From Silos to Synapses: Why Federated Learning is the Network Healthcare AI Has Been Waiting For

From 24 to 26 February 2026, we hosted the AI Radiology summit, bringing together selected industry-leading experts, academics, clinicians, and policy advocates for 3 days of intense collaborative discussions to shape the future of AI in healthcare.

The conversation around artificial intelligence in healthcare is no longer just about the algorithms themselves. It’s about the networks that will power them. This was the resounding theme of the panel featuring Dr. Joe Zhang, Joanna Wills, Sam Khashman, Henriettae Ståhlbrandt, and Bryce Travers, moderated by Dr. Christina Triantafyllou.

The room was filled with great expertise and a shared sense of purpose. The conversation laid bare a profound paradox: we have plenty of health data that could deliver immense value to patients and the health care system, but it’s scattered across thousands of disconnected silos, each with its own governance, incentives, and way of doing things.

As one Bryce Travers aptly put it, “My trusts have no real incentive to share data with each other.” Henriettae Ståhlbrandt highlighted the cultural hurdles, noting the need to “trust the private companies that they also have the greater good in mind,” underlining that these aren't purely technical problems; they are problems of trust, incentives, alignment, and infrastructure.

The Two Networks We Need to Build

The discussion delineates the two critical networks we lack for AI to flourish:

1. The Human Network: The network of relationships among clinicians, data scientists, and leaders. It’s built on shared challenges, such as reducing backlogs, and a common goal: improving patient care. Collaboration often starts where close working relationships already exist. You can’t just plug in a piece of software; you have to bring people along on the journey.
2. The Data Network: The technical fabric that enables information to flow securely and collaboratively between organisations. The panel recognised its immense potential: connecting data from diverse populations across regions to build more robust, representative, and equitable AI models.

But building this data network is where the friction lies. We heard about the high "startup costs" of collaboration, the lack of immediate advantages for individual trusts, and the sheer complexity of navigating the commercial and governance landscape of a system as fragmented as the NHS.

This is where the conversation takes a turn that feels incredibly timely for us at the AI Centre.

A New Kind of Network for a New Kind of AI

Throughout the panel, we heard the underlying requirements for a solution. We need networks that can:

• Learn from heterogeneity: Not everyone can or should standardise their scanners and protocols overnight. We need systems that can work with data as it is, in the real world.
• Respect local governance: Each trust has a duty to protect its patients' data. Any solution must put control back in the hands of the data owners.
• Align incentives: A model in which one party bears all costs and another reaps all the rewards is unsustainable. We need a framework for a true public-private partnership that shares value.
• Start from a place of trust: The technology must be transparent and secure enough to overcome the inherent differences between the public and private sectors.

This is the very blueprint for the Federated Learning and Interoperability Platform (FLIP), which we have been developing at the AI centre and have just made public during this event.

FLIP: Turning the Panel’s Vision into Reality

It was therefore incredibly exciting that, on the very same day as this panel, we were able to announce the launch of FLIP (Federated Learning and Interoperability Platform), a fully open-source platform developed by the AI Centre in partnership with KCL, GSTT, deepc, OneLondon, and Flower Labs.

FLIP is, in essence, the data network infrastructure the panel called for. Instead of moving patient data to a central location to train an algorithm, FLIP moves the algorithm to the data. The data stays within the hospital's secure walls, behind its own governance firewall. Only the anonymous model updates—the "learning"—are shared and aggregated to build a powerful, generalisable AI.

This directly addresses so many of the core challenges raised:

• No incentive to share? With FLIP, trusts don't share data in the traditional sense. They retain full ownership and control, eliminating the primary risk. The incentive to participate shifts from "giving something away" to "gaining a better model" trained on a much wider, more representative dataset. Also, the trusts get a platform to showcase the value of their data, giving them more leverage on negotiations with interested third-party entities.
• Standardisation is a barrier? FLIP is designed for a non-standardised world. It can learn from diverse protocols, scanners, and populations across trusts without requiring any changes. FLIP provides layers of abstraction and algorithms for data harmonization, effectively creating a mirror of the trust’s data that can be processed homogeneously from the user’s point of view.
• Can public-private partnership benefit both? FLIP provides the secure, transparent, and equitable infrastructure for these partnerships. A private company can use the platform to collaboratively train a model across multiple NHS trusts. The NHS gets a cutting-edge tool and a stake in its development; the company gets a validated, robust product. The value exchange can be agreed upon by the institutions, and FLIP sits in the middle, providing a neutral, trustworthy bridge between the parties. They don’t need to build their collaboration infrastructure from scratch and can benefit from a transparent, open-source platform to intermediate and ensure security.

By providing an open-source, scalable platform, we empower the human networks and give them the tools they need to build the data networks of the future. There are paths for AI to deliver huge value for patients without compromising their security and autonomy. We invite you to explore FLIP and join us in building the future of healthcare AI, together.

Find out more:

• https://github.com/londonaicentre/FLIP
• Platforms | AI Centre for Value Based healthcare