Making Cross-Continental Federated Learning Repeatable with FLIP: a Multi-Application Study

Rafael Garcia-Dias; Alexandre Triay Bagur; Chayanin Tangwiriyasakul; Virginia Fernandez; Parhom Esmaeili; Piyalitt Ittichaiwong; Yang Li; Lawrence Adams

cs.LGSep 28, 2026

Making Cross-Continental Federated Learning Repeatable with FLIP: a Multi-Application Study

Authors: Rafael Garcia-Dias, Alexandre Triay Bagur, Chayanin Tangwiriyasakul, Virginia Fernandez, Parhom Esmaeili, Piyalitt Ittichaiwong, Yang Li, Lawrence Adams, +15 more

Organizations: School of Biomedical Engineering & Imaging Sciences, King’s College London, London, UK · Bangkok Dusit Medical Services, Bangkok, Thailand

Abstract

Federated learning (FL) in healthcare remains challenging, as the overhead of rebuilding governance guarantees for every collaboration stops most projects at the proof-of-concept stage. Here we present FLIP (Federated Learning Interoperability Platform), an open-source, multi-application platform that makes FL training and evaluation repeatable. FLIP implements common FL workflows as a set of composable services: cohort queries against per-site structured databases, on-demand DICOM retrieval from institutional PACS, per-site project approval, and reusable FL job types. To demonstrate FLIP, we ran two distinct use cases, federated fine-tuning and federated evaluation, on synthetic chest X-ray cohorts across two client nodes based in the United Kingdom (UK) and Thailand. In FLIP, each institution independently approves its participation in each project and operates its own node under local IT security processes. This study makes an operational rather than an algorithmic claim. It does not compare federated with centralised training; for that question, we refer the reader to existing systematic reviews and meta-analyses. The central result is evidence that such platforms enable international FL collaboration and improve repeatability, auditability, and site-specific governance. We also present a comprehensive comparison of existing platforms to help researchers and operators choose the right platform for their use case.

Figures & tables

Figure 1: FLIP architecture. The Central Hub (left, cloud-hosted) handles coordination and the FL server, while each Trust Node (middle) hosts an OMOP database, XNAT, and an FL client attached to GPU compute. Data leaving the TN are limited to model weight updates, aggregate cohort counts, per-round scalar metrics, and execution logs; no row-level records or image data are transferred.

Figure 2: Screenshot of the cohort query for the training set, with the aggregated per-site record counts on the right. A WHERE clause separates the training and holdout sets. The platform accepts SQL SELECT statements only; the full query is available in the FLIP repository.

Platform	Sites (countries)	Data	Workloads shown
FLIP (ours)	2 (2: UK, TH)	synthetic CXR	fine-tuning, evaluation
FeTS [ 17 ]	71 (6 continents)	real MRI, 6314 cases	segmentation
RACOON [ 3 ]	6 of 38 (1: DE)	real CT, 582 scans	segmentation
MedPerf [ 43 ]	9 (13)	real, multi-site	evaluation only
CODA [ 39 ]	8 of 9 (1: CA)	1.09M patients	analytics; FL simulated
JIP [ 41 ]	11 (1: DE)	real, multi-centre	imaging

Table 3: Deployment evidence reported by each platform. Scale and data realism are the weakest aspects of this study, and we report them explicitly rather than leave them to be inferred. Sites are those reported as deployed . Where a platform’s published federated training ran on a subset of them, or on a simulated partition of one dataset, this is stated, since the two are routinely conflated.

Class	UK Synthetic (n=478)			Thai Synthetic (n=468)
	Pretr.	F-tuned	$p$	Pretr.	F-tuned	$p$
Effusion	0.998	1.000	0.12 n.s.	1.000	1.000	1.00 n.s.
Consolidation	0.989	0.997	0.027	0.998	1.000	0.31 n.s.
Infiltration	0.863	0.998	$1.8{\times}10^{-14}$	0.878	1.000	$1.2{\times}10^{-15}$
Nodule or Mass	0.970	1.000	$5.9{\times}10^{-6}$	0.955	1.000	$7.2{\times}10^{-7}$
Pneumothorax	0.971	1.000	0.0065	0.997	1.000	0.19 n.s.

Table 4: Per-class AUROC of pretrained and FedAvg fine-tuned Ark+ on UK and Thai holdout sets.

p

-values are from two-sided DeLong tests, Benjamini–Hochberg (BH) corrected within each site. The testing family is the 5 classes for the single model pair at that site (

m{=}5

), so the BH threshold for the

k

-th smallest

p

-value is

0.05k/5

, ranging from

0.01

0.05

. Four comparisons survive correction at the UK site and two at the Thai site, whereas effusion (both sites) and Thai consolidation and pneumothorax do not (n.s.). Bold values mark the best model per class and site at

Q{=}0.05

Figure 3: FLIP interface showing federated training results, UK–Thailand deployment.

Phase	Depl. UK–TH	Depl. same-net	Sim. UK	Sim. TH
Rounds 1–49 ( $T_{\mathrm{round}}$ )	$745.8\pm 0.7$	$181.7\pm 0.8$	$275.4\pm 0.5$	$1{,}605.2\pm 2.3$
Round 0 (init.)	$3{,}830.0$	$334.6$	$294.7$	$1{,}696.2$
Aggregation ( $T_{\mathrm{agg}}$ )	$0.19\pm 0.06$	$0.18\pm 0.04$	$0.22\pm 0.05$	$0.39\pm 0.09$
Checkpoint persist. ( $T_{\mathrm{persist}}$ )	$5.52\pm 0.04$	—	—	—
Inter-round gap ( $T_{\mathrm{gap}}$ )	$0.19\pm 0.02$	$0.19\pm 0.02$	$0.21\pm 0.02$	$0.29\pm 0.03$
Model transfer ( $T_{\mathrm{down}}$ )	$<0.2$	—	—	—

Table 5: Round-loop timing over 50 rounds (mean

\pm

std, seconds; gap

n{=}49

). Each simulator serialises both clients on one GPU, so its round time is approximately the sum of both sites’

T_{\mathrm{compute}}

. Client turnaround runs from task dispatch to acceptance of the contribution, and therefore spans

T_{\mathrm{down}}{+}T_{\mathrm{compute}}{+}T_{\mathrm{up}}

. Dashes mark phases not yet re-derived from the corresponding job export, as only the UK–TH deployment has been decomposed at this granularity.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Platform	Layer	PubMed	GitHub	arXiv	medRxiv
APPFL	deployed	yes	yes	no	no
CODA	deployed	yes	–	no	no
FLA 3	deployed	–	yes	yes	no
FeatureCloud	deployed	–	yes	yes	no
FeTS	deployed	yes	yes	yes	no
FL4Health	deployed	–	yes	no	no

Appendix

Table 6: Retrieval of each comparator by the four searches, measured against the saved candidate pools. yes the pool contained an artefact for that platform; no it did not; – the platform has no artefact of that kind to retrieve (no indexed paper, or no public repository). The PubMed and arXiv columns are assessed over the 1,485-record and 1,190-preprint pools respectively, the GitHub column over the 2,330-repository union, and the medRxiv column over the 167-preprint match pool. A platform paper here means a record describing the platform, not a study merely citing it.

Explore similar work

Multi-center Medical Data Mining with FL-Net - A One-stop Shop for Federated Learning
Sep 17, 2026Simon Süwer, Julian Klemm, Elisa Acitelli +38Federated Learning Multicenter Evaluation
Development and Design of FLKit: A Structured Onboarding Toolkit for Federated Learning in Health and Life Sciences
Jun 22, 2026Ashkan Pirmani, Ilse Vermeulen, Goran Vinterhalter +7Federated Learning Life Sciences Research
1STADIUS, Department of Electrical Engineering (ESAT), KU Leuven, Leuven, Belgium · 2Biomedical Research Institute (BIOMED), Hasselt University, Hasselt, Belgium · 3Data Science Institute (DSI), Hasselt University, Hasselt, Belgium +1
Embedding-Based Federated Learning with Runtime Governance for Iron Deficiency Prediction
May 20, 2026Fan Zhang, Simon Deltadahl, Majid Lotfian Delouee +10Federated Learning Fedavg
Dept. of Applied Mathematics and Theoretical Physics University of Cambridge Cambridge, UK · Translational AI Laboratory, Dept. of Laboratory Medicine Amsterdam UMC Amsterdam, The Netherlands · Precision Health University Research Institute Queen Mary Univ. of London London, UK +3

Sep 17, 2026, cs.LGJ/K move · Enter open · S save

Multi-center Medical Data Mining with FL-Net - A One-stop Shop for Federated Learning

Simon Süwer, Julian Klemm, Elisa Acitelli, Mathieu Almeida+37

Federated Learning Multicenter Evaluation Clinical Trials Harmonization

Open paper AskPDF arXiv

Federated learning enables collaborative training without sharing patient-level data, but most studies remain simulations. Based on five requirements derived from the literature, we analyzed 14 FL frameworks and found that none fully satisfied these requirements. We present FL-Net, a novel federated clinical research framework to fulfill all requirements. It integrates modular data harmonization, data discovery, disclosure control, securely built versioned FL-Net-Tools and containerized federated workflow execution into a persistent network. It enables the re-use of harmonized data and workflows across studies. FL-Net's end-to-end capabilities were evaluated through harmonization, cross-study patient discovery across MIMIC and US-130, and reproducible, audited federated workflows with up to 50 concurrent clients. FL-Net is being developed within the dAIbetes and Microb-AI-ome EU projects and will cover over 800,000 patients across 10 hospitals in 9 countries covering longitudinal and single point in time data, FL-Net provides a practical foundation for interoperable, reproducible, and privacy-preserving multicenter clinical research.

Abstract

Explore similar work

Multi-center Medical Data Mining with FL-Net - A One-stop Shop for Federated Learning

Development and Design of FLKit: A Structured Onboarding Toolkit for Federated Learning in Health and Life Sciences

Embedding-Based Federated Learning with Runtime Governance for Iron Deficiency Prediction