Source-linked AI summary
The Future of Digital Health with Federated Learning
Nicola Rieke, Jonny Hancox, Wenqi Li, Fausto Milletari, Holger Roth, Shadi Albarqouni, Spyridon Bakas, Mathieu N. Galtier, Bennett Landman, Klaus Maier-Hein, Sebastien Ourselin, Micah Sheller, Ronald M. Summers, Andrew Trask, Daguang Xu, Maximilian Baust, M. Jorge Cardoso
TL;DR
Medical data is siloed, sensitive, and difficult to share, limiting its use for machine learning. This paper examines how Federated Learning could support collaborative digital health while addressing implementation challenges, finding that it can perform comparably to centralized training and better than isolated institutional models.
Problem
Medical machine learning lacks comprehensive, representative datasets because healthcare data is siloed, sensitive, regulated, and difficult to share safely.
Method
The paper synthesizes Federated Learning’s benefits, impact, and key implementation challenges for medical applications and digital health.
Results
Federated Learning can achieve performance comparable to co-located data and superior to models trained on isolated single-institution data.
Takeaways & Limitations
Federated Learning may enable collaborative healthcare modeling without exchanging or centralizing sensitive medical datasets, with potential benefits for digital health and patient care.
Takeaways & Limitations
Medical data heterogeneity challenges Federated Learning because common strategies such as FedAvg are prone to fail when participant data is non-IID.
Abstract
from arXiv · showhide
Data-driven Machine Learning has emerged as a promising approach for building accurate and robust statistical models from medical data, which is collected in huge volumes by modern healthcare systems. Existing medical data is not fully exploited by ML primarily because it sits in data silos and privacy concerns restrict access to this data. However, without access to sufficient data, ML will be prevented from reaching its full potential and, ultimately, from making the transition from research to clinical practice. This paper considers key factors contributing to this issue, explores how Federated Learning (FL) may provide a solution for the future of digital health and highlights the challenges and considerations that need to be addressed.
1 Introduction
Digital healthcare has benefited from powerful machine learning models, but sensitive, regulated health data remains difficult to obtain, curate, and share. Federated Learning addresses these barriers by enabling collaborative training without exchanging underlying datasets, with potential to support large-scale precision medicine while requiring careful technical consideration.
- Deep Learning models have driven disruptive innovations across radiology, pathology, genomics, and other digital-health fields.Modern models can contain millions of parameters.
- Supervised medical tools require large annotated datasets spanning diverse anatomies, pathological patterns, and input types, but health data is sensitive, regulated, and subject to consent and ethical-approval requirements.These constraints make medical data difficult to obtain, curate, and use for research.
- Federated Learning enables collaborative algorithm training without exchanging underlying datasets, addressing key data-governance and privacy barriers.After training, a consensus model benefits from knowledge accumulated across institutions.
- Federated Learning can achieve performance comparable to training on co-located data and superior to models trained on isolated single-institution datasets.
- Successful Federated Learning implementation could enable large-scale precision medicine, supporting models that reflect individual physiology, reduce bias, and detect rare diseases while respecting governance and privacy.The paper examines these benefits alongside the technical considerations and implementation challenges specific to medical applications.
2 Data-driven Medicine Requires Federated Efforts
Data-driven medicine requires large, diverse datasets, but isolated medical data creates representational bias and centralisation raises privacy, legal, regulatory, and technical barriers. Federated Learning enables collaborative training and validation without moving data, potentially improving access, clinical quality, and stakeholder control while requiring careful implementation.
- Data limitations: Large, single-source or isolated datasets can produce demographic and technical imbalances that reduce prediction accuracy for particular groups or sites.Examples include skew from gender, age, acquisition protocol, and equipment manufacturer differences.
- Data limitations: Centralising or releasing healthcare data creates ethical, privacy, legal, regulatory, and technical challenges, including reidentification risks despite anonymisation.Healthcare data also requires safe anonymisation, access control, and transfer, which may be non-trivial or impossible.
- Federated solution: Federated Learning provides controlled, indirect access to large and comprehensive datasets by keeping data with its controllers while supporting collaborative training and validation.Data controllers retain governance and access control, including the possibility of revoking access; federated validation could support large-scale studies across institutions.
- Federated solution: Federated participants generally receive aggregated model parameters rather than other institutions’ data, but the models themselves may still create privacy concerns.In client-server architectures, a federated server can manage aggregation and distribution while participating institutions remain unknown to one another.
- Healthcare stakeholders: Federated efforts could extend high-quality ML-supported decisions to remote and rare-disease patients while preserving hospital data control and expanding researchers’ access to real-world data.Hospitals retain possession and traceability, whereas researchers—especially smaller labs and start-ups—could develop applications without curating their own datasets.
3 Technical Considerations
Federated Learning combines knowledge from non-co-located data held by participating entities into a global model through varied communication architectures and compute plans. Applying FL in medicine introduces distinct requirements, including differences in participant scale, computing infrastructure, connectivity, and data types.
- FL combines knowledge learned from non-co-located data residing within participating entities into a global model.
- FL approaches can be realised through different communication architectures and corresponding compute plans.These architectures and plans are illustrated in Figures 1 and 2.
- Medical FL settings differ from mobile-device federations in participant scale, institutional computing infrastructure, and connectivity stability.Mobile applications could involve potentially millions of participants, whereas hospitals cannot form consortia at the same scale; medical institutions may have more powerful infrastructure and stable connectivity.
- Medical FL must also account for variation in data type and definition.
3.1 Federated Learning Definition
Federated Learning is a collaborative learning paradigm in which parties train a shared model without exchanging or centralising their datasets. Each party retains private data locally while contributing through local optimisation and parameter sharing.
- 3.1 Federated Learning Definition: FL enables multiple parties to train collaboratively without exchanging or centralising datasets.It uses a global loss formed from a weighted combination of local losses computed on each party’s private data.
- 3.1 Federated Learning Definition: Private data X_k remains unshared among parties throughout learning.The local loss weights satisfy w_k > 0.
- 3.1 Federated Learning Definition: Participants refine the global consensus model through local optimisation rounds before sharing updated parameters directly or via a parameter server.More local-training rounds without sharing updates or synchronisation provide less guarantee that the procedure minimises the global objective.
3.2 Relation to Similar Strategies
Federated Learning builds on collaborative model-sharing and distributed-compute strategies, including transfer learning, ensembling, multi-task learning, and parallel training. These approaches can complement FL, while distributed computation enables substantially faster training.
- 3.2 Relation to Similar Strategies: Federated Learning is rooted in collaborative learning approaches that share models or distribute computation.Transfer Learning exemplifies model-sharing, while large-scale model training increasingly uses multiple devices and nodes.
- 3.2 Relation to Similar Strategies: Transfer Learning trains a model on a large dataset before optimising it on target data, often outperforming training from scratch when the target dataset is small.The initial and target datasets need not come from the same domain or data-source type, and the data need not be co-located.
- 3.2 Relation to Similar Strategies: Ensembling combines independently trained models or predictions, while multi-task learning leverages shared representations for related tasks; both can be combined with FL.These collaborative learning strategies are independent of the concept of FL.
- 3.2 Relation to Similar Strategies: Parallelising training across multiple devices or nodes enables fast computation, including ImageNet neural-network training in 1 hour or less than 80 seconds.Distributed computation has been studied extensively and supports training on large-scale models.
3.3 Challenges and Considerations
Federated learning offers privacy-preserving collaboration for healthcare, but deployment must address privacy risks, trust, information leakage, heterogeneous data, reproducibility, interpretability, and governance. These challenges coexist with an opportunity to integrate knowledge across institutions and capture greater data variability across demographics.
- Privacy Vs. Performance: FL does not eliminate privacy risks, and healthcare regulations may differ across federations, requiring context-specific mitigation strategies.Privacy-preserving techniques can provide stronger protection, but leakage or perceived leakage may remain unacceptable under strict governance policies.
- Level of Trust: Non-trusted federations require security strategies because participants may degrade performance, disrupt systems, or extract information.Potential measures include encrypting model submissions, authenticating parties securely, and tracing actions.
- Information leakage: Shared model updates can indirectly expose private training data through model inversion, gradient analysis, or adversarial attacks.Limiting update granularity and adding noise for differential privacy may reduce leakage, but this remains an active research area.
- Data heterogeneity: Medical data varies across types, devices, acquisition protocols, and demographics, challenging FL methods that assume independent and identically distributed data.This heterogeneity occurs both generally and within defined medical tasks.
- Reproducibility and interpretability: Healthcare FL requires reproducibility, traceability, and accountability across complex multi-party hardware, software, and network environments.Researchers also cannot directly inspect images used for training, motivating secure intra-node viewing or global-model explainability tools, while interpretability remains open.
- Future opportunities: FL can integrate knowledge from multiple institutions, capturing greater data variability and enabling analysis across different patient demographics.Future healthcare AI may depend more on collaboration among institutions than on large centralized databases held by a single hospital or laboratory.
4 Conclusion
Federated Learning is presented as a promising approach for collaboratively developing powerful, accurate, safe, robust, and unbiased digital-health models without exchanging or centralizing datasets. Although technical questions remain unresolved, FL may create new research and business opportunities and improve global patient care.
- 4 Conclusion: FL enables multiple parties to train collaboratively without exchanging or centralizing datasets, addressing sensitive medical-data egress concerns.Its privacy-preserving collaborative structure supports access to broader data distributions while keeping datasets decentralized.
- 4 Conclusion: FL may support powerful, accurate, safe, robust, and unbiased models by improving access to data approximating the true global distribution.The paper frames this potential as especially relevant because machine-learning methods benefit from broadly representative data.
- 4 Conclusion: Despite FL’s potential to open research and business avenues and improve patient care globally, unresolved technical questions ensure it will remain an active research area throughout the next decade.The article discusses both FL’s healthcare benefits and the considerations that still need to be addressed.