The problem with clinical AI.
Sleep apnea affects over 175 million Europeans, and most of them are undiagnosed. The data that could train a world-class detector is already in hospital archives — polysomnography recordings of patients who slept all night while instruments read their breathing, heart rate, and oxygen saturation.
But that data cannot move. GDPR Article 9 treats it as special-category data. National law adds more. Ethical boards add more. Even within a single hospital group, moving records between sites is a multi-month legal project. The practical result: clinical AI models are either trained on small single-site datasets, or they never get built.
It wasn't the first time anyone tried federated learning in healthcare. What made Ambiti8n unique was that it was a practical, real-world system actually put into operation — not a research demo.
Maria Rossi · Clinical AI Lead · Ospedale di BariWhat we built.
Using the Ambiti8n AI Factory, two hospital data vaults in Torino (Italy) and Warsaw (Poland) trained a shared CNN-LSTM model without sharing records. Each site runs an openEHR + OMOP pipeline locally. A federated orchestrator at Gdańsk Tech's TASK HPC centre coordinates training rounds. Gradients travel. Patients don't.
- The model: A CNN-LSTM that classifies apnea/hypopnea events from multi-channel polysomnography signals, trained via federated rounds using FedX.
- The data: Real clinical records, extracted locally through openEHR archetypes and harmonised to OMOP CDM. Never leaves the hospital.
- The serving: Packaged with AI-Act-ready documentation and deployed to CloudFerro's Sherlock sovereign inference platform in Poland.
The assembly line — what we actually composed.
None of the individual components are new. What's new is that they are pre-integrated, governed-by-design, and reusable. Tomorrow's consortium doesn't start from a whiteboard — they pick from the factory floor.
Built by eight partners across five countries.
Federated learning and inference
Early app
Sovereign cloud and orchestration
Sovereign cloud
Federated HPC
LLM · Compliance
Model catalog
ObservabilityWhat this proves.
The IPCEI-CIS cloud-edge continuum isn't a vision document anymore. It's a working production environment. A consortium can take real, legally sensitive clinical data in two countries, build a jointly trained model on it, serve that model on an EU-sovereign cloud, and produce an AI-Act-compliant audit trail — in four months, using reusable components.
Healthcare was the first vertical because it's the hardest. The same assembly line is now ready for manufacturing, finance, and energy — sectors where the data-sovereignty problem is structurally identical.