The future of digital health with federated learning

The future of digital health with federated learning Nicola Rieke Jonny Hancox Wenqi Li Fausto Milletarì Holger R. Roth Shadi Albarqouni Spyridon Bakas Mathieu N. Galtier Bennett A. Landman Klaus Maier-Hein Sébastien Ourselin Micah Sheller Ronald M. Summers Andrew Trask Daguang Xu Maximilian Baust M. Jorge Cardoso

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

The future of digital health with federated learning 2020 npj Digital Medicine 10.1038/s41746-020-00323-1 rieke2020-the-future-of-digital-health-with rieke2020-the-future-of-digital-health-with.txt, the project's staged copy 9dbf0a30957c1af8ae54f7a7d9a200b9bcf3d664 cc-by machine-learning

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them federated learning as an answer to data governance and privacy constraints in healthcare machine learning Quotation withheld because this record is not cleared for redistribution. how to train on medical data that cannot be centralised Quotation withheld because this record is not cleared for redistribution. a consensus perspective on what FL offers digital health and what it still requires Quotation withheld because this record is not cleared for redistribution. prediction Quotation withheld because this record is not cleared for redistribution. medical data sitting in silos that machine learning cannot reach Quotation withheld because this record is not cleared for redistribution. review review Quotation withheld because this record is not cleared for redistribution. mechanisms MOVE THE MODEL, NOT THE DATA: training happens locally inside each institution and only model characteristics leave the firewall, so the consensus model is built without the patient records ever being pooled Quotation withheld because this record is not cleared for redistribution. mechanisms WHY ANONYMISATION CANNOT BE THE ANSWER: identity survives the removal of the obvious fields - a face is reconstructable from a CT scan, and a handful of record elements permits reidentification Quotation withheld because this record is not cleared for redistribution. mechanisms DATA IS WITHHELD FOR COMMERCIAL REASONS AS WELL AS LEGAL ONES: curating a high-quality set costs time and money, which gives it business value and makes sharing less likely - a motive no privacy technology addresses Quotation withheld because this record is not cleared for redistribution. FL is stated to still require rigorous technical consideration to avoid compromising safety or patient privacy - the paradigm is offered as promising rather than solved Quotation withheld because this record is not cleared for redistribution. The authors say plainly that not all technical questions have been answered and expect FL to remain an open research area for a decade Quotation withheld because this record is not cleared for redistribution. The performance claim is attributed to other work rather than demonstrated here - this is a perspective paper citing comparisons, not running them Quotation withheld because this record is not cleared for redistribution. Public medical data carries varying license restrictions that sometimes limit its exploitation, which bounds what any centralised alternative could have used anyway Quotation withheld because this record is not cleared for redistribution. Federated learning Federated learning Quotation withheld because this record is not cleared for redistribution. data governance data governance Quotation withheld because this record is not cleared for redistribution. consensus model consensus model Quotation withheld because this record is not cleared for redistribution. anonymisation anonymisation Quotation withheld because this record is not cleared for redistribution. reidentification reidentification Quotation withheld because this record is not cleared for redistribution. privacy privacy Quotation withheld because this record is not cleared for redistribution. precision medicine precision medicine Quotation withheld because this record is not cleared for redistribution. license restrictions license restrictions Quotation withheld because this record is not cleared for redistribution. egress egress Quotation withheld because this record is not cleared for redistribution. business value business value Quotation withheld because this record is not cleared for redistribution. firewalls firewalls Quotation withheld because this record is not cleared for redistribution. Data-driven machine learning (ML) Data-driven machine learning (ML) Quotation withheld because this record is not cleared for redistribution. accurate and robust statistical models accurate and robust statistical models Quotation withheld because this record is not cleared for redistribution. data silos data silos Quotation withheld because this record is not cleared for redistribution. privacy concerns privacy concerns Quotation withheld because this record is not cleared for redistribution. without access to sufficient data without access to sufficient data Quotation withheld because this record is not cleared for redistribution. full potential full potential Quotation withheld because this record is not cleared for redistribution. transition from research to clinical practice transition from research to clinical practice Quotation withheld because this record is not cleared for redistribution. Centralised Data Lake Centralised Data Lake Quotation withheld because this record is not cleared for redistribution. FL aggregation server FL aggregation server Quotation withheld because this record is not cleared for redistribution. federation of training nodes federation of training nodes Quotation withheld because this record is not cleared for redistribution. global model global model Quotation withheld because this record is not cleared for redistribution. partially trained models partially trained models Quotation withheld because this record is not cleared for redistribution. central server central server Quotation withheld because this record is not cleared for redistribution. FL peer to peer FL peer to peer Quotation withheld because this record is not cleared for redistribution. Centralised training Centralised training Quotation withheld because this record is not cleared for redistribution. local, independent training local, independent training Quotation withheld because this record is not cleared for redistribution. ML from non-co-located data ML from non-co-located data Quotation withheld because this record is not cleared for redistribution. data controller data controller Quotation withheld because this record is not cleared for redistribution. governance processes governance processes Quotation withheld because this record is not cleared for redistribution. privacy policies privacy policies Quotation withheld because this record is not cleared for redistribution. data access data access Quotation withheld because this record is not cleared for redistribution. validation phase validation phase Quotation withheld because this record is not cleared for redistribution. large-scale, ininstitutional validation large-scale, ininstitutional validation Quotation withheld because this record is not cleared for redistribution. rare diseases rare diseases Quotation withheld because this record is not cleared for redistribution. high-dimensional, storage-intense medical data high-dimensional, storage-intense medical data Quotation withheld because this record is not cleared for redistribution. centralised pool centralised pool Quotation withheld because this record is not cleared for redistribution. growing global data set growing global data set Quotation withheld because this record is not cleared for redistribution. data storage requirements data storage requirements Quotation withheld because this record is not cleared for redistribution. FL topologies FL topologies Quotation withheld because this record is not cleared for redistribution. Centralised Centralised Quotation withheld because this record is not cleared for redistribution. Decentralised Decentralised Quotation withheld because this record is not cleared for redistribution. Hierarchical Hierarchical Quotation withheld because this record is not cleared for redistribution. FL compute plans FL compute plans Quotation withheld because this record is not cleared for redistribution. Sequential training/cyclic transfer learning Sequential training/cyclic transfer learning Quotation withheld because this record is not cleared for redistribution. memorise information memorise information about this quote: THE QUOTE CONTINUES A SENTENCE THAT THIS COPY CUTS IN TWO - a figure caption (the FL compute plans panels) stands between it and the words it follows, which are not adjacent to it in the extraction. Quotation withheld because this record is not cleared for redistribution. differential privacy differential privacy about this quote: THE QUOTE CONTINUES A SENTENCE THAT THIS COPY CUTS IN TWO - a figure caption (the FL compute plans panels) stands between it and the words it follows, which are not adjacent to it in the extraction. Quotation withheld because this record is not cleared for redistribution. learning from encrypted data learning from encrypted data about this quote: THE QUOTE CONTINUES A SENTENCE THAT THIS COPY CUTS IN TWO - a figure caption (the FL compute plans panels) stands between it and the words it follows, which are not adjacent to it in the extraction. Quotation withheld because this record is not cleared for redistribution. volume-based volume-based Quotation withheld because this record is not cleared for redistribution. fee-for-service-based fee-for-service-based Quotation withheld because this record is not cleared for redistribution. value-based healthcare value-based healthcare Quotation withheld because this record is not cleared for redistribution. focused treatment focused treatment Quotation withheld because this record is not cleared for redistribution. accuracy and robustness of healthcare AI accuracy and robustness of healthcare AI Quotation withheld because this record is not cleared for redistribution. costs costs Quotation withheld because this record is not cleared for redistribution. patient outcomes patient outcomes Quotation withheld because this record is not cleared for redistribution. FL FL Quotation withheld because this record is not cleared for redistribution. learning paradigm learning paradigm Quotation withheld because this record is not cleared for redistribution. multiple parties multiple parties Quotation withheld because this record is not cleared for redistribution. global loss function global loss function Quotation withheld because this record is not cleared for redistribution. weighted combination of K local losses weighted combination of K local losses Quotation withheld because this record is not cleared for redistribution. private data private data Quotation withheld because this record is not cleared for redistribution. data heterogeneity data heterogeneity Quotation withheld because this record is not cleared for redistribution. inhomogeneous data distribution inhomogeneous data distribution Quotation withheld because this record is not cleared for redistribution. independently and identically distributed (IID) data independently and identically distributed (IID) data Quotation withheld because this record is not cleared for redistribution. FedAvg FedAvg Quotation withheld because this record is not cleared for redistribution. FedProx FedProx Quotation withheld because this record is not cleared for redistribution. part-data-sharing strategy part-data-sharing strategy Quotation withheld because this record is not cleared for redistribution. FL with domainadaptation FL with domainadaptation Quotation withheld because this record is not cleared for redistribution. global optimal solution global optimal solution Quotation withheld because this record is not cleared for redistribution. optimal for an individual local participant optimal for an individual local participant Quotation withheld because this record is not cleared for redistribution. model training optimality model training optimality Quotation withheld because this record is not cleared for redistribution. Privacy-preserving techniques for FL Privacy-preserving techniques for FL Quotation withheld because this record is not cleared for redistribution. levels of protection levels of protection Quotation withheld because this record is not cleared for redistribution. accuracy of the final model accuracy of the final model Quotation withheld because this record is not cleared for redistribution. future techniques future techniques Quotation withheld because this record is not cleared for redistribution. ancillary data ancillary data Quotation withheld because this record is not cleared for redistribution. low-risk low-risk Quotation withheld because this record is not cleared for redistribution. Non-trusted Non-trusted Quotation withheld because this record is not cleared for redistribution. security strategies security strategies Quotation withheld because this record is not cleared for redistribution. advanced encryption of model submissions advanced encryption of model submissions Quotation withheld because this record is not cleared for redistribution. secure authentication of all parties secure authentication of all parties Quotation withheld because this record is not cleared for redistribution. traceability of actions traceability of actions Quotation withheld because this record is not cleared for redistribution. verification systems verification systems Quotation withheld because this record is not cleared for redistribution. execution integrity execution integrity Quotation withheld because this record is not cleared for redistribution. model confidentiality model confidentiality Quotation withheld because this record is not cleared for redistribution. protections against adversarial attacks protections against adversarial attacks Quotation withheld because this record is not cleared for redistribution. traditional training traditional training Quotation withheld because this record is not cleared for redistribution. training process is exposed to multiple parties training process is exposed to multiple parties Quotation withheld because this record is not cleared for redistribution. risk of leakage via reverseengineering risk of leakage via reverseengineering Quotation withheld because this record is not cleared for redistribution. multiparty computations multiparty computations Quotation withheld because this record is not cleared for redistribution. training configurations training configurations Quotation withheld because this record is not cleared for redistribution. hyperparameter tuning hyperparameter tuning Quotation withheld because this record is not cleared for redistribution. contribution from each participant contribution from each participant Quotation withheld because this record is not cleared for redistribution. computational resources consumed computational resources consumed Quotation withheld because this record is not cleared for redistribution. quality of the data used for local training quality of the data used for local training Quotation withheld because this record is not cleared for redistribution. compensation compensation Quotation withheld because this record is not cleared for redistribution. revenue model revenue model Quotation withheld because this record is not cleared for redistribution. secure intra-node viewing facility secure intra-node viewing facility Quotation withheld because this record is not cleared for redistribution. explainability and interpretability of the global model explainability and interpretability of the global model Quotation withheld because this record is not cleared for redistribution. redundant nodes redundant nodes Quotation withheld because this record is not cleared for redistribution. secure encryption methods secure encryption methods Quotation withheld because this record is not cleared for redistribution. data leakage data leakage Quotation withheld because this record is not cleared for redistribution. node schedulers node schedulers Quotation withheld because this record is not cleared for redistribution. “honest broker” system “honest broker” system Quotation withheld because this record is not cleared for redistribution. trusted third party trusted third party Quotation withheld because this record is not cleared for redistribution. peer-topeer system peer-topeer system Quotation withheld because this record is not cleared for redistribution. gatekeeper function gatekeeper function Quotation withheld because this record is not cleared for redistribution. trustlessbased architecture trustlessbased architecture Quotation withheld because this record is not cleared for redistribution. secure protocol secure protocol Quotation withheld because this record is not cleared for redistribution. computational overheads computational overheads Quotation withheld because this record is not cleared for redistribution. FL → Federated learning FL expands to Federated learning. Acronym expansion is absent from the eight edge types. Quotation withheld because this record is not cleared for redistribution. FL → ML from non-co-located data FL enables machine learning from non-co-located data. Enablement is absent from the eight edge types and is not safely reducible to causation here. Quotation withheld because this record is not cleared for redistribution. Non-trusted → security strategies Non-trusted FL collaborations require security strategies. Prerequisite-for is absent from the eight edge types. Quotation withheld because this record is not cleared for redistribution. traceability and accountability processes → execution integrity Traceability and accountability in non-trusted federations require execution integrity. Prerequisite-for is absent from the eight edge types. Quotation withheld because this record is not cleared for redistribution. “honest broker” system → trusted third party An honest-broker federation is administered through a trusted third party acting as intermediary. Administered-through or mediated-by is absent from the eight edge types. Quotation withheld because this record is not cleared for redistribution. anonymisation - TREATED HERE AS INSUFFICIENT BY CONSTRUCTION, not as imperfect in practice: removing name and date of birth does not preserve privacy, a face can be reconstructed from a CT scan, and genomic data is as unique as a fingerprint Quotation withheld because this record is not cleared for redistribution. Federated learning - collaborative model training without exchanging or centralising the parties' data sets Quotation withheld because this record is not cleared for redistribution. Aggregation-server FL - nodes receive a global model, train locally, and return partial models for central aggregation Quotation withheld because this record is not cleared for redistribution. Peer-to-peer FL - each node exchanges partial models with peers and performs its own aggregation Quotation withheld because this record is not cleared for redistribution. Centralised training - sites donate data to a shared lake for extraction and independent local training Quotation withheld because this record is not cleared for redistribution. FL topology - centralised, decentralised, or hierarchical communication architecture Quotation withheld because this record is not cleared for redistribution. FL compute plan - trajectory followed by a model across participating partners Quotation withheld because this record is not cleared for redistribution. Data heterogeneity - non-IID cross-institutional distributions that challenge federated aggregation and local optimality Quotation withheld because this record is not cleared for redistribution. Privacy-performance trade-off - stronger federated protections may affect final-model accuracy Quotation withheld because this record is not cleared for redistribution. Non-trusted federation - collaboration needing encryption, authentication, traceability, verification, integrity, and adversarial defenses Quotation withheld because this record is not cleared for redistribution. Federated information leakage - model changes exposed to multiple parties create reverse-engineering risk even without raw-data sharing Quotation withheld because this record is not cleared for redistribution. Honest-broker administration - trusted intermediary controls the federation and abstracts internal mechanisms from clients Quotation withheld because this record is not cleared for redistribution. Peer-to-peer administration - sites interact directly with no gatekeeper and must synchronise protocol changes Quotation withheld because this record is not cleared for redistribution. Federated learning —distinct-from→ anonymisation Federated learning anonymisation strength of the relation: 0.8 Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the span describes federated learning's privacy aim; distinct-from 'anonymisation' is loose. Quotation withheld because this record is not cleared for redistribution. without access to sufficient data —causes→ transition from research to clinical practice without access to sufficient data transition from research to clinical practice direction on the measure: eliminates · the relation itself is separately evidenced · strength of the relation: 1 · note: The paper says insufficient access prevents ML from making the research-to-clinic transition. Quotation withheld because this record is not cleared for redistribution. Centralised training —distinct-from→ FL aggregation server Centralised training FL aggregation server the relation itself is separately evidenced · strength of the relation: 1 · note: The figure explicitly presents the difference between federated workflows and the non-FL centralised-data-lake workflow. Quotation withheld because this record is not cleared for redistribution. learning from encrypted data —causes→ privacy learning from encrypted data privacy direction on the measure: increases · the relation itself is separately evidenced · strength of the relation: 0.7 · about this quote: THE QUOTE CONTINUES A SENTENCE THAT THIS COPY CUTS IN TWO - a figure caption (the FL compute plans panels) stands between it and the words it follows, which are not adjacent to it in the extraction. · note: The same proposed-mechanism hedge applies, so paper weight is 0.7 while confidence remains 0.95. Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says learning from encrypted data HAS BEEN PROPOSED to further enhance privacy; hedged. Quotation withheld because this record is not cleared for redistribution. FL —is-a→ learning paradigm FL learning paradigm the relation itself is separately evidenced · strength of the relation: 1 · note: The paper directly defines FL as a learning paradigm. Quotation withheld because this record is not cleared for redistribution. global optimal solution —distinct-from→ optimal for an individual local participant global optimal solution optimal for an individual local participant the relation itself is separately evidenced · strength of the relation: 0.7 · note: The paper says the global optimum may not be locally optimal; weight 0.7 preserves both modal hedges. Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says the global optimum MAY not be optimal for a local participant; hedged. Quotation withheld because this record is not cleared for redistribution. training process is exposed to multiple parties —causes→ risk of leakage via reverseengineering training process is exposed to multiple parties risk of leakage via reverseengineering direction on the measure: increases · the relation itself is separately evidenced · strength of the relation: 1 · note: The paper directly says multi-party exposure increases reverse-engineering leakage risk. Quotation withheld because this record is not cleared for redistribution. secure encryption methods —causes→ data leakage secure encryption methods data leakage direction on the measure: reduces · the relation itself is separately evidenced · strength of the relation: 1 · note: Secure encryption is named as an architecture measure to prevent data leakage. Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span names designing secure encryption methods to prevent leakage as a CHALLENGE; causes is loose. Quotation withheld because this record is not cleared for redistribution. FL —causes→ accuracy and robustness of healthcare AI FL accuracy and robustness of healthcare AI direction on the measure: increases · the relation itself is separately evidenced · strength of the relation: 0.7 · note: Weight 0.7 preserves the paper's 'has the potential to increase' wording. Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says FL HAS THE POTENTIAL to increase accuracy and robustness; hedged. Quotation withheld because this record is not cleared for redistribution. FL —causes→ costs FL costs direction on the measure: reduces · the relation itself is separately evidenced · strength of the relation: 0.7 · note: The projected cost reduction shares the paper's potential-language hedge. Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the span shown is about paradigm shifts in healthcare payment; the cost claim is not visible. Quotation withheld because this record is not cleared for redistribution. FL —causes→ patient outcomes FL patient outcomes direction on the measure: increases · the relation itself is separately evidenced · strength of the relation: 0.7 · note: The projected improvement in patient outcomes is weighted 0.7 for the paper's modal framing. Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says FL HAS THE POTENTIAL to improve patient outcomes; hedged. Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1038/s41746-020-00323-1.