I answer medical questions with data I am never allowed to see.

Real patient records. Federation simulated

Keeping the data apart costs you. Fitting each place pays it back

Privacy is supposed to cost you. Keep the data apart, get a worse model. That much is true, and the bottom bar is what it costs. The top bar is what you get back when you let the model fit each place instead of forcing one model on all of them. What is measured is one question: two years ahead, will this person get worse? All three models answer it on the same patients. The score is area under the curve, where 0.50 is a coin toss and 1.00 is perfect.

How it was run. These records were already gathered into one research registry. I split them back into country groups and made each group train on its own, passing only the model between them. So it measures what federation does to the score, not a live federation.

What the top bar beats, and by how much. Against the same setup with the country by country fitting switched off, it is 7.2% better at telling apart who will get worse from who will not, and 31% better at finding them without flagging half the group. Against the middle bar, everything pooled in one place, it is 3.8% better.

Those first two numbers compare the top bar with its own switched off version, which is not the bar drawn at the bottom. The bottom bar is the plainest way to do this and it is drawn because it is where everybody starts.

Measured on

The middle bar is the ceiling everyone assumes. It is not the ceiling. What made the difference was letting the model fit each site, instead of forcing one model on all of them. Plain federated averaging, the bottom bar, is worse than pooling. That is why the field keeps saying privacy is expensive.

One caveat I will make myself, before you make it for me. The pooled model is a single global model. The winner is fitted to each country. So what won here is that fitting, not the sharing on its own. What the sharing did was make the fitting possible without gathering the records first. That is the claim I stand behind, and it is the more useful one. The next thing to prove is that it holds when the groups are real and separate, not split apart after the fact. Pirmani et al., npj Digital Medicine 8(478), 2025. Code is public.

What I built, and what it does

Start here if none of this is your field. A model normally learns by having all the data in one place. Mine does the opposite: it travels to each center or registry in turn, learns there, and moves on. The records never go anywhere. Everything below is built on that one idea.

Two rules run through all of it. If a site needs its own engineer before it can begin, it will never begin. And if it needs me in the room to run, it is not finished.

FL4E

Working prototype, run on Azure Drag it, or use the arrow keys Designed and built. 2024

Sharing is treated as all or nothing, so groups that cannot do all of it walk away. I introduced the degree of federation: every center decides for itself whether its data leaves, and the middle is a real place to stand. Turn the dial.

Fully federated 0.846

A benchmark, not the clinical study. Public data, small on purpose, so the shape of the effect is easy to check. Fed-Heart-Disease, 740 records across 4 centers, logistic regression, federated averaging, five repeats. Area under the curve. JMIR Formative Research 8:e55496, 2024.

Global Data Sharing Initiative

In use across registries Press a stream, or tab to it Architected. 2023

Two ways in already existed. Send your data, or fill in a form. Any registry that could do neither was simply out. I built the third one, the federated route, so those registries could run the analysis themselves and send back only the result. Five registries came in that way. They added 45.5% on top of everything the other two routes had collected. In the first year of covid, with a rare disease, every record was worth having. Switch a stream off to see what one way in would have cost.

The worldwide multiple sclerosis and COVID-19 effort. 18 registries, 11,284 records in total. JMIR Medical Informatics 11(1):e48030, 2023.

These two are real in different ways. The first one we built and tested ourselves. The second was run by eighteen registries on their own machines, for a study that was published. Every number under both came out of the paper cited beside it.

Publications

None of this works alone

I follow one thread the whole way along. What a patient leaves behind, how it gets out of the system holding it, who is allowed to touch it, what the clinician actually meant when they typed it, the model, and then the evidence that goes back to a patient. Most people work on one link in that chain. My work has had to connect all of them, and the hard part was never the model.

Get in touch

Email