CANAL-AI

CANAL-AI A Franco-Japanese federated-learning pilot

Caen–Nagoya Alliance for AI in Health Data

Status: regulatory phase · No patient data processed yet

Only model parameters cross the border. Patient data never moves.

We train a shared medical AI across the data warehouses of Caen University Hospital (France) and Nagoya University Hospital (Japan) to predict the cardiovascular side-effects of hormone therapy in prostate cancer.

Support the pilot How it works
Symbolic illustration of Caen and Nagoya, the two cities of the CANAL-AI consortium, joined across a single river under a shared sunset
Two cities, one federation Symbolic illustration — Mont-Saint-Michel (Normandy) and Nagoya Castle (Aichi) as cultural emblems of the two regions joined by the pilot. AI-rendered composite, not a depiction of the hospital sites.

Announcements

  1. Dr Nishida selected for the joint travel grant of the French Embassy in Japan and the Goto Kiyoko – Paul Bourdarie Cancer Foundation

    Read more

  2. Dr Dolladille visits Nagoya University

    Dr Charles Dolladille visited Nagoya University on 13 and 14 April 2026 to meet the CANAL-AI team on site. On day 2, Adeline Lassaux and Sonia Ciechelski from the French Embassy in Japan joined the discussions about the project and the cooperation opportunities that follow from it.

    Read more
    Dr Charles Dolladille with the CANAL-AI team — including Dr Nishida — at Nagoya University, 13 April 2026

  3. Project website launched

    Read more

  4. SECOM Science and Technology Foundation grant awarded

    Read more

  5. Caen Ethics Committee approves the CANAL-AI protocol

    Read more

  6. Letter of Intent Nagoya ↔ Caen signed

    Read more

A pilot, by design

One pilot. A network to come.

Two sites are a deliberate choice: enough to demonstrate a federation designed for GDPR and APPI conformity end-to-end, small enough to document every step — governance, harmonisation, secure aggregation — in detail.

Our core deliverables are a reusable protocol and an APPI–GDPR operational blueprint that will let additional hospitals plug in without rebuilding the legal, ethical, and technical groundwork.

Beyond Caen–Nagoya, any hospital willing to keep patient data on-site and participate in federated learning is a candidate partner — see Become a partner hospital.

Why cardio-oncology needs federated learning

Androgen-deprivation therapy (ADT) is a cornerstone of prostate-cancer management but carries clinically meaningful cardiovascular and metabolic risks. GnRH agonists, GnRH antagonists, and ARSI do not share the same cardiovascular risk profile; HERO (NEJM, 2020) and PRONOUNCE (Circulation, 2021) are mixed, with the antagonist advantage concentrated in men with pre-existing CVD.

Single-country models fail to generalise across populations with different practices. International data is the answer; GDPR and APPI are the barrier. Federated learning is the bridge.

FedBioMed

How it works

Built on FedBioMed, scoped to what two hospitals can actually ship.

We will use FedBioMed, the open-source federated-learning framework developed by INRIA and Université Côte d'Azur. Each hospital runs a local FedBioMed node on harmonised OMOP-CDM data. An aggregator at a neutral location coordinates training rounds — parameters travel, data do not.

Federated-learning architecture Caen and Nagoya University Hospitals each run a local FedBioMed node; encrypted model updates travel to a central aggregator and back; patient data never leaves either site. Caen Univ. Hospital data stays on site Nagoya Univ. Hospital data stays on site FedBioMed Aggregator gradients only — no patient data Future partner encrypted model updates encrypted model updates Round 1 / N

Local training

Each hospital trains a model on its own patient data — locally, inside its own firewall. Nothing leaves the building yet.

Secure send

Only the resulting model parameters are sent to the FedBioMed aggregator, encrypted via Joye-Libert secure aggregation.

Average

The aggregator averages parameters across sites to produce an updated shared model — no raw data, just maths.

Update and repeat

The updated model returns to every hospital. After many rounds, each one holds a model that has learned from all — without any patient data ever leaving.

The stack, with glosses

Primary model
PyTorch MLP on OMOP-derived tabular features, trained with FedAvg. FedProx as a non-IID fallback.
Baseline
Federated logistic regression via FedBioMed's scikit-learn training plan (FedSGDClassifier, log-loss).
Privacy layer
Secure aggregation (Joye-Libert). Differential privacy via OPACUS evaluated as an additional layer, with reported (ε, δ) budgets.
Harmonisation
OMOP-CDM at each site with a custom ETL. Publishing this ETL is one of our deliverables.
Explainability
SHAP + permutation importance, computed locally on each site's held-out cohort; rankings aggregated across sites; clinician review by cardio-oncology experts.
Evaluation
Leave-one-site-out validation — AUROC, AUPRC, calibration (Brier, ICI), decision-curve analysis.
Endpoint
Major Adverse Cardiovascular Events (MACE) within 12 months of ADT initiation — a 4-point composite (myocardial infarction, ischaemic stroke, heart-failure hospitalisation, cardiovascular death) aligned with the HERO definition.

Regulatory status and ethics

Research-use-only pilot. Caen Scientific & Ethics Committee favourable opinion received 30 March 2026. Nagoya IRB application and GDPR DPIA in progress. No patient data has been processed to date.

Any clinical deployment would require separate CE-marking (EU MDR 2017/745, Class IIa software, Rule 11) and PMDA SaMD clearance in Japan. French processing will follow MR-003/MR-004; Japanese processing will follow APPI Article 76 — final modalities subject to IRB approval.

Glossary of abbreviations Show / hide definitions
ADT
Androgen-Deprivation Therapy — the hormone therapy that lowers testosterone.
ARSI
Androgen-Receptor Signalling Inhibitors.
GnRH
Gonadotropin-Releasing Hormone.
MACE
Major Adverse Cardiovascular Events.
CVD
Cardiovascular Disease.
OMOP-CDM
International common data model for electronic health records.
ETL
Extract, Transform, Load — data-engineering pipeline.
EDS
Entrepôt de Données de Santé — Caen health-data warehouse.
GDPR
EU personal-data protection regulation.
APPI
Japanese personal-data protection law.
DPIA
Data Protection Impact Assessment.
IRB
Institutional Review Board — research ethics committee.
MR-003 / MR-004
French CNIL reference frameworks for health-data research.
CE / MDR / SaMD
EU Medical Device Regulation; Software as a Medical Device.
PMDA
Japanese Pharmaceuticals and Medical Devices Agency.
Federated learning
Training across sites by exchanging model parameters, not data.
FedBioMed
INRIA / UCA open-source federated-learning framework.
FedAvg
Standard federated averaging algorithm.
FedProx
FedAvg variant robust to heterogeneous (non-IID) sites.
SecAgg (Joye-Libert)
Cryptographic sum of encrypted updates.
OPACUS
PyTorch differential-privacy library (DP-SGD).
SHAP
SHapley Additive exPlanations — feature-importance scores.
AUROC / AUPRC
Area under ROC / precision-recall curves — discrimination metrics.
Brier / ICI
Calibration metrics.
DCA
Decision Curve Analysis — clinical-utility metric.

Deliverables

Outputs and deliverables

Each deliverable is tagged with its target audience, target year, licence, and persistent identifier on release.

  1. Q3 2027

    Methods paper — open-access; audience: methodologists, cardio-oncologists; preprint + peer-reviewed venue.

  2. Q4 2027

    APPI–GDPR operational framework — CC BY 4.0; archived on Zenodo with DOI; audience: hospital DPOs, IRBs, FL consortia.

  3. Q1 2028

    OMOP-to-FedBioMed ETL — MIT licence on GitHub + Zenodo DOI; audience: data engineers and biomedical-informatics teams.

  4. 2028

    Clinical paper — open-access; audience: oncologists, cardiologists, regulators.

12–24 month plan

Roadmap

Indicative milestones; detailed timing is subject to IRB and convention outcomes.

  1. Q2–Q3 2026

    Nagoya IRB submission. DPIA. Data-access convention at Caen. FedBioMed kickoff.

  2. Q4 2026

    OMOP variable list locked. Cohort definition frozen. Synthetic-data prototype end-to-end.

  3. Q1–Q2 2027

    First federated round. MLP baseline with SecAgg. Sample-size and power report.

  4. Q3–Q4 2027

    Model comparison, methods paper submitted, APPI–GDPR blueprint released.

  5. 2028

    External validation at a third site. Clinical paper. Open-source pipeline release.

Risks we take seriously

Transparency with funders and partners:

  • IRB timing

    Nagoya IRB and the DPIA set the pace. Mitigation: blueprint work proceeds in parallel.

  • Sample size and event rate

    Two sites may yield limited MACE events at 12 months. Mitigation: power report, pragmatic baselines, plans for a third site.

  • Non-IID heterogeneity

    Sites differ in case-mix and practice. Mitigation: FedProx, stratified evaluation, leave-one-site-out validation.

Principal investigators

Dr Charles Dolladille

Dr Charles Dolladille

MD, PhD — cardiologist and pharmacologist. MCU-PH, Caen University Hospital; INSERM U1086 ANTICIPE.

Dr Basile Chrétien

Dr Basile Chrétien

PharmD, MPH, MSc — pharmacologist. PhD candidate, Nagoya University.

Dr Kazuki Nishida

Dr Kazuki Nishida

MD, PhD — biostatistician with machine-learning expertise, Nagoya University.

Institutional partners

  • Nagoya University

    Graduate School of Medicine — scientific and institutional lead on the Japanese side.

    Official site

  • Nagoya University Hospital

    Japanese clinical and data-warehouse partner — hosts the FedBioMed node on the Nagoya side.

    Official site

  • Caen University Hospital

    Entrepôt de Données de Santé, Department of Pharmacology, INSERM U1086 ANTICIPE.

    Official site

  • Université de Caen Normandie

    Institutional partner on the French side, including legal and regulatory support.

    Official site

  • FedBioMed — INRIA / Université Côte d'Azur

    Federated-learning framework and methodological partner (in-kind).

    Official site

With thanks to

CANAL-AI would not exist without the support of the people and institutions below.

Pilot funding

SECOM is our first supporter — thank you. CANAL-AI is designed to grow into a multi-site federation, and we are actively seeking additional funders to join them.

Community

Patient and public involvement

CANAL-AI is a methodological pilot using retrospective hospital data — no patients are recruited or examined for this work. As the consortium grows toward prospective validation, a patient advisory group jointly convened in France and Japan will be established to review lay summaries, the cardio-oncology endpoint definition, and any future deployment pathway. Plain-language summaries of the project are available in English, French, and Japanese on this site.

Open science

Open data and FAIR outputs

CANAL-AI commits to FAIR principles for every output that does not require patient-level data. The OMOP-to-FedBioMed ETL, the APPI–GDPR operational framework, and the analysis code will be released under permissive licences (CC BY 4.0 for documentation; MIT for code), each with a persistent DOI via Zenodo.

Patient-level data remain at each hospital under the federated-learning architecture and are never released. A Data Management Plan (DMP) following the Horizon Europe template will be published before the first federated training round, and a project-level persistent identifier will be registered at award start.

Funding

Fund the pilot

The pilot runs 2026–2028 and has a target budget of roughly €450k over 24 months. SECOM has committed the first tranche; additional funders are warmly welcome.

What funding unlocks

  • Site onboarding

    OMOP mapping, FedBioMed node setup, IRB processing, staff time at each partner hospital.

  • Compute

    Training rounds, secure aggregation infrastructure, differential-privacy evaluation.

  • Open-source blueprint

    Publication of the APPI–GDPR blueprint and the OMOP-to-FedBioMed ETL (CC BY 4.0).

Funders are credited on this site, on the blueprint publication, and as named supporters on the methods paper.

Talk to us about funding

Partnership

Become a partner hospital

CANAL-AI is designed to grow. If your hospital has a data warehouse and is willing to participate in federated learning, you may be a natural next node. Patient data never leaves your institution.

What a partner site needs

  • Access to a hospital data warehouse (or an equivalent structured EHR extract).
  • Willingness to harmonise relevant variables to OMOP-CDM.
  • Local IRB support and the willingness to participate in federated learning.
  • A workstation capable of running a FedBioMed node. No patient data ever leaves your institution.
Talk to us about joining

Contact

Reach the team

For scientific, institutional, or funding enquiries, fill in the form below — your message will reach the three principal investigators in one email.

Your message is delivered to the team through Web3Forms. The team's email addresses are never shown on this site.