Source-linked AI summary
Federated Learning for Smart Healthcare: A Survey
Dinh C. Nguyen, Quoc-Viet Pham, Pubudu N. Pathirana, Ming Ding, Aruna Seneviratne, Zihuai Lin, Octavia A. Dobre, Won-Joo Hwang
TL;DR
Smart healthcare’s centralized AI systems struggle with scalability, privacy, distributed data, and limited local datasets. This survey synthesizes federated-learning designs, applications, projects, and challenges, concluding that FL is still nascent but expected to support large-scale, privacy-aware healthcare operations.
Problem
Centralized healthcare AI faces scalability, privacy, distributed-data, and limited-dataset challenges, while comprehensive coverage of FL applications and taxonomy remains absent.
Method
The paper provides a comprehensive survey of FL motivations, requirements, advanced designs, healthcare applications, real-world projects, lessons, and future directions.
Results
The survey covers federated EHR management, remote health monitoring, medical imaging, and COVID-19 detection, and analyzes recent FL-healthcare projects.
Takeaways & Limitations
FL is expected to support large-scale, collaborative, privacy-aware healthcare operations, although its smart-healthcare application remains in its infancy.
Takeaways & Limitations
Non-IID medical datasets can degrade or destabilize FL training because label distributions differ across healthcare institutions.
Abstract
from arXiv · showhide
Recent advances in communication technologies and Internet-of-Medical-Things have transformed smart healthcare enabled by artificial intelligence (AI). Traditionally, AI techniques require centralized data collection and processing that may be infeasible in realistic healthcare scenarios due to the high scalability of modern healthcare networks and growing data privacy concerns. Federated Learning (FL), as an emerging distributed collaborative AI paradigm, is particularly attractive for smart healthcare, by coordinating multiple clients (e.g., hospitals) to perform AI training without sharing raw data. Accordingly, we provide a comprehensive survey on the use of FL in smart healthcare. First, we present the recent advances in FL, the motivations, and the requirements of using FL in smart healthcare. The recent FL designs for smart healthcare are then discussed, ranging from resource-aware FL, secure and privacy-aware FL to incentive FL and personalized FL. Subsequently, we provide a state-of-the-art review on the emerging applications of FL in key healthcare domains, including health data management, remote health monitoring, medical imaging, and COVID-19 detection. Several recent FL-based smart healthcare projects are analyzed, and the key lessons learned from the survey are also highlighted. Finally, we discuss interesting research challenges and possible directions for future FL research in smart healthcare.
1 INTRODUCTION
Smart healthcare is moving from centralized AI toward federated learning because growing, distributed health data create scalability and privacy challenges. The survey addresses this gap by reviewing FL principles, designs, applications, projects, lessons, and future directions.
- Motivation: Centralized healthcare AI faces communication latency, poor scalability, privacy risks, and distributed data across large IoMT networks.These limitations motivate distributed, privacy-preserving intelligence at the network edge.
- Federated learning: Federated learning trains models by aggregating local updates from multiple healthcare clients without directly accessing their raw data.The approach is presented as a promising solution for privacy protection and cost-effective smart healthcare.
- Research gap: Existing surveys do not comprehensively cover FL applications in smart healthcare or provide a holistic taxonomy of emerging healthcare uses.Earlier work addresses general FL, security, mobile edge integration, architectures, or digital-health requirements, but leaves newer FL designs insufficiently explored.
- Survey scope: The survey reviews FL motivations and requirements, advanced designs, healthcare applications, real-world projects, lessons learned, and future research directions.Its application coverage includes health data management, remote monitoring, medical imaging, and COVID-19 detection.
2 FL FOR HEALTHCARE: KEY PRINCIPLE AND CATEGORIES
FL-smart healthcare uses an iterative process in which selected clients train locally, upload updates, and receive an aggregated global model. The survey categorizes healthcare FL according to participating clients’ feature and sample spaces.
- Key principle: The FL process initializes a healthcare task and learning parameters, selects participating clients, and distributes an initial model.Client selection determines which healthcare participants join the federated training process.
- Key principle: During each communication round, clients train local models on their own datasets, calculate updates, and upload those updates to the server.The server receives model updates such as neural-network gradients rather than raw healthcare data.
- Key principle: The server aggregates client updates, for example with FedAvg, then broadcasts the updated global model for subsequent rounds.FedAvg averages local gradient parameters element-wise using weights proportional to client dataset sizes.
- Categories: Horizontal FL combines clients with the same feature space but different sample spaces using a shared AI model and aggregated updates.This category supports collaboration without direct access to local data.
- Categories: Federated transfer learning addresses clients with different sample and feature spaces by mapping features into a shared representation.Encryption techniques such as random masks can further protect gradients during exchange.
3 MOTIVATIONS AND REQUIREMENTS OF USING FL IN SMART HEALTHCARE
The survey frames FL adoption in smart healthcare around its motivations and technical requirements. It first explains why existing systems are insufficient, then details the requirements for effective FL use.
- Section scope: The section identifies limitations of current healthcare systems before presenting FL’s motivations and benefits.This establishes the rationale for examining federated approaches in smart healthcare.
- Section scope: The section describes the technical requirements associated with using FL in smart healthcare.These requirements complement the motivation analysis by addressing practical deployment needs.
- Section scope: The discussion connects healthcare-system limitations with FL-based solutions and their deployment requirements.The section is organized to explain both why FL is needed and what its use entails.
3.1 Motivations
Smart healthcare needs approaches that address centralized systems’ privacy, data-access, and training-performance limitations. FL keeps health data local while supporting collaborative training across medical sites.
- Motivations: Centralized healthcare AI can expose sensitive information to unauthorized access, cloud providers, or data modification without user consent.These privacy concerns arise from sharing health data with clouds or data centers.
- Motivations: Data exchange between medical sites is difficult because institutional policies and growing privacy concerns restrict access to other sites’ datasets.This creates a dataset-shortage challenge for smart healthcare AI training.
- Motivations: Limited datasets at a single medical site can reduce training accuracy because of imbalanced features and insufficient data sizes.The passage identifies data augmentation as one possible response to this limitation.
- Benefits: FL keeps local health data at medical sites and devices while sending local updates such as model gradients to the central server.This reduces exposure of sensitive user information during collaborative training.
- Benefits: FL offers a trade-off among accuracy, utility, privacy enhancement, and scalability by retaining model generalizability with nominal accuracy loss.Its distributed learning feature also helps reduce communication costs associated with moving large raw datasets.
3.2 Requirements
Reliable FL-based smart healthcare requires trusted coordination, secure client-server communication, capable participating devices, available client datasets, and attention to operational metrics. Privacy remains incomplete because model updates may still reveal health-related information.
- Privacy Protection: FL cannot fully address privacy in smart healthcare, so dedicated privacy-protection mechanisms are needed to enhance it.This limitation follows the risk that model updates may still expose health-user information.
- Trusted Server: A trusted server must coordinate data training and aggregate local model gradients transparently and reliably under agreements with healthcare organizations.Although clients keep data locally, model updates can still contain reconstructable health-related information.
- Reliable Client-Server Communications: Secure client-server communication is required because adversaries may steal or modify local updates, disrupting training or biasing aggregation.These protections also support user trust in collaborative healthcare tasks such as federated medical image analysis.
- Resource Availability: Limited computation and energy resources can prevent lightweight medical devices from participating across the multiple communication rounds needed for desired training performance.Smart watches are given as an example of devices that may be unable to join training in the long run.
- Data Availability: Participating clients need locally constructed datasets and extracted data features to conduct their own training based on their working environments.For example, smartphones may collect human motion data within their working areas.
- Operational Metrics: Efficient FL design should evaluate model size, local-data size, convergence time, and required accuracy across varying device, network, and computation conditions.These metrics depend on both clients and the server.
4 ADVANCED FL DESIGNS FOR SMART HEALTHCARE
Advanced FL designs for smart healthcare address resource management, security and privacy threats, incentives, and personalization across heterogeneous healthcare settings. The survey also highlights accuracy, privacy, communication, dataset, and reliability tradeoffs that constrain practical deployment.
- Resource-aware FL: Resource-aware FL optimizes device scheduling and joint computing–communication resources to improve training efficiency across healthcare services.Multi-armed bandits address scheduling under massive devices and unknown channel states, while other work jointly allocates computation and radio resources.
- Secure FL: Secure FL counters poisoning, inference, backdoor, malicious-server, communication, and free-riding attacks through reliable device selection, decentralized aggregation, and blockchain.Blockchain can manage user reputations and coordinate global-model calculation through peer-to-peer block consensus, reducing reliance on a single aggregation server.
- Secure FL: Turbo-Aggregate reduces aggregation overhead from O(N^2) to O(Nlog(N)), tolerates up to 50% dropping-out, and runs up to 40 times faster than benchmarks.These advantages were reported from numerical experiments, while validation on real health datasets remains an open need.
- Privacy-enhanced and incentive-aware FL: Privacy-enhanced and incentive-aware FL combine differential privacy, cost-aware contracts, and game-theoretic participation mechanisms to balance protection, contribution, and learning performance.Differential privacy accounts for computation, communication, and privacy costs, while Stackelberg games model offers, participation decisions, and server utility.
- Personalized FL: Personalized FL addresses global-model limitations caused by individual differences, heterogeneous data, insufficient personalized data, and inconsistent labels across devices.An α-weighted update increased classification accuracy by about 16.7% on Animals-10 and by 9.153% and 11.01% on two human activity-recognition datasets.
- Lessons learned: The survey identifies healthcare-specific deployment constraints, including unreliable user-selection metrics, privacy–accuracy tradeoffs, non-healthcare evaluation datasets, and costly communication.Bandwidth-efficient and compressed FL methods are therefore needed for large health data and model updates.
5 FL APPLICATIONS IN HEALTHCARE
FL applications in smart healthcare span EHR management, remote monitoring, medical imaging, and COVID-19 detection. Across these domains, studies use distributed training to preserve privacy while addressing heterogeneous data and achieving competitive or improved predictive performance.
- Federated Remote Health Monitoring: Remote monitoring systems train global or personalized models across homes and wearable devices without sharing users’ local data.One human-activity approach addressed imbalanced and non-IID data through class-balanced local datasets and synchronized updates.
- Federated Remote Health Monitoring: 95.41% accuracy, an increase of 7.49% over standalone CNN, was achieved for human-activity recognition with low communication costs in balanced and imbalanced settings.The approach used CNN-based personalized models and class-balanced data synthesis.
- Federated Medical Imaging: Medical-imaging FL aligns data across institutions and uses local GANs to generate or translate images while preserving privacy.A prostate-cancer imaging scheme achieved an accuracy score of 0.9722, improving performance by 0.13% over non-FL schemes.
- Federated COVID-19 Detection and Diagnosis: COVID-19 detection uses federated CNN and GAN models that synchronize local parameters through a cloud aggregation server.ResNet18 achieved 98.06% detection performance in federated X-ray image learning settings.
- Federated EHRs Management: FL supports EHR analytics across multiple hospitals while keeping raw health data local and exchanging model updates.Differential privacy can mask local updates, with competitive accuracy compared to traditional FL schemes.
6 REAL-WORLD PROJECTS OF FL IMPLEMENTATION IN SMART HEALTHCARE
The survey reviews real-world FL projects for collaborative medical imaging and UK healthcare deployment. These projects combine institutional collaboration with infrastructure for secure, traceable, or scalable federated training.
- Healthcare Collaboration: A University of Pennsylvania project involving 19 other institutions investigated FL for collaborative medical imaging.Intel supported hospital and cloud deployment using Xeon Scalable processors and Intel SGX.
- Healthcare Collaboration, UK: The UK NHS FL platform trains AI algorithms at local hospitals under central-server management using Owkin Connect and NVIDIA Clara.Blockchain provides traceability and monitoring for healthcare data used by the platform.
- Medical Imaging: Preliminary results indicated improved classification accuracy for FL-based medical image analytics in imaging-related disease diagnosis.The survey presents this as an outcome of the collaborative UK experiment.
- COVID-19 Collaboration: A multinational COVID-19 chest-CT segmentation project involved institutions from China, Italy, and Japan and used 1,704 scans.The project illustrates FL deployment across geographically distributed medical institutions.
7 RESEARCH CHALLENGES AND FUTURE DIRECTIONS
Future FL-healthcare research must improve communication efficiency and establish standardized evaluation practices. These needs become more pressing as many IoMT devices exchange updates with aggregation servers.
- Communication Efficiency: Communication-resource allocation and efficient scheduling are needed to support uplink updates and downlink broadcasts for many IoMT devices.Communication directly affects FL learning performance in healthcare services.
- Standardized Evaluation: The field lacks a standard and universal evaluation framework for comparing different FL-healthcare approaches on the same problem.Different blockchain-based FL approaches are difficult to compare because they target different healthcare settings.
7.3 Quality of Federated Healthcare Training Data
Federated healthcare training quality is constrained by differences in computational capabilities and data quality across hospital sites. Incentive mechanisms are proposed as a way to encourage higher-quality data and reliable updates.
- Training Data Quality: Heterogeneous computational capabilities and data qualities across hospitals can greatly degrade training quality.The challenge concerns both the resources available at participating sites and the quality of their training data.
- Incentive Mechanisms: Incentive mechanisms can motivate healthcare organizations to provide high-quality data and reliable updates to the aggregation server.Game theory and blockchain are identified as important tools for designing these mechanisms.
7.4 Health Dataset Issues for Robust FL-based Health Data Analytics
Robust federated healthcare analytics must accommodate heterogeneous client datasets and evolving wireless environments. The survey highlights diverse data modalities and the need for approaches that handle heterogeneous models and future 6G requirements.
- Health Dataset Heterogeneity: Healthcare clients may hold heterogeneous datasets spanning text, images, audio, and time series, with differing medical attributes.Examples include blood type, heart rate, face images, and body temperature.
- Evaluation Scope: Most existing healthcare FL approaches are evaluated on a single dataset with limited features, despite realistic cross-client diversity.The survey contrasts diabetic retinopathy and EHR evaluations as examples of this narrow assessment practice.
- Heterogeneous Models: Heterogeneous FL should support clients using different models through private ensemble learning without centralizing their data.The cited inference strategy lets participants operate an ensemble of heterogeneous models without explicitly joining data at one location.
- Future Network Requirements: Future 6G healthcare systems must address stricter wireless requirements while integrating technologies such as blockchain, compressive sensing, and THz communications.The passage notes that 6G-related healthcare applications remain prospective because 5G is not yet fully available and commercially deployed worldwide.
7.6 FL with Provable Privacy Guarantee
FL can improve privacy protection in smart healthcare, but sensitive health data remains exposed to multiple privacy threats. The survey identifies attack categories that require dedicated safeguards.
- Privacy Risks: Smart healthcare FL still faces privacy issues because health-related data are highly sensitive.The survey frames privacy protection as an unresolved requirement rather than a fully achieved property of FL.
- Threat Categories: The identified privacy threats include membership inference attacks, unintentional information leakage, and generative adversarial network attacks.These categories describe ways adversaries may exploit FL systems or their outputs.
- Membership Inference: An attacker may misuse the global FL model to check whether a data sample is present in the training data.This illustrates how model access can expose information about individual health records.
7.7 Security Issues in FL-based Smart Healthcare
FL-based smart healthcare systems must defend against malicious clients, transmission tampering, and attacks on aggregated models. They must also address non-iid medical data, which can degrade or destabilize training.
- Client-Side Security: Malicious clients can send poisonous updates or fake information to degrade federated model aggregation.Attackers may also contaminate feature information during local training.
- Transmission and Server Threats: Adversaries may modify local updates during transmission or steal information from the aggregated global model.The survey links these attacks to data modification and information leakage concerns.
- Security Measures: Differential privacy and secure aggregation are proposed to protect training data and mask local updates against attacks.Secure aggregation uses double masking and key sharing among clients and the central server.
- Non-iidness: Non-iid medical datasets can cause training quality degradation or divergence when label distributions differ across hospitals.Geographic differences in disease prevalence can produce unequal local label distributions.
- Mitigation Strategies: Proposed responses include fair dataset allocation, local batch normalization for feature shifts, and quantitative metrics for assessing non-iid data.These methods target unequal data distributions and heterogeneous client features before or during model aggregation.
8 CONCLUSIONS
This survey reviews federated learning for smart healthcare, covering its foundations, designs, applications, projects, lessons, challenges, and future directions. It concludes that healthcare FL remains early-stage but is expected to mature toward intelligent, privacy-enhanced, large-scale collaboration.
- Survey Scope: The survey explains FL concepts, motivations, and technical requirements for applying collaborative AI in smart healthcare.It positions this work as a comprehensive survey addressing a gap in the open literature.
- Coverage: It reviews FL designs, healthcare applications, real-world projects, lessons learned, challenges, and future research directions.Applications include federated EHR management, remote health monitoring, medical imaging, and COVID-19 detection.
- Conclusion: FL-based smart healthcare is still at its infancy but is expected to mature for intelligent and privacy-enhanced health services.The conclusion characterizes the field as emerging rather than fully established.
- Future Outlook: FL is expected to support large-scale collaborative healthcare and a shift from centralized health-data analytics toward distributed healthcare operations.The stated future direction emphasizes privacy awareness in distributed operations.