Source-linked AI summary
Federated Learning for Healthcare Domain - Pipeline, Applications and Challenges
Madhura Joshi, Ankit Pal, Malaikannan Sankarasubbu
TL;DR
Healthcare federated learning addresses the need to learn from distributed, sensitive medical data without explicitly centralizing it. This survey synthesizes healthcare federated-learning architectures, methods, challenges, and applications. It concludes by identifying implementation limitations and future research directions.
Problem
Healthcare data are distributed and sensitive, while single-source data can bias models and privacy constraints limit aggregation and access.
Method
The paper surveys healthcare federated-learning research, explains system components and architectures, examines challenges, and classifies applications across prognosis, diagnosis, and clinical workflow.
Results
The survey presents federated learning as a healthcare approach for using distributed data while preserving data privacy, and organizes its methods, applications, and challenges.
Takeaways & Limitations
The paper provides practitioners with an overview of healthcare federated-learning methods, applications, implementation tools, and unresolved issues.
Abstract
from arXiv · showhide
Federated learning is the process of developing machine learning models over datasets distributed across data centers such as hospitals, clinical research labs, and mobile devices while preventing data leakage. This survey examines previous research and studies on federated learning in the healthcare sector across a range of use cases and applications. Our survey shows what challenges, methods, and applications a practitioner should be aware of in the topic of federated learning. This paper aims to lay out existing research and list the possibilities of federated learning for healthcare industries.
1 INTRODUCTION
Healthcare data are valuable for machine learning but are distributed, sensitive, and subject to privacy constraints. This survey presents federated learning as an approach for studying healthcare applications without explicitly centralizing training data.
- Single-source healthcare data can bias predictions toward particular equipment, demographics, and acquisition protocols while increasing computational and time costs.
- Medical data are sensitive, distributed across stakeholders, and often cannot be aggregated or accessed freely because of privacy and regulatory requirements.
- Federated learning learns from data held by health-related institutions without explicitly centralizing or sharing the training data.
- The survey explains federated-learning components, communication architectures, and system building blocks for healthcare.
- It examines privacy, data, and communication challenges and classifies healthcare applications into prognosis, diagnosis, and clinical workflow.
2.1 Architecture of Communication
Federated learning can use either a centralized server to coordinate client updates or direct client-to-client communication. The centralized design is common but introduces a server bottleneck, whereas decentralization avoids a single point of failure and requires mutual trust.
- Centralized: In centralized federated learning, a server aggregates local models, coordinates clients, and sends updated models back to participating devices.
- Centralized: Centralized communication makes the server a system bottleneck and creates a single-point-of-failure drawback.
- Decentralized: Decentralized federated learning lets clients communicate directly, update the global model without a central server, and avoid single-point failure.
- Decentralized: The decentralized process updates local gradients, sends them to selected parties, and propagates updated models among participants.
- Decentralized: Decentralized learning relies on mutual trust between users, creating a challenge in single-sided-trust environments.
2.2 Scale of Federated Learning system
Federated learning systems differ by scale: cross-silo systems connect a few data-rich organizations, while cross-device systems connect many clients with smaller datasets and unreliable device conditions.
- Cross-Silo: Cross-silo federated learning connects a small number of organizations or data centers that contribute large amounts of data and computational power.
- Cross-Silo: Cross-silo settings address cases where organizations cannot share data because of privacy and security restrictions.
- Cross-Device: Cross-device federated learning uses scalable numbers of clients with smaller amounts of data, commonly on mobile or IoT devices.
- Cross-Device: Cross-device systems must manage many devices and common failures such as unstable connections and power loss.
2.3 Federated Server
Federated servers coordinate client participation, model-update exchange, and aggregation across training rounds. The section surveys aggregation methods and evaluates federated systems using performance, communication, time, privacy, and robustness considerations.
- A cross-device server manages communication among clients and the global model, with authentication required for clients to join training sessions.
- In decentralized training, devices interact directly and contribute equally, but designing the system with suitable communication overhead is difficult.
- Clients train locally and share model updates rather than raw data, while communication exchanges parameters and computation performs training and aggregation.
- Each federated round aggregates participating devices’ updates into a global model, and subsequent rounds continue until the server’s maximum is reached.
- FedAvg combines client-side stochastic-gradient updates through server model averaging and has been studied across five architectures and four dataset types.FedMA instead builds shared neural-network models layer by layer by matching and averaging comparable hidden components.
- Federated evaluation considers training speed, model performance, transferred-data quantity, accuracy, communication, time consumption, privacy, and robustness.
2.4 Federated Client
Federated learning clients train on private local data and share model updates for aggregation, with horizontal, vertical, and transfer-learning arrangements supporting different data distributions and healthcare applications.
- Federated client architecture: Clients train local models on private data and send updates to a federated server for aggregation; decentralized systems allow clients to communicate directly.The server aggregates updates into a global model across training rounds.
- Horizontal federated learning: Horizontal federated learning uses shared features but different samples, with parties exchanging model updates to create a global model.FedAvg is presented as a typical horizontal federated learning algorithm.
- Vertical federated learning: Vertical federated learning uses different features but similar samples, requiring synchronized datasets and a specialized training architecture.Parties cannot independently train on the full feature set and labels.
- Privacy-preserving participation: Federated learning avoids direct disclosure of client input data or model outputs while supporting participation from medical institutions and other devices.The approach is described as keeping healthcare datasets private during local learning.
- Healthcare applications: Surveyed healthcare applications include federated transfer learning, privacy-preserving fMRI classification, EEG classification, breast-density categorization, and multi-institutional training.Other reviewed systems address aggregation, communication costs, untrustworthy conditions, and emotion monitoring.
- Reported findings: Reported studies include equivalent performance to centralized models on vertically partitioned data and 99 percent model quality using data from ten universities.The survey also reports FL models trained across 58 hospitals and external validation for COVID-19 CT anomaly identification.
2.5 Machine Learning Pipeline
The federated learning pipeline moves from task initialization and model selection through training, communication, optimization, testing, deployment, and termination. Model choice depends on the federated task and data, while training repeatedly exchanges and aggregates local model parameters.
- The federated learning system comprises eight steps: task initialization, selection, configuration, model training, client-server communication, scheduling and optimization, versioning, testing, deployment, and termination.
- Federated learning uses three major machine learning stages: model selection, model training, and hyperparameter tuning.
- Model selection: Model selection chooses an optimal model according to the task and the size, quality, and type of federated data.
- Model training: Training repeatedly fits local models, shares their parameters with a server, aggregates them, updates local models with the global model, and repeats the loop.
- Hyperparameter tuning: Hyperparameters such as learning rate, training epochs, mini-batch size, and optimizer are tuned according to application constraints including computing power, memory, and bandwidth.
3 CHALLENGES & ISSUES
The survey groups federated-learning challenges into privacy and attacks, data limitations and bias, and communication constraints. Healthcare deployments must also handle heterogeneous clients, noisy or improperly annotated data, energy use, and unreliable participation.
- The survey classifies challenges into privacy and system attacks, data limitations including bias and poisoning, and communication issues.
- Communication challenges: Communication must reduce rounds and resource use while retaining participating devices and preserving system effectiveness.
- Privacy and attacks: Distribution shifts between previous patient data and new patients can undermine prediction validity and enable adversarial examples.
- Data challenges: Healthcare data collection is affected by instrumental and environmental noise in clinical records and medical images, including MRI artifacts.
- Data challenges: Improper annotation can produce imbalanced datasets, class imbalance, bias, and data sparsity, while data biases can drive unfairness in federated models.
- Reliability challenges: Federated clients are heterogeneous in dataset size and distribution, and may be unreliable during communication.
- Resource challenges: On-device training can consume substantial battery energy and deter users from participating.
4 APPLICATION
Healthcare federated learning applications address distributed, sensitive, and heterogeneous data across prognosis, diagnosis, and clinical workflows. Surveyed studies use federated approaches to collaborate without directly exchanging raw data.
- Prognosis: ENIGMA Shape provides federated analysis of biomedical data across clients without directly exchanging data, consistent with ENIGMA’s standard pipelines.The framework was demonstrated as an implementation for safely accessing and meta-analyzing biomedical data while maintaining individual privacy.
- Prognosis: Healthcare FL applications address data-sharing barriers in neuroimaging, adverse-drug-reaction prediction, and disease-related biomarker discovery.These studies use distributed data while avoiding raw-data transfer or centralized aggregation.
- Prognosis: An FL study involving 20 institutes used vital signs, laboratory results, and chest X-rays to predict oxygen needs during the SARS-COV-2 pandemic.The resulting model was called EXAM, or EMR CXR AI Model.
- Diagnosis and clinical workflow: Federated approaches are applied to privacy-sensitive clinical tasks including continuous blood-pressure estimation, medical imaging, clinical decision support, and genomic prediction.The surveyed examples include decentralized ABP calculation, federated imaging systems, and privacy-aware drug-sensitivity prediction.
- Clinical workflow: Cross-institutional collaboration can enlarge datasets for genome-wide studies, but exchanging personal information creates privacy risks.The survey links these applications to concerns including discrimination, denial of benefits, and higher insurance premiums after leaks.
5 TOOLS
The survey reviews federated-learning tools including TensorFlow Federated and FATE, emphasizing their programming interfaces, supported algorithms, and deployment limitations.
- TensorFlow Federated: TensorFlow Federated provides localized-information computation through Federated Core and Federated Learning APIs.Federated Core is the foundation layer, while the higher-level Federated Learning API supports existing training and evaluation algorithms.
- TensorFlow Federated: TFF supports only FedAvg, provides no privacy mechanisms, and can be deployed only on a single computer through simulation.These constraints limit its direct use for privacy-preserving, distributed production deployments.
- FATE: FATE is an open-source secure-computing framework supporting homomorphic encryption, secure multi-party computation, and several machine-learning model families.Its supported models include logistic regression, deep learning, tree-based algorithms, and transfer learning.
- FATE: FATE requires practitioners to modify source code to incorporate federated algorithms because its interfaces operate at the algorithm level.The survey identifies this as difficult for non-expert users.
5.3 PySyft
PySyft is presented as a privacy-preserving deep-learning library, while LEAF provides federated-learning benchmarks and dataset utilities; both have documented scope limitations.
- PySyft: PySyft combines differential privacy, homomorphic encryption, and multi-party computation with PyTorch and TensorFlow.It is designed for answering questions using data that users cannot see.
- PySyft: PySyft can run on one or several computers and uses a WebSocket API for client communication.The framework supports both PyTorch and TensorFlow.
- PySyft: PySyft lacks comprehensive documentation for its interfaces and architecture and does not support on-device training on mobile or IoT.The survey notes that tutorials exist, but system-wide documentation is incomplete.
- LEAF: LEAF benchmarks federated learning, multi-task learning, meta-learning, and on-device learning across six databases.Its utilities split datasets into IID or non-IID parties and provide reference training implementations.
- LEAF: LEAF has enough clients for cross-device scenarios but may be too limited for questions where dataset size is especially relevant.It supports standardized algorithms such as FedAvg, not decentralized, split, or vertical federated learning.
5.5 Paddle FL
PaddleFL is an open-source framework for large-scale federated applications, while Clara Train SDK offers domain-optimized APIs for clinical AI and federated training.
- Paddle FL: PaddleFL supports federated applications in natural language processing, computer vision, and recommendation systems.It is designed to deploy federated-learning systems as distributed clusters at large scale.
- Paddle FL: PaddleFL includes federated methods, user-specified models and algorithms, distributed training setup, and an FL task generator.Users can create their own models and training algorithms in addition to the offered techniques.
- Paddle FL: PaddleFL remains early in production, with limited documentation and samples, weaker standardized-benchmark performance, and no vertical federated learning support.The survey also reports limited API flexibility for topology customization and message flow.
- Clara Train SDK: Clara Train SDK provides domain-optimized APIs for AI-assisted annotation and supports transfer learning, federated learning, and AutoML.Its TensorFlow-based framework includes pre-skilled models for clinical AI improvement.
- Clara Train SDK: Clara Train SDK’s configurable MMAR function packages models and components for federated learning and monitors local training across one or multiple GPUs.The MMAR function supports model and component transfer during federated workflows.
5.7 Fed ML
This section presents federated learning tools that support different deployment settings, algorithms, models, and privacy requirements. FedML emphasizes simulation and distributed experimentation, while IBM Federated Learning targets enterprise adoption.
- FedML: FedML provides an FL system and benchmarking forum supporting on-device training, distributed computing, and single-machine emulation.It also includes algorithms, prototypes, and databases for decentralized, vertical, and split learning.
- FedML: FedML supports standalone simulation, distributed computing, and on-device training across varied hardware specifications.
- IBM Federated Learning: IBM Federated Learning targets enterprise environments where safe rollout, failure tolerance, and rapid model specification are critical.It uses existing machine learning libraries so enterprise users can access state-of-the-art algorithms without learning new languages.
- IBM Federated Learning: IBM Federated Learning supports classification, regression, linear classifiers, XGBoost adaptation, and differential privacy across models.Differential privacy spans models from Naive Bayes to neural-network structures.
6 CONCLUSION & FUTURE DIRECTIONS
Federated learning has become increasingly relevant to healthcare, motivating a comprehensive survey of its systems, methods, applications, and issues. The paper concludes that important technical concerns remain, while future work targets FL-enabled 5G/6G medical devices and services.
- Conclusion: Federated learning, originally developed for mobile and edge-device use cases, has recently gained popularity in healthcare applications.
- Conclusion: The healthcare field requires a comprehensive overview of existing federated learning systems because research interest has grown across industry and academia.
- Future Directions: Not all federated learning technological concerns have been solved, leaving FL as an ongoing research focus.
- Future Directions: Future research will integrate FL into 5G/6G medical devices, including intelligent implants and wearables, and investigate new healthcare services enabled by 6G.The passage specifically frames large-scale FL-based healthcare as a future direction.
- Conclusion: The survey covers healthcare federated learning methods, applications, implementation approaches, and associated issues.