FLIP demonstrated at AWS Public Sector AI Symposium

AWS Public Sector AI Symposium London — 9 September

Alexandre Triay Bagur, Senior AI Engineer, presented NHS Federated Learning: Training AI Without Moving Patient Data in the healthcare track of the AWS Public Sector AI Symposium at Convene 133 Houndsditch, in partnership with OneLondon.

The first half of the session set out the scale of the underlying data engineering problem. A single London NHS Trust runs eight or more structured clinical systems alongside separate stores of narrative documents, letters and imaging — together holding 20 years of real-world data on 5.2 million patients, including some 10 billion rows of clinical concepts and 300 million documents. Aggregated across the capital, this becomes a living data fabric covering 11 million patients, built on shared code and common data models rather than one-off extracts.

Much of that clinical detail is locked in free text, and the session briefly introduced MESA (Medical-concept Extraction with Schema Alignment), the AI Centre’s open-source toolkit for fine-tuning clinical language models on AWS, now in active deployment across five cancer centres. MESA is available at github.com/londonaicentre/MESA.

Radiology is the harder case: 17.2 million imaging studies spanning two decades, at petabyte scale and impractical to centralise on any realistic timeline. This is what FLIP was built for. Rather than moving images, FLIP places a node alongside a secure enclave inside each Trust and coordinates training across those nodes from a central hub. Apache 2.0 licensed and built on AWS with NVIDIA FLARE, Flower and deepc, it is deployed across Guy’s and St Thomas’, King’s College Hospital, King’s College London, the London Secure Data Environment and Bangkok Dusit Medical Services in Thailand.

The session walked through a worked example: federated fine-tuning of the Ark+ chest X-ray foundation model across the UK and Thailand. With the pre-trained backbone frozen, only a small classification head was trained — a useful illustration of how little needs to cross a network boundary to do something clinically meaningful.

Alex closed on security and data sovereignty, which FLIP enforces structurally rather than by policy alone. Each Trust approves every project separately; only aggregate statistics, filtered model updates and task status ever leave the hospital; and Trusts poll outbound only, so the hub has no network route in. The National Data Opt-Out is applied where the data is prepared, so it propagates to every query, project and model. FLIP is a Health Research Authority REC-approved Research Database overseen by a Data Allocation Committee, and has been assessed against Cyber Essentials, the NHS Data Security and Protection Toolkit, the NCSC Cyber Assessment Framework and the SATRE specification for federated analytics.

View Alex's presentation slides via this link here.

What comes next

Work now underway includes extending deployable node infrastructure-as-code to further cloud providers, building language model capabilities directly into node infrastructure for automated labelling from radiology reports, standing up new nodes over the coming six months, and onboarding new projects and strategic partnerships in London, the UK and overseas. Trained models are packaged for deployment through integrated radiology platforms, closing the loop from federated training to clinical evaluation.

FLIP is open source. Documentation is at londonaicentreflip.readthedocs.io and the code at github.com/londonaicentre/FLIP.