. Finally, Hao et al. [50] propose a method to partition neural network models
for distributed training. The method enables clients to employ a cloud platform for part of the
training process, reducing the required computational power. The authors also propose a data per-
turbation method to guarantee data privacy when sharing partial models. These studies explore
how different architectures can be used to distribute the resources required by the FL process, in-
cluding computational resources and local models. They focus on two general types of applications.
The first is the time-series analysis of medical data, particularly for sepsis outcome prediction or
activity monitoring of patients. The second is medical image segmentation and classification of
diagnostic prediction of different types of tumors.
Given the popularity of blockchain as distributed architectures for data authenticity, a group
of studies also explore their applicability to FL. Shae and Tsai [51] propose to use a blockchain to
store anonymous public models that can be freely aggregated by any member of the architecture
that requires a learning model. Lugan et al. [52] further explore the properties of blockchains
and aims to guarantee the authenticity of local learning models provided by individual clients.
Clients also participate in a verification process that guarantees that new contributions improve
the accuracy of the global learning model before their publication. Marulli et al. [53] aim to increase
the confidentiality guarantees of learning models maintained in the blockchain with the use of
homomorphic encryption. Rahman et al. [54] employ a blockchain to share partial training data
so clients can download and train advanced models. Partial models employ differential privacy to
guarantee privacy. All the studies employ the blockchain to authenticate and audit contributions
from peers to the global parameter pool used to generate the global learning model. These studies
focus on a varied number of case studies. Particular applications include image analysis for tissue
recognition or disease diagnostics. In particular, Rahman et al. [54] explore a wide range of learning
model applications tailored for COVID-19 management, such as infection contact management and
outbreak management.
In summary, the studies above explore different distributed designs applied to FL. A few studies
focus on FL’s basic algorithms and propose methods to further increase their decentralization and
scalability. In turn, a large part of this work aims to apply well-known distributed designs to the
problem of FL. These include well-known peer-to-peer architectures as well as microservice-based
fog architectures. Furthermore, these studies examine specific aspects of the FL architecture but
focus on enabling their distribution and scalability.
These studies explore different architecture proposals to support the required communication
and coordination features required by FL. A distributed environment based on Cloud and Fog com-
puting concepts can provide the required scalability in an environment with a stable number of
collaborating institutions. However, a fully distributed architecture, such as a BitTorrent swarm,
enables a scalable solution to support dynamic scenarios like multiple institutions or individual
patients collaborating on a global scale. Also, it can increase the reliability of the overall system
by eliminating known problems such as single points of failure. However, a fully distributed ar-
chitecture brings additional challenges that may require complex algorithms for management or
decision making, for example. One particular challenge relevant to FL is the authenticity of trans-
actions conducted throughout the training process. In that sense, blockchains represent a relevant
architecture that natively provides authenticity guarantees to transactions. Even so, there are still
guarantees necessary to FL with medical data requiring the implementation of specialized mecha-
nisms. The following topics address work related to these specific aspects.
3.1.3 Aggregation Methodologies. Model aggregation is a fundamental step on FL. In summary,

the aggregation step generates a global learning model by combining the parameters received
from each participant in the FL process. The methodology used to select and combine parameters
greatly influences the resulting global model accuracy. Consequently, several studies evaluate and
propose solutions to improve the aggregation process and obtain models with higher accuracy.
Six studies aim to improve the aggregation results by including metrics that classify local mod-
els before being combined, enabling the algorithm to select only those that improve the resulting
model. Boughorbel et al. [55] use the concept of uncertainty to classify the usefulness of local
models. This property allows the aggregation algorithm to select only local models with high
generalization to train the global model. Huang et al. [56] employ a cross-entropy analysis to eval-
uate the quality of local models before sending them to the aggregation server, which defines the
threshold that characterizes a good model. The method reduces the computational and network
requirements for training and improves the resulting model’s accuracy. Cao et al. [57] propose a
metric that quantifies the added knowledge from individual learning models in a federated context.
The aggregation entity can then request only the local models with higher added knowledge to im-
prove training accuracy and reduce network requirements. Chen et al. [58] evaluate the learning
model parameters and propose a “sign” metric that indicates how much improvement an update
contributes to the global model. Updates that decrease the global model accuracy are not sub-
mitted for aggregation. Similarly, Zhang et al. [59] propose a method to select which clients will
participate in the global model training. Clients only upload the local model if it is guaranteed
to contribute to the global model accuracy. Furthermore, the global model also discards updates
received from clients that take too long to train a certain number of epochs for the model. Yang
et al. [60] propose a method that leverages labeled and unlabeled data available in training datasets.
Clients employ the current global FL model to generate pseudo-labels for unlabeled data, enabling
it to improve the global model.
Other studies focus on the aggregation algorithm, evaluating and proposing alternatives to the
traditional gradient descent, which is widely used in FL. Jiang et al. [61] propose the use of an
attention mechanism to classify and select the parameters that will be aggregated in the global
model. Results indicate that the attention aggregation outperforms the gradient descent. Chen et al.
[62] investigate the problem of activation divergence that occurs during the aggregation of models
in FL, which can drastically increase their training time and reduce their accuracy. The authors
propose a solution for this problem with a method that maximizes the entropy of activation vectors
across all FL participants.
Two studies aim at modifying the local models to guarantee that dataset imbalances do not
influence the resulting global model. Yan et al. [63] employ a generative adversarial network
(GAN) to create a synthetic dataset that encompasses the representative characteristics of the orig-
inal training data. The synthetic dataset eliminates data variations that can influence the global
model while preserving the privacy of training data. Wu et al. [64] employ a generative convolu-
tional autoencoder to guarantee class balance among distributed datasets used for in-home health
monitoring.
Finally, two studies focus on improving the aggregation methodology to make it robust against
malicious users trying to degrade the resulting global model. Su and Xu [65] propose a modifica-
tion of the gradient descent method capable of tolerating input data from adversarial workers. The
method can maintain learning model accuracy even with a fraction of malicious clients present in
the infrastructure. In turn, Chen et al. [66] leverage the trusted execution environment available
on modern computer platforms (e.g., Intel SGX) to verify and sign the results from individual con-
tributions to FL. Such a method guarantees model authenticity during aggregation and prevents
insider threats, patients’ health data leakage, and model tempering at the aggregation engine and
edge devices.
The articles in this category demonstrate that the main focus of work related to aggregation is
to obtain the highest possible accuracy from the global model resulting from the process. To that
end, three main lines of investigation exist: (i) applying metrics to rank and select only the models
with higher knowledge derived from the training dataset; (ii) identify issues inherent from the
distributed training process and propose solutions to reduce the deviation that it may cause; and
(iii) mitigate possible attacks stemming from malicious participants trying to hinder the accuracy
of the aggregated model.
The distributed nature of FL requires these additional methods to coordinate the training pro-
cess since a centralized analysis of the entire dataset is not possible. Solutions that evaluate the
accuracy and generality of local models before submitting them for global aggregation provide a
solution without distributed coordination among peers. This characteristic reduces the network re-
quirements from the overall infrastructure while reducing the impact of imbalanced local datasets.
Nevertheless, local mechanisms still rely on local estimates that can deviate from the characteris-
tics of the global dataset. In turn, solutions that exchange dataset quality information among FL
peers can further improve the quality of local models before the aggregation process. However,
these solutions impose additional network requirements to the architecture and may not be suit-
able for scenarios considering the participation of constrained devices (e.g., smartphones). Some
proposals also consistently rely on using local datasets to verify the accuracy of local and global
models. This approach may incur a verification bias similar to what may occur during model train-
ing with local datasets. Such analyses need to account for these biases, possibly using synthetic
or sampled datasets, as some studies proposed. Overall, most studies show that FL imposes addi-
tional processing and communication costs to compensate for the distributed nature assumed by
the training process.
3.1.4 Confidentiality Guarantees. Medical data is sensitive by nature. Consequently, the confi-
dentiality of individual data items used to train models is one of the most stringent requirements
for FL in medical applications. Usually, stewards use traditional mechanisms based on perimeter
security to ensure that only trustworthy parties have access to clinical data. However, scenarios
with data sharing among potentially untrustworthy parties require different security mechanisms,
such as the ones proposed by the zero-trust model [67]. The fundamental assumption of this model
is that no network traffic is trustworthy until its proper authentication and authorization. It is com-
patible with modern system architectures, for example, distributed medical applications based on
IoT [68]. This trust model is particularly relevant to FL applied to medical data because an institu-
tion or patient must distribute sensitive information to parties outside the boundaries of traditional
trust enforcement mechanisms. As a result, no assumptions can be made about the data security
of collaborating partners or communication channels. Furthermore, collaborating parties should
assume that to avoid leakage of sensitive information, the more obfuscated it becomes during the
FL process, the better.
Current studies on FL applied to medical data explore the assumptions of zero-trust models.
Thus, they focus on evaluating and guaranteeing the confidentiality of training datasets to avoid
leaking private data to untrustworthy parties. Early studies on the subject include the work from
Hitaj et al. [69], which explores the use of a GAN to reconstruct the private data used to distributed
train learning models. Their results show that a GAN can defeat most basic privacy mechanisms
designed in traditional FL architectures and the need to develop stronger security mechanisms to
protect medical data.
Three studies look into the application of individual security techniques to guarantee the con-
fidentiality of model parameters and, as a result, the data used to train them. Bonawitz et al. [70]
explore the use of secret sharing to secure model parameters before transmitting local models to
collaborating parties. Dong et al. [71] further investigate the use of secret sharing and conduct
an extensive evaluation on its confidentiality guarantees. Song et al. [72] analyze the efficacy of
homomorphic encryption to protect learning models on FL applied to healthcare applications.
Differential privacy is a method that enables a given database to be modified without changing
its properties and, thus, the results of statistical analyses conducted over the data [73]. It is a partic-
ularly promising technique to secure FL architectures. Li et al. [74] provide an in-depth analysis of
differential privacy applied to FL for healthcare. Pfohl et al. [75] evaluate the accuracy of learning
models trained with three systems: centralized, FL, and FL with differential privacy. Their results
indicate that differential privacy negatively impacts the accuracy of learning models when their
parameters are not correctly adjusted.
Furthermore, a considerable body of work explores the application of differential privacy com-
bined with other techniques. Truex et al. [76] combine secure multiparty communication and dif-
ferential privacy to guarantee the confidentiality of training data during the aggregation of local
models. Zhang et al. [77] combine learning model partitioning and differential privacy to provide
confidentiality guarantees on FL. The proposal aims to maintain a balanced trade-off between
computational cost and privacy guarantees. Ma et al. [78] enhance tensor factorization [23] with
differential privacy to guarantee the confidentiality of training data when applied to computational
phenotyping of EHR. Triastcyn and Faltings [79] protect the parameters from local models with
the Bayesian differential privacy technique. Results indicate that the method further reduces data
leakage compared to traditional differential privacy, but it requires additional assumptions to be
taken on the FL architecture. Li et al. [80] propose an FL architecture that combines differential
privacy and the mixture of experts. The resulting architecture is resilient to inversion attacks on
local models and degradation of model aggregation due to non-independent and identically dis-
tributed (IID) data. Gong et al. [81] propose to dynamically adapt the level of noise introduced
with differential privacy methods applied to FL. The noise level varies with the significance of each
neuron to the overall knowledge, which directly impacts the accuracy of the global model.
Some studies also focus on combining homomorphic encryption with other techniques to im-
prove data and ML models’ confidentiality. Such an approach allows data processing while both
models and data remain encrypted on FL. Hao et al. [82] employ homomorphic encryption and
augmented learning with error to protect the parameters of local learning models during training.
The method reduces the network communication required to secure the aggregation process on
FL without sacrificing model accuracy. Dong et al. [83] combine secret sharing and homomorphic
encryption to protect the parameters from learning models shared in a federated architecture. The
method also employs ternary gradients to encode model parameters to further reduce training
data leakage.
Finally, a group of studies proposes techniques to manipulate learning models to reduce the
chance of leakage during transmission. Cui and Liu [84] propose a hybridization method that re-
combines parameters from local learning models during aggregation and mitigates the possibility
of the derivation of confidential data from the global learning model. Shao et al. [85] employ a sto-
chastic channel-based method that identifies paths on the neural network with lower knowledge
acquired during training. The method prunes these paths before sharing the model, reducing the
chance of confidential data being derived. Ge et al. [86] propose an architecture that divides the
learning model into private and public parts. The public part is trained with a subset of the pri-
vate data to reduce data leakage during aggregation. In turn, the private part combines the global
learning model shared by the central aggregation entity with additional training with the complete
private dataset. Jeon and Kim [87] propose a method that splits local datasets into small batches
individually sent to the aggregation entity. The method reduces the chance that the global model
can be overfitted, which would enable the derivation of the training data from the model.
The literature analysis demonstrates that FL applied to EHR requires extensive work to mitigate
risks to the confidentiality of data. To that end, studies explore various cryptographic techniques
tailored for distributed applications. In particular, previous work demonstrates that homomorphic
encryption [88] and differential privacy [89] can protect the datasets used for neural network train-
ing with minimal impact to model accuracy. These early results also indicate that these techniques
are viable alternatives to mitigate possible attacks in federated learning architectures without com-
promising model accuracy. Results from studies combining different techniques also show that
these techniques can complement each other to increase confidentiality guarantees in FL environ-
ments. The literature also shows that most studies focus on privacy guarantees on neural network
models because they already offer an intrinsic level of privacy due to their non-interpretability.
However, interpretable models can be a requirement for medical applications where the reasoning
behind a given decision or diagnosis is a requirement. For such a scenario, a few studies presented
promising results applying homomorphic encryption and differential privacy to train interpretable
models such as random forests using FL.
3.1.5 Federated Data Analaysis. The final group of studies on the literature corpus focuses on
the particular aspect of federated data analysis. These studies do not aim to generate a learning
model using ML methods over distributed datasets. Rather, they aim to perform an initial analysis
of the distributed data to facilitate the distributed learning procedure.
Some studies aim to process datasets with sensitive information and create anonymous versions
of them to facilitate the distribution. Lee et al. [90] employ data hashing techniques to generate
fingerprints of private databases and enable searches for specific parameters without compromis-
ing the confidentiality of the dataset. Their work enables institutions to explore a dataset prior to
requesting private data for a data analysis study. Huang et al. [91] employ autoencoders and k-
means clustering to generate an anonymous dataset that can be shared and grouped into a global
structure that does not compromise confidential data. The resulting dataset enables distributed
model training with lower confidentiality requirements for communication mechanisms. Fioretto
and Van Hentenryck [92] employ a two-step process to anonymize datasets that will be shared
for ML. The process employs a predictor mechanism that describes all the parameters and data
formats required by a specific learning task. This predictor is distributed to dataset owners and
employed to generate anonymized datasets with the data already converted to the required format.
The second group of studies aims to enable the evaluation of distributed datasets to reduce
data impurity prior to its usage for learning tasks. Liu et al. [93] employ a GAN to generate the
missing data on local databases before use on an ML algorithm. The authors aim to improve the
accuracy of local learning models used in the aggregation process of FL. Pezoulas et al. [94] propose
an architecture for curation and federated analysis of medical datasets. Their goal is to enable
dataset cleaning, model harmonization, and distributed analysis over data from multiple medical
institutions without sacrificing confidentiality.
These studies demonstrate that there are issues beyond the scope of the FL process applied to
medical and EHR data. The selected literature corpus presents work on generating anonymous
copies of sensitive datasets to facilitate distribution and preparation prior to learning tasks. These
studies aim to find solutions that enable FL peers to exchange information about local datasets
and mitigate known problems such as class imbalances. They show promising results in achiev-
ing this goal without compromising data privacy. However, such methods may require additional
data exchange among peers, thus increasing the overall network cost incurred by the FL process.
Nevertheless, the possibility to align the private datasets to the specific requirements of learning
tasks is essential to guarantee the quality of the locally trained models, which will directly impact
the result of the overall process. Finally, results from these studies can be enhanced with privacy
methods previously reviewed, such as homomorphic encryption.
3.2 Machine Learning Methods and Applications

Machine learning methods have a wide application in healthcare [95]. The results from our litera-
ture review also demonstrate this fact with the variety of case studies explored with FL. This sec-
tion analyzes the machine learning methods used in studies about FL applied to EHR data and their
main application. Table 2 summarizes the machine learning methods and applications observed in
the literature corpus. Some articles did not describe a specific case study on the application of
machine learning to healthcare data. Hence, the table excludes these particular studies.
Deep learning is the most common technique observed in the literature corpus, being used
in almost 79% of the reported case studies. One particular area that benefited greatly from deep
learning in computer vision. This trend is reflected in the literature corpus, in which all studies
that explore medical image analysis employ deep learning. One area of particular interest is the
analysis of neurological magnetic resonance imaging (MRI) data. Studies from Sheller et al. [24],
Roy et al. [46], Li et al. [74], and Sheller et al. [33] employ FL to train models for segmentation of
brain MRI data, in particular to identify regions with tumors. Some studies also propose to analyze
MRI data to identify patterns representative of specific neurological conditions. In particular, Silva
et al. [27] apply the ENIGMA Shape Analysis for prediction of neurological diseases in general, Lu
et al. [45] employ shallow neural networks to diagnose Alzheimer’s disease on patients with mild
cognitive impairment, and Li et al. [80] use a multi-layer perceptron to identify autism spectrum
disorder. Finally, Cao et al. [57] focus on medical images for cardiac diagnostics, proposing the use
of deep neural networks to segment coronary vessels.
Another common application of deep learning with EHR data is the prediction of patient out-
comes. In particular, studies from Liu et al. [25], Cui and Liu [84], Huang et al. [56], Pfohl et al.
[75], Sharma et al. [29], Shao et al. [85], Gong et al. [81], and Vaid et al. [34] focus on predicting the
mortality of patients using EHR data from ICUs. This case study is of particular interest to work
Table 2. Machine Learning Methods and Applications with EHR Data
Support Vector Machines
EHR Data Classification
Medical Image Analysis

EHR Data Extraction
Outcome Prediction
Disease Diagnostics
Bayesian Networks
Deep Learning
Regression
Clustering
Kim et al. [23]
Jochems et al. [22]
Liu et al. [25]
Shae and Tsai [51]
Brisimi et al. [44]
Sheller et al. [24]
Lee et al. [90]
Ma et al. [78]
Lu et al. [45]
Pfohl et al. [75]
Boughorbel et al. [55]
Liu et al. [26]
Huang et al. [91]
Cui and Liu [84]
Sharma et al. [29]
Shao et al. [85]
Liu et al. [93]
Roy et al. [46]
Sanyal et al. [47]
Huang et al. [56]
Silva et al. [27]
Triastcyn and Faltings [79]
Hao et al. [82]
Lugan et al. [52]
Truex et al. [76]
Fioretto and Van Hentenryck [92]
Jiang et al. [28]
Li et al. [74]
Chen et al. [62]
Ge et al. [86]
Ribero et al. [30]
Cao et al. [57]
Li et al. [80]
Pezoulas et al. [94]
Deist et al. [31]
Dong et al. [83]
Gong et al. [81]
Yan et al. [63]
Wu et al. [64]
Hao et al. [50]
Chen et al. [58]
Rahman et al. [54]
Lim et al. [43]
Guo et al. [38]
Chen et al. [39]
Ju et al. [40]
Zhang et al. [41]
Szatmari et al. [32]
Sheller et al. [33]
Zhang et al. [59]
Yang et al. [60]
Xue et al. [48]
Polap et al. [49]
Can and Ersoy [42]
Vaid et al. [34]
Lee et al. [35]
Liu et al. [36]
on FL because high-quality datasets are available to conduct experiments in a simulated federated

environment. Such databases include the MIMIC-III [96] and the eICU [97], which are widely used
for research on the application of ML for ICU EHR data analysis. Various of the above studies em-
ploy these databases with the same goal but present different experimental parameters (e.g., data
division among nodes) related to the particular aspect of FL being investigated. One particular
study, from Boughorbel et al. [55], employs deep learning to predict the outcomes of preterm birth
using EHR data.
Some studies use deep learning to evaluate and classify the information contained in EHR data.
Most of these studies aim to extract ICD (international classification of diseases) from EHRs
and classify patients according to such data. For example, Shae and Tsai [51] and Chen et al. [62]
employ deep learning to generate models to classify the records from an EHR database. The work
from Kim et al. [23] and Ma et al. [78] focus on the computational phenotyping of EHR datasets to
identify classifications of each health record. Finally, Ge et al. [86] employ a convolutional neural
network to analyze unstructured medical texts and classify them according to multiple parame-
ters (e.g., diagnosis and treatment). Some work goes further on the classification of EHR data and
aims to achieve disease diagnostics. Lee et al. [90] and Liu et al. [93] employ EHR data from ICUs
of multiple institutions to generate automated diagnostics models with deep learning. The work
from Triastcyn and Faltings [79] focuses specifically on diseases related to blindness, employing a
convolutional neural network to detect them.
One final group of studies that focuses on deep learning aims to use models for hand-writing
[52, 82, 83] and speech [28] recognition. They do not focus specifically on analyzing EHR data,
but they explore the usage of hand-written health records data to extract and interpret medical
information.
The remainder of the literature corpus explores FL with other machine learning techniques.
Jochems et al. [22] employ a Bayesian network to evaluate the post-treatment survival of cancer
patients. Pezoulas et al. [94] use a Multinominal Naïve Bayes to determine the clinical outcome
of lymphoma patients. Brisimi et al. [44] employ sparse support vector machines with binary su-
pervised classification problems to predict hospitalizations due to cardiac events. Sanyal et al. [47]
focus on time-series analysis techniques for real-time evaluation of data from IoT medical devices
(e.g., multi-parametric monitors). Deist et al. [31] apply linear regression methods to predict the
two-year survival of lung cancer patients. Finally, Ribero et al. [30] use K-means clustering to
recommend treatments based on datasets about drug administration.
Regarding the combination of multiple techniques, the work from Liu et al. [26] employ a
deep averaging network and SVM to predict diagnoses using plain text EHR records. Fioretto and
Van Hentenryck [92] use linear regression and support vector machines for diabetes prediction.
Finally, Truex et al. [76] explore multiple machine learning algorithms to classify EHR data from
nurseries.
3.3 Discussion
This section builds upon the results observed so far in our literature analysis and draws some
insights on the current state and future perspectives for FL applied to EHR data. Based on the
data collected from our literature corpus, we propose a taxonomy to classify the observed work.
Figure 3 illustrates the proposed taxonomy, including the percentage of studies that fall into each
category.
The taxonomy reveals that deep learning is the most explored ML methodology when using
FL with EHR data. As previously mentioned in our analysis, this is an expected result given the
current popularity of deep learning methods and their variations. Furthermore, it is important to
Fig. 3. Literature taxonomy for FL applied to EHR.
observe that various techniques employed in FL have high compatibility with deep learning-based
models. For example, neural networks provide an intrinsic obfuscation to the data used for training
because the interpretation of parameters is not a straightforward task. This fact makes methods
based on neural networks ideal candidates for FL architectures. Nevertheless, as demonstrated by
Hitaj et al. [69], this property does not automatically result in completely confidential models nor
protects data while training and, thus, more advanced security mechanisms must be included in
the architecture.
A few studies also explore other common learning models for medical applications, such as
Bayesian networks or support vector machines. Some of these methods inherently expose the
meaning of their parameters and require additional mechanisms to guarantee the confidentiality of
sensitive medical data. On the other hand, these mechanisms enable the expansion of FL to a wide
variety of ML techniques beyond deep learning, including other classical methods from medical
data analysis. Such work requires an in-depth analysis of candidate ML methods and manipulating
their parameters to obfuscate the respective training data.
Regarding the aspects from FL approached by different studies, there is a distinct interest in
studies related to the confidentiality guarantees provided to the training data and resulting mod-
els. This interest is expected because of the high sensitivity of medical data and its regulatory
protection requirements (e.g., GDPR or HIPPA). As a result, current work evaluates the degree of
confidentiality provided by different FL approaches. Furthermore, it proposes mechanisms to mit-
igate parameter data leakage or model tampering that may expose protected information. These
mechanisms focus on classical distributed confidentiality methods, such as differential privacy or
secret sharing, to protect the distribution of learning models among collaborating partners. How-
ever, there are open questions on extending these cryptographic techniques to support ML meth-
ods such as decision trees, random forests, and KNN. These techniques may impose significant
performance penalties for the ML models since they add another layer for security and privacy.
Therefore, their performance needs to be improved dramatically.
Another subject of particular interest is the evaluation of case studies of EHR data analysis.
This is a result of the novelty of the subject. There is a high interest in evaluating the advantages
of FL applied to medical data and what scenarios present the most gains. These studies indicate
that current case studies align with the traditional applications of ML to medical data, including
prediction of patient outcomes, automated analysis of medical images, and automated diagnosis
based on parameter analysis.
Current work presents an equal interest in architectures for FL, aggregation methods for model
generation, and federated data analysis. The first two subjects are directly related to the communi-
cation cost and quality of the learning models extracted from the learning process. These subjects
are of high relevance to FL architectures in general. Thus, it is expected that other subjects have a
higher frequency in our results, given the focus of our literature search. Nevertheless, some studies
approach these two subjects by looking at the specific characteristics of medical data and learn-
ing models. In turn, federated data analysis encompasses a series of studies that aim at evaluating
and preparing datasets prior to the use of ML methodologies. This area is of particular interest
because many applications require the analysis of distributed datasets to enable medical decision-
making without the particular requirement of learning models. Consequently, there is space for
further work on federated data analysis for medical data that can take advantage of the body of
knowledge developed from work with FL.
Finally, regarding the applications of FL on EHR data, there is even interest in almost all subjects
observed in the literature corpus. Outcome prediction, in particular, presented a higher interest
than the average. As mentioned earlier, this results from the availability of high-quality public
datasets that enable studies on outcome prediction based on ICU information. As a result, vari-
ous works use such an application as a case study to evaluate FL mechanisms applied to medical
data.
The observations discussed in this section stem from the analysis of the collected literature
corpus. They enable us to better understand the current state of FL architectures applied in general
and their requirements when EHR data is the focus of their application. Based on these insights,
the following section discusses an overview of the main components of an architecture for FL for
EHR data.
4 ARCHITECTURE FOR FEDERATED ANALYSIS AND LEARNING WITH EHR DATA

This section presents an architecture for federated analysis and learning with EHR data. The main
goal of this architecture is to enable healthcare institutions (e.g., hospitals and laboratories) with
access to private medical datasets to employ them in distributed data analysis and ML studies
without compromising patient confidentiality. These institutions, defined as data owners, can
maintain datasets with medical data of different characteristics, such as imaging exams, diagnostics
results, and drug prescriptions. These datasets comprise sensitive information protected by laws
and regulations, and individual patients should provide their explicit consent to make their data
available for possible analysis in possible studies. We consider the acquisition of such consent the
be an orthogonal issue to the design of the FL architecture, which can be tackled in future work.
Nevertheless, we will consider mitigating threats to data privacy during data analysis or model
training, as this is a key aspect of FL.
The architecture’s design aims to leverage the best practices and methodologies observed in
recent FL literature, as observed throughout Section 3. Figure 4 illustrates the components and their
distribution among the different entities that take part in the learning process. These components
are discussed in more detail next.
Data owners are presumed to be aware that participation in the FL infrastructure entails collab-
orating with privacy-preserving analyses and learning models. In exchange, the entity can request
the execution of analysis and learning algorithms on other private entities to conduct medical data
experiments. Other incentive mechanisms may exist within the FL framework, for example, finan-
cial or regulatory. Furthermore, current literature suggests that a blockchain infrastructure may
be used as a collaborative infrastructure in which each partner becomes a node of the network and
can use the learning models shared by other peers.
Fig. 4. Overview of an architecture for federated analysis and learning with EHRdata.
Due to the sensitive nature of the information contained in local datasets, the data owners will
never directly share information with other parties. Furthermore, the data formats and standards
used to store EHR data in each dataset can be disjoint among collaborators of the learning process.
Each data owner uses their private datasets to train a private model, which can also include
parameters obtained from models shared by other data owners from the FL manager. The data
owner uses these private models to conduct internal data analyses and are not supposed to be
directly shared with other FL entities.
Information sharing among data holders requires confidentiality guarantees for learning model
parameters and analyses results. To that end, a data confidentiality module analyzes outgoing
information (e.g., analysis results and learning models) to guarantee that it does not contain any
sensitive information. To that end, this module can implement different methodologies based on
a zero-trust security model [67]. Collaborating institutions will use public networks to exchange
analysis results and learning models that may reveal private information. As such, security mech-
anisms must provide strong confidentiality guarantees, even if the institutions can establish trust
among themselves for sharing medical data. Current literature shows that homomorphic encryp-
tion and differential privacy are techniques with substantial results to avoid data leaks without
prohibitive computational costs [98]. If necessary, all collaborating parties should agree on the
parameters of the confidentiality mechanisms (e.g., cryptographic keys).
The dataset processing module implements mechanisms that enable external learning algo-
rithms to access and interpret the data contained in local datasets. Since the local data of each
private entity may be in a particular format, tasks such as type conversion or image process-
ing may be needed to enable a particular learning algorithm to process it. This module should
receive a template of the required data formats and identify the necessary conversions based on
private data. To that end, a mechanism based on metadata can provide information on the proper-
ties of a dataset, enabling automated mechanisms to process data for compatibility with learning
algorithms. The learning algorithm verification module implements mechanisms to analyze
algorithms received from third parties and guarantee they will not purposely expose sensitive
information from private datasets. One possibility to implement such verification is to include a
“manifesto” with each learning algorithm that specifies input and outputs. Learning algorithms can
then be checked to guarantee that they follow the manifesto specification. If a further guarantee is
required, the manifesto can be human-verified to guarantee it does not violate any confidentiality
restriction.
The model training module is responsible for executing the training algorithm, taking the data
from the local datasets, and generating the resulting learning models. It executes both private and
external learning algorithms, but, in the latter case, it presumes that the modules described above
will verify the necessary conditions to avoid private data leakage. In turn, the federated data
analysis module focuses on private or remote requests for data analyses that do not require learn-
ing algorithms. As mentioned before, federated data analysis is of particular interest to medical
applications because it enables decision-making with broader knowledge about a given condition.
The relevant security modules will also verify the external requests for federated data analysis to
guarantee they will not leak any sensitive information. It is worth noting that data analysis or
model training tasks require data holders to provide a certain amount of computational power to
provide the required partial results. Thus, an incentive mechanism becomes necessary to justify
this computational power. One straightforward incentive is the possibility of data holders using
the analysis results to their own needs. However, such benefit requires that a given study provides
relevant results to every data holder, which may be unfeasible. Thus, there is space for further de-
velopment of incentive mechanisms for participation in FL studies, aiming at offsetting the costs
incurred to data holders for knowledge generation.
All data holders communicate among themselves using a communication infrastructure that
enables the secure exchange of learning models and analyses results through insecure networking
environments. This is an important assumption given that the data holder will not be within the
same administrative domain and must communicate using the Internet. The communication infras-
tructure also connects all data holders with the federated learning manager, which is the central
element that coordinates the distributed learning and analyses processes. A common approach to
implementing the manager is a centralized element (e.g., a cloud instance). However, decentral-
ized approaches can also provide the functionality required to the architecture. For example, the
literature corpus mentioned the use of swarm and blockchain infrastructures. Such peer-to-peer
algorithms might be explored to deploy the complete functionality required from the public entity
as part of the private entities themselves.
The manager implements the Model Aggregation methodologies required to collect all public
models from private entities and merge them into a global learning model. Such methodologies
will depend on the learning algorithms employed by the private entities. It is expected that the
entity willing to train a certain learning model will provide the respective aggregation algorithm
to the public entity. The federated data analysis module of the public entity focuses on the aggrega-
tion of data analyses results provided by private entities to generate the final analysis result. The
public entity also implements its learning algorithm verification module, which can store and pro-
vide manifestos and participate in distributed analysis mechanisms requiring consensus among
all participants. Current literature shows a variety of studies proposing methodologies to im-
prove the aggregation process to improve the accuracy of global models. Results show that model
accuracy varies depending on the type of learning model and the characteristics of local datasets.
As a result, support for different types of aggregation mechanisms may be necessary to guarantee
compatibility with different learning algorithms.
The manager maintains a repository of public models generated with the aggregation method-
ology. Depending on the model’s goals, it can be forwarded to the individual party that requested
the training. They can also be redistributed to all participants in the learning process for their
private model training or benchmarking with private data. The public entity also maintains a
repository with all analysis and learning algorithms being processed by the infrastructure.
The description of the architecture and components presented above is not intended to be in-
depth or complete. Our goal in this section is to provide an overview of the components that may
be part of a more extensive infrastructure destined to analyze and learn EHR data in a federated
fashion. Many elements required by these components are active research topics or still require
studies to properly investigate them. As such, this description is intended to indicate which topics
are still open issues in the viability of large-scale FL with EHR data.
5 FINAL REMARKS
Federated Learning presents itself as a promising solution to train machine learning models with
large and diverse datasets without compromising information confidentiality. This characteristic
is significant for healthcare applications, in which EHR data is very sensitive by nature and often
cannot be easily shared. While there are some successful case studies for the use of FL in production
scenarios, various research questions remain open before these architectures become available for
widespread usage for healthcare applications.
This paper explored the landscape of research about FL applied to EHR data based on a liter-
ature corpus of 67 recent articles curated using a well-defined systematic review methodology.
The analysis of these articles demonstrates various efforts to propose and improve methods to
guarantee the confidentiality of training data. This result is straightforward given the stringent
privacy restrictions related to the medical data from individual patients. Multiple studies also
explore the application of FL on different types of medical data and how the resulting models
compare against traditional training methods. Other topics include model aggregation methods
tailored for the requirements of medical data and architectures to enable FL with multiple med-
ical institutions. These insights indicate that FL applied to medical data is an active research
area, and its fundamental aspects are still under development. It is important to note that the
literature corpus analyzed in this study focuses on FL studies in healthcare contexts. However,
FL techniques applied in other areas (e.g., Industry 4.0) may become relevant to the healthcare
context in the future. Analyzing such articles and evaluating their applicability to healthcare sce-
narios is relevant. However, it lies beyond the scope of this review and is considered as future
work.
Regarding research directions, the confidentiality of training data remains an open issue that
requires further investigation and proposals. Particularly, there is space to explore confidential-
ity mechanisms for ML methods different from deep learning when used with federated training.
There is space to investigate the effectiveness of aggregation methods with expanded characteris-
tics, for example, with resilience to training problems on local models that may impact the accuracy
of the global result. Finally, there is the need to further explore methods to enable the normaliza-
tion of local datasets and to guarantee that different local datasets can provide the expected inputs
and parameters to training models. Such dataset normalization used in an automated fashion is an
open issue that is key for the widespread availability of FL in general.
REFERENCES
[1] D. W. Bates, S. Saria, L. Ohno-Machado, A. Shah, and G. Escobar. 2014. Big data in health care: Using analytics to
identify and manage high-risk and high-cost patients. Health Affairs 33 (2014), 1123–1131.
[2] J. Wiens and E. S. Shenoy. 2017. Machine learning for healthcare: On the verge of a major shift in healthcare epidemi-
ology. Clinical Infectious Diseases 66 (2017), 149–153.
[3] N. H. Shah, A. Milstein, and S. C. Bagley. 2019. Making machine learning models clinically useful. JAMA 322 (2019),
1351–1352.
[4] S. Keyhani, P. L. Hebert, J. S. Ross, A. Federman, C. W. Zhu, and A. L. Siu. 2008. Electronic health record components
and the quality of care. JSTOR Medical Care 46 (2008), 1267–1272.
[5] J. King, V. Patel, E. W. Jamoom, and M. F. Furukawa. 2014. Clinical benefits of electronic health record use: National
findings. Health Services Research 49 (2014), 392–404.
[6] B. Shickel, P. J. Tighe, A. Bihorac, and P. Rashidi. 2018. Deep EHR: A survey of recent advances in deep learning
techniques for electronic health record (EHR) analysis. IEEE Journal of Biomedical and Health Informatics 22 (2018),
1589–1604.
[7] T. Chilimbi, Y. Suzue, J. Apacible, and K. Kalyanaraman. 2014. Project Adam: Building an efficient and scalable deep
learning training system. In 11th USENIX Symposium on Operating Systems Design and Implementation (OSDI’14), 2014,
pp. 571–582.
[8] E. Vayena, A. Blasimme, and I. G. Cohen. 2018. Machine learning in medicine: Addressing ethical challenges. PLOS
Medicine 15 (2018), 1–4.
[9] E. Horvitz and D. Mulligan. 2015. Data, privacy, and the greater good. Science 349 (2015), 253–255.
[10] Q. Yang, Y. Liu, T. Chen, and Y. Tong. 2019. Federated machine learning: Concept and applications. ACM Transactions
on Intelligent Systems and Technologies 10 (2019), 12:1–12:19.
[11] R. Gu, C. Niu, F. Wu, G. Chen, C. Hu, C. Lyu, and Z. Wu. 2021. From server-based to client-based machine learning:
A comprehensive survey. ACM Computing Surveys 54 (2021).
[12] A. Hard, K. Rao, R. Mathews, S. Ramaswamy, F. Beaufays, S. Augenstein, H. Eichner, C. Kiddon, and D. Ramage. 2018.
Federated learning for mobile keyboard prediction. https://arxiv.org/abs/1811.03604. arXiv:1811.03604.
[13] P. Bellavista, L. Foschini, and A. Mora. 2021. Decentralised learning in federated deployment environments: A system-
level survey. ACM Computing Surveys 54 (2021).
[14] J. Xu, B. S. Glicksberg, C. Su, P. Walker, J. Bian, and F. Wang. 2019. Federated learning for healthcare informatics.
https://arxiv.org/abs/2003.08119. arXiv:1911.06270.
[15] N. Rieke, J. Hancox, W. Li, F. Milletari, H. Roth, S. Albarqouni, S. Bakas, M. N. Galtier, B. Landman, K. Maier-Hein,
S. Ourselin, M. Sheller, R. M. Summers, A. Trask, D. Xu, M. Baust, and M. J. Cardoso. 2020. The future of digital health
with federated learning. https://arxiv.org/abs/2003.08119. arXiv:2003.08119.
[16] F. Zerka, S. Barakat, S. Walsh, M. Bogowicz, R. T. H. Leijenaar, A. Jochems, B. Miraglio, D. Townend, and P. Lambin.
2020. Systematic review of privacy-preserving distributed machine learning from federated databases in health care.
JCO Clinical Cancer Informatics 4 (2020), 184–200.
[17] L. Li, Y. Fan, M. Tse, and K.-Y. Lin. 2020. A review of applications in federated learning. Computers & Industrial Engi-
neering 149 (2020), 106854.
[18] V. Mothukuri, R. M. Parizi, S. Pouriyeh, Y. Huang, A. Dehghantanha, and G. Srivastava. 2021. A survey on security
and privacy of federated learning. Future Generation Computer Systems 115 (2021), 619–640.
[19] C. Zhang, Y. Xie, H. Bai, B. Yu, W. Li, and Y. Gao. 2021. A survey on federated learning. Knowledge-Based Systems 216
(2021), 106775.
[20] H. Zhang and M. Ali Babar. 2013. Systematic reviews in software engineering: An empirical investigation. Elsevier
Information and Software Technology 55 (2013), 1341–1354.
[21] T. Yang, G. Andrew, H. Eichner, H. Sun, W. Li, N. Kong, D. Ramage, and F. Beaufays. 2018. Applied federated learning:
Improving Google keyboard query suggestions. https://arxiv.org/abs/1812.02903. arXiv:1812.02903.
[22] A. Jochems, T. M. Deist, I. El Naqa, M. Kessler, C. Mayo, J. Reeves, S. Jolly, M. Matuszak, R. Ten Haken, J. van Soest,
C. Oberije, C. Faivre-Finn, G. Price, D. de Ruysscher, P. Lambin, and A. Dekker. 2017. Developing and validating a
survival prediction model for NSCLC patients through distributed learning across 3 countries. International Journal
of Radiation Oncology, Biology, Physics (2017), 344–352.
[23] Y. Kim, J. Sun, H. Yu, and X. Jiang. 2017. Federated tensor factorization for computational phenotyping. https://arxiv.
org/abs/1704.03141. arXiv:1704.03141.
[24] M. J. Sheller, G. A. Reina, B. Edwards, J. Martin, and S. Bakas. 2019. Multi-institutional deep learning modeling without
sharing patient data: A feasibility study on brain tumor segmentation. Brainlesion (2019), 92–104.
[25] D. Liu, T. Miller, R. Sayeed, and K. D. Mandl. 2018. FADL: Federated-autonomous deep learning for distributed elec-
tronic health record. https://arxiv.org/abs/1811.11400. arXiv:1811.11400.
[26] D. Liu, D. Dligach, and T. Miller. 2019. Two-stage federated phenotyping and patient representation learning. https:
//arxiv.org/abs/1908.05596. arXiv:1908.05596.
[27] S. Silva, B. A. Gutman, E. Romero, P. M. Thompson, A. Altmann, and M. Lorenzi. 2019. Federated learning in dis-
tributed medical databases: Meta-analysis of large-scale subcortical brain data. In IEEE 16th International Symposium
on Biomedical Imaging (ISBI’19). 270–274.
[28] D. Jiang, Y. Song, Y. Tong, X. Wu, W. Zhao, Q. Xu, and Q. Yang. 2019. Federated topic modeling. In International
Conference on Information and Knowledge Management (CIKM’19). 1071–1080.
[29] P. Sharma, F. E. Shamout, and D. A. Clifton. 2019. Preserving patient privacy while training a predictive model of
in-hospital mortality. https://arxiv.org/abs/1912.00354. arXiv:1912.00354.
[30] M. Ribero, J. Henderson, S. Williamson, and H. Vikalo. 2020. Federating recommendations using differentially private
prototypes 2020. https://arxiv.org/abs/2003.00602. arXiv:2003.00602.
[31] T. M. Deist, F. J. Dankers, P. Ojha, M. Scott Marshall, T. Janssen, C. Faivre-Finn, C. Masciocchi, V. Valentini, J. Wang,
J. Chen, Z. Zhang, E. Spezi, M. Button, J. Jan Nuyttens, R. Vernhout, J. van Soest, A. Jochems, R. Monshouwer, J. Bussink,
G. Price, P. Lambin, and A. Dekker. 2020. Distributed learning on 20 000+ lung cancer patients – the personal health
train. Elsevier Radiotherapy and Oncology (2020), 189–200.
[32] T.-I. Szatmari, M. K. Petersen, M. J. Korzepa, and T. Giannetsos. 2020. Modelling audiological preferences using feder-
ated learning. In 28th ACM Conference on User Modeling, Adaptation and Personalization (UMAP’20). 187–190.
[33] M. J. Sheller, B. Edwards, G. A. Reina, J. Martin, S. Pati, A. Kotrotsou, M. Milchenko, W. Xu, D. Marcus, R. R. Colen, and
S. Bakas. 2020. Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient
data. Nature Scientific Reports 10 (2020), e24207.
[34] A. Vaid, S. K. Jaladanki, J. Xu, S. Teng, A. Kumar, S. Lee, S. Somani, I. Paranjpe, J. K. D. Freitas, T. Wanyan, K. W. Johnson,
M. Bicak, E. Klang, Y. J. Kwon, A. Costa, S. Zhao, R. Miotto, A. W. Charney, E. Böttinger, Z. A. Fayad, G. N. Nadkarni,
F. Wang, and B. S. Glicksberg. 2021. Federated learning of electronic health records to improve mortality prediction
in hospitalized patients with Covid-19: Machine learning approach. JMIR Medical Informatics 9 (2021), e24207.
[35] H. Lee, Y. J. Chai, H. Joo, K. Lee, J. Y. Hwang, S.-M. Kim, K. Kim, I.-C. Nam, J. Y. Choi, H. W. Yu, M.-C. Lee, H. Masuoka,
A. Miyauchi, K. E. Lee, S. Kim, and H.-J. Kong. 2021. Federated learning for thyroid ultrasound image analysis to
protect personal information: Validation study in a real health care environment. JMIR Medical Informatics 9 (2021),
e25869.
[36] J. C. Liu, J. Goetz, S. Sen, and A. Tewari. 2021. Learning from others without sacrificing privacy: Simulation comparing
centralized and federated machine learning on mobile health data. JMIR Mhealth and Uhealth 9 (2021), e23728.
[37] C. A. da Costa, C. F. Pasluosta, B. Eskofier, D. B. da Silva, and R. da Rosa Righi. 2018. Internet of Health Things: Toward
intelligent vital signs monitoring in hospital wards. Artificial Intelligence in Medicine 89 (2018), 61–69.
[38] Y. Guo, F. Liu, Z. Cai, L. Chen, and N. Xiao. 2020. FEEL: A federated edge learning system for efficient and privacy-
preserving mobile healthcare. In 49th International Conference on Parallel Processing (ICPP’20). 1–11.
[39] Y. Chen, X. Qin, J. Wang, C. Yu, and W. Gao. 2020. Fedhealth: A federated transfer learning framework for wearable
healthcare. IEEE Intelligent Systems 35 (2020), 83–93.
[40] C. Ju, D. Gao, R. Mane, B. Tan, Y. Liu, and C. Guan. 2020. Federated transfer learning for EEG signal classification. In
42nd Annual International Conference of the IEEE Engineering in Medicine Biology Society (EMBC’20). 3040–3045.
[41] M. Zhang, Y. Wang, and T. Luo. 2020. Federated learning for arrhythmia detection of non-IID ECG. In IEEE 6th Inter-
national Conference on Computer and Communications (ICCC’20). 1176–1180.
[42] Y. S. Can and C. Ersoy. 2021. Privacy-preserving federated deep learning for wearable IoT-based biomedical monitor-
ing. ACM Transactions on Internet Technology 21 (2021).
[43] W. Y. B. Lim, S. Garg, Z. Xiong, D. Niyato, C. Leung, C. Miao, and M. Guizani. 2020. Dynamic contract design for
federated learning in smart healthcare applications. IEEE Internet of Things Journal (2020), 1–10.
[44] T. S. Brisimi, R. Chen, T. Mela, A. Olshevsky, I. C. Paschalidis, and W. Shi. 2018. Federated learning of predictive models
from federated electronic health records. Elsevier International Journal of Medical Informatics 112 (2018), 59–67.
[45] S. Lu, Y. Zhang, Y. Wang, and C. Mack. 2019. Learn electronic health records by fully decentralized federated learning.
[46] A. G. Roy, S. Siddiqui, S. Pölsterl, N. Navab, and C. Wachinger. 2019. Braintorrent: A peer-to-peer environment for
decentralized federated learning. https://arxiv.org/abs/1905.06731. arXiv:1905.06731.
[47] S. Sanyal, D. Wu, and B. Nour. 2019. A federated filtering framework for Internet of Medical Things. https://arxiv.org/
abs/1905.01138. arXiv:1905.01138.
[48] Z. Xue, P. Zhou, Z. Xu, X. Wang, Y. Xie, X. Ding, and S. Wen. 2021. A resource-constrained and privacy-preserving
edge computing enabled clinical decision system: A federated reinforcement learning approach. IEEE Internet of Things
Journal (2021), 1–17.
[49] D. Polap, G. Srivastava, and K. Yu. 2021. Agent architecture of an intelligent medical system based on federated
learning and blockchain technology. Journal of Information Security and Applications 58 (2021), 102748.
[50] M. Hao, H. Li, G. Xu, Z. Liu, and Z. Chen. 2020. Privacy-aware and resource-saving collaborative learning for health-
care in cloud computing. In 2020 IEEE International Conference on Communications (ICC’20). 1–6.
[51] Z. Shae and J. Tsai. 2018. Transform blockchain into distributed parallel computing architecture for precision medicine.
In IEEE 38th International Conference on Distributed Computing Systems (ICDCS’18). 1290–1299.
[52] S. Lugan, P. Desbordes, E. Brion, L. X. R. Tormo, A. Legay, and A. Macq. 2019. Secure architectures implementing
trusted coalitions for blockchained distributed learning (TCLearn). IEEE Access 7 (2019), 181789–181799.
[53] F. Marulli, E. Bellini, and S. Marrone. 2020. A security-oriented architecture for federated learning in cloud environ-
ments. In Web, Artificial Intelligence and Network Applications (WAINA’20). 730–741.
[54] M. A. Rahman, M. S. Hossain, M. S. Islam, N. A. Alrajeh, and G. Muhammad. 2020. Secure and provenance enhanced In-
ternet of Health Things framework: A blockchain managed federated learning approach. IEEE Access 8 (2020), 205071–
205087.
[55] S. Boughorbel, F. Jarray, N. Venugopal, S. Moosa, H. Elhadi, and M. Makhlouf. 2019. Federated uncertainty-aware
learning for distributed hospital EHR data. https://arxiv.org/abs/1910.12191. arXiv:1910.12191.
[56] L. Huang, Y. Yin, Z. Fu, S. Zhang, H. Deng, and D. Liu. 2018. LoAdaBoost: Loss-based AdaBoost federated machine
learning on medical data. https://arxiv.org/abs/1811.12629. arXiv:1811.12629.
[57] T.-D. Cao, T. Truong-Huu, H. Tran, and K. Tran. 2020. A federated learning framework for privacy-preserving and
parallel training. https://arxiv.org/abs/2001.09782. arXiv:2001.09782.
[58] H. Chen, H. Li, G. Xu, Y. Zhang, and X. Luo. 2020. Achieving privacy-preserving federated learning with irrelevant
updates over e-health applications. In 2020 IEEE International Conference on Communications (ICC’20). 1–6.
[59] W. Zhang, T. Zhou, Q. Lu, X. Wang, C. Zhu, H. Sun, Z. Wang, S. K. Lo, and F.-Y. Wang. 2021. Dynamic fusion-based
federated learning for Covid-19 detection. IEEE Internet of Things Journal (2021), 1–8.
[60] D. Yang, Z. Xu, W. Li, A. Myronenko, H. R. Roth, S. Harmon, S. Xu, B. Turkbey, E. Turkbey, X. Wang, W. Zhu, G. Car-
rafiello, F. Patella, M. Cariati, H. Obinata, H. Mori, K. Tamura, P. An, B. J. Wood, and D. Xu. 2021. Federated semi-
supervised learning for Covid region segmentation in chest CT using multi-national data from China, Italy, Japan.
Medical Image Analysis 70 (2021), 101992.
[61] J. Jiang, S. Ji, and G. Long. 2020. Decentralized knowledge acquisition for mobile internet applications. Springer World
Wide Web (2020).
[62] W. Chen, K. Bhardwaj, and R. Marculescu. 2020. FedMAX: Mitigating activation divergence for accurate and
communication-efficient federated learning. https://arxiv.org/abs/2004.03657. arXiv:2004.03657.
[63] Z. Yan, J. Wicaksana, Z. Wang, X. Yang, and K.-T. Cheng. 2020. Variation-aware federated learning with multi-source
decentralized medical image data. IEEE Journal of Biomedical and Health Informatics (2020), 1–14.
[64] Q. Wu, X. Chen, Z. Zhou, and J. Zhang. 2020. FedHome: Cloud-edge based personalized federated learning for in-home
health monitoring. IEEE Transactions on Mobile Computing (2020), 1–14.
[65] L. Su and J. Xu. 2019. Securing distributed gradient descent in high dimensional statistical learning. ACM SIGMETRICS
Performance Evaluation Review 47 (2019), 83–84.
[66] Y. Chen, F. Luo, T. Li, T. Xiang, Z. Liu, and J. Li. 2020. A training-integrity privacy-preserving federated learning
scheme with trusted execution environment. Elsevier Information Sciences 522 (2020), 69–79.
[67] J. Kindervag, S. Balaouras, K. Mak, and J. Blackborow. 2016. No More Chewy Centers: The Zero Trust Model of Informa-
tion Security, Technical Report, Forrester Research.
[68] B. Chen, S. Qiao, J. Zhao, D. Liu, X. Shi, M. Lyu, H. Chen, H. Lu, and Y. Zhai. 2021. A security awareness and protection
system for 5G smart healthcare based on zero-trust architecture. IEEE Internet of Things Journal 8 (2021), 10248–10263.
[69] B. Hitaj, G. Ateniese, and F. Perez-Cruz. 2017. Deep models under the GAN: Information leakage from collaborative
deep learning. In SIGSAC Conference on Computer and Communications Security (SIGSAC’17). 603–618.
[70] K. Bonawitz, V. Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth. 2017.
Practical secure aggregation for privacy-preserving machine learning. In SIGSAC Conference on Computer and Com-
munications Security (CCS’17). 1175–1191.
[71] Y. Dong, X. Chen, L. Shen, and D. Wang. 2019. Privacy-preserving distributed machine learning based on secret
sharing. In Information and Communications Security (ICICS’19). 684–702.
[72] L. Song, C. Ma, P. Wu, and Y. Zhang. 2019. PPD-DL: Privacy-preserving decentralized deep learning. In Artificial
Intelligence and Security (ICAIS’19). 273–282.
[73] C. Dwork. 2008. Differential privacy: A survey of results. Lecture Notes in Computer Science 4978 (2008), 1–19.
[74] W. Li, F. Milletarí, D. Xu, N. Rieke, J. Hancox, W. Zhu, M. Baust, Y. Cheng, S. Ourselin, M. J. Cardoso, and A. Feng.
2019. Privacy-preserving federated brain tumour segmentation. In Machine Learning in Medical Imaging (MLMI’19).
133–141.
[75] S. R. Pfohl, A. M. Dai, and K. Heller. 2019. Federated and differentially private learning for electronic health records.
[76] S. Truex, N. Baracaldo, A. Anwar, T. Steinke, H. Ludwig, R. Zhang, and Y. Zhou. 2019. A hybrid approach to privacy-
preserving federated learning. In 12th ACM Workshop on Artificial Intelligence and Security (AISec’19). 1–11.
[77] J. Zhang, J. Wang, Y. Zhao, and B. Chen. 2019. An efficient federated learning scheme with differential privacy in
mobile edge computing. In Machine Learning and Intelligent Communications (MLICOM’19). 538–550.
[78] J. Ma, Q. Zhang, J. Lou, J. C. Ho, L. Xiong, and X. Jiang. 2019. Privacy-preserving tensor factorization for collaborative
health data analysis. In 28th ACM International Conference on Information and Knowledge Management (CIKM’19).
1291–1300.
[79] A. Triastcyn and B. Faltings. 2019. Federated learning with Bayesian differential privacy. In IEEE International Confer-
ence on Big Data (Big Data’19). 2587–2596.
[80] X. Li, Y. Gu, N. Dvornek, L. H. Staib, P. Ventola, and J. S. Duncan. 2020. Multi-site fMRI analysis using privacy-
preserving federated learning and domain adaptation: Abide results. Medical Image Analysis 65 (2020), 101765.
[81] M. Gong, K. Pan, Y. Xie, A. Qin, and Z. Tang. 2020. Preserving differential privacy in deep neural networks with
relevance-based adaptive noise imposition. Elsevier Neural Networks 125 (2020), 131–141.
[82] M. Hao, H. Li, X. Luo, G. Xu, H. Yang, and S. Liu. 2019. Efficient and privacy-enhanced federated learning for industrial
artificial intelligence. IEEE Transactions on Industrial Informatics (2019), 1–11.
[83] Y. Dong, X. Chen, L. Shen, and D. Wang. 2020. EaSTFLy: Efficient and secure ternary federated learning. Elsevier
Computers and Security 94 (2020), 101824.
[84] J. Cui and D. Liu. 2019. Federated machine learning with anonymous random hybridization (FeARH) on medical
records. https://arxiv.org/abs/2001.09751. arXiv:2001.09751.
[85] R. Shao, H. He, H. Liu, and D. Liu. 2019. Stochastic channel-based federated learning for medical data privacy preserv-
ing. https://arxiv.org/abs/1910.11160. arXiv:1910.11160.
[86] S. Ge, F. Wu, C. Wu, T. Qi, Y. Huang, and X. Xie. 2020. FedNER: Privacy-preserving medical named entity recognition
with federated learning. https://arxiv.org/abs/2003.09288. arXiv:2003.09288.
[87] J. Jeon and J. Kim. 2020. Privacy-sensitive parallel split learning. In International Conference on Information Networking
(ICOIN’20). 7–9.
[88] N. Dowlin, R. Gilad-Bachrach, K. Laine, K. Lauter, M. Naehrig, and J. Wernsing. 2016. CryptoNets: Applying Neural
Networks to Encrypted Data with High Throughput and Accuracy. Technical Report, Microsoft Research.
[89] M. Abadi, A. Chu, I. Goodfellow, H. B. McMahan, I. Mironov, K. Talwar, and L. Zhang. 2016. Deep learning with
differential privacy. In 2016 ACM SIGSAC Conference on Computer and Communications Security (CCS’16). ACM, 2016,
308–318.
[90] J. Lee, J. Sun, F. Wang, S. Wang, C.-H. Jun, and X. Jiang. 2018. Privacy-preserving patient similarity learning in a
federated environment: Development and analysis. JMIR Medical Informatics 6 (2018), e20.
[91] L. Huang, A. L. Shea, H. Qian, A. Masurkar, H. Deng, and D. Liu. 2019. Patient clustering improves efficiency of
federated machine learning to predict mortality and hospital stay time using distributed electronic medical records.
Elsevier Journal of Biomedical Informatics 99 (2019), 103291.
[92] F. Fioretto and P. Van Hentenryck. 2019. Privacy-preserving federated data sharing. In 18th International Conference
on Autonomous Agents and MultiAgent Systems (AAMAS’19). 638–646.
[93] D. Liu, T. A. Miller, and K. D. Mandl. 2019. Confederated machine learning on horizontally and vertically separated
medical data for large-scale health system intelligence. https://arxiv.org/abs/1910.02109. arXiv:1910.02109.
[94] V. C. Pezoulas, K. D. Kourou, F. Kalatzis, T. P. Exarchos, E. Zampeli, S. Gandolfo, A. Goules, C. Baldini, F. Skopouli,
S. D. Vita, A. G. Tzioufas, and D. I. Fotiadis. 2020. Overcoming the barriers that obscure the interlinking and analysis
of clinical data through harmonization and incremental learning. IEEE Open Journal of Engineering in Medicine and
Biology 1 (2020), 83–90.
[95] B. Norgeot, B. S. Glicksberg, and A. J. Butte. 2019. A call for deep-learning healthcare. Nature Medicine 25 (2019),
14–15.
[96] A. E. Johnson, T. J. Pollard, L. Shen, H. L. Li-wei, M. Feng, M. Ghassemi, B. Moody, P. Szolovits, L. A. Celi, and R. G.
Mark. 2016. Mimic-III, a freely accessible critical care database. Scientific Data 3 (2016), 160035.
[97] T. J. Pollard, A. E. W. Johnson, J. D. Raffa, L. A. Celi, R. G. Mark, and O. Badawi. 2018. The eICU collaborative research
database, a freely available multi-center database for critical care research. Scientific Data 5 (2018), 180178.
[98] B. Liu, M. Ding, S. Shaham, W. Rahayu, F. Farokhi, and Z. Lin. 2021. When machine learning meets privacy: A survey
and outlook. ACM Computing Surveys 54 (2021).
Received March 2021; revised September 2021; accepted November 2021

Federated Learning for Healthcare- Systematic Review and Architecture Proposal

Uploaded by

Copyright:

Available Formats

Federated Learning for Healthcare- Systematic Review and Architecture Proposal

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Federated Learning for Healthcare- Systematic Review and Architecture Proposal

Uploaded by

Copyright:

Available Formats

Federated Learning for Healthcare: Systematic Review and

RODOLFO STOFFEL ANTUNES and CRISTIANO ANDRÉ DA COSTA, Universidade do Vale

CCS Concepts: • Applied computing → Health informatics; • Computer systems organization →

ACM Reference format:

Table 1. Databases and Keywords Used to Extract the

Fig. 1. Distribution of articles by publication year.

Fig. 2. Distribution by topics of articles in the literature corpus.

3.1 State-of-the-Art Evolution

3.1.3 Aggregation Methodologies. Model aggregation is a fundamental step on FL. In summary,

3.2 Machine Learning Methods and Applications

Table 2. Machine Learning Methods and Applications with EHR Data

Support Vector Machines

EHR Data Classification

Medical Image Analysis

on FL because high-quality datasets are available to conduct experiments in a simulated federated

Fig. 3. Literature taxonomy for FL applied to EHR.

4 ARCHITECTURE FOR FEDERATED ANALYSIS AND LEARNING WITH EHR DATA

Received March 2021; revised September 2021; accepted November 2021

You might also like