Applied machine learning scientist. Real world evidence, and health data systems

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

I make evidence out of health records that cannot be moved.

The records sit in many places, in many countries, and the rules keep them there. So I send the model to the data instead. That is called federated learning. It is how I do it. It is not what I do. And I keep it simple enough that a small team can run it without me.

Research handed me the question. Building the systems that answer it showed me where it breaks. In September 2026 I take that into a company, on a two year mandate with Johnson & Johnson Innovative Medicine.

See what that produced Bring me a data problem Read the CV

Ashkan Pirmani, who makes evidence out of health records that cannot be gathered in one place, at KU Leuven and Hasselt University
  • Doctorates Dual PhD, electrical engineering at KU Leuven and biomedical sciences at Hasselt University. Both defended in 2024 2024
  • Next Won a two year grant to take my own work out of the university and into a company. It pays for one researcher at a time 2026
  • Award Won a personal grant at the largest multiple sclerosis conference in the world, for the method behind the paper below 2025
  • Paper Showed in simulation that records kept in their own countries can beat the same records pooled. npj Digital Medicine. First author 2025

From a question to a decision, and where I sit in it

I led it Co-owned Outside my decision

  1. The question A clinician asks it. I make it answerable.
  2. The same patient One definition, across sites that record it differently.
  3. A permitted route Custodians and ethics boards say yes or no. I design the route they can say yes to.
  4. Running at each site Their machine, their staff, their IT people.
  5. The data problems Found by looking with the people who entered it.
  6. The model The smallest part, and rarely what stops a study.
  7. The evidence Designing the study, running it, and writing what came out.
  8. The clinical decision Clinicians, guidelines, regulators. Never me.

01Result

Real patient records. Federation simulated

Two years ahead, will this person get worse?

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. All three answer the same question, 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.

0.50, a coin toss1.00, perfect

Federated, and fitted to each country. 0.8398
All data pooled in one place. 0.8092
Federated, one shared model for everyone. 0.7840

Standard deviations were 0.0019, 0.0012 and 0.0019, across ten repetitions. All three rows are the paper, table 1.

What the top bar beats, and by how much. Take the same setup and switch the country by country fitting off. Against that, 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. In the usual names, ROC-AUC and AUC-PR. Against the middle bar, everything pooled in one place, it is 3.8% better. That switched off version is not the bar at the bottom.

Ashkan Pirmani presenting at a lectern. The slide behind him reads: Good for all. Not always good enough for one. A general purpose framework can be technically functional, and still practically inaccessible, for the very next project in the same team.
The whole argument, on one slide. A model that is good for everyone on average can still be no use to any single country. That is what the chart above measures, and it is what I have been saying out loud since before I could prove it.

Measured on

26,246patients
283,115clinical episodes
146centers contributed
32country groups

The decision I owned. Stop training one model for everyone. Fit each country instead, with a network built to carry a shared half and a private half at once. I conceived it, ran it, and wrote it. One of seventy three names on the paper, and the first.

The middle bar is the ceiling everyone assumes. It is not. And plain federated averaging, the bottom bar, really is worse than pooling, which 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.

02Work

What I built, and what it answered

Start here if none of this is your field. A health question needs more patients than any one place has. The records exist. Reaching all of them is the problem. So the question stays open. Everything below is a way to answer it anyway.

Most real world evidence does not work this way, and that is fine. A company licenses a big database, it sits in one place, and it is already legal to query. Try that first. It runs out in three places. When the patients are rare. When they are outside the United States. When the thing you want to measure was never written into a claim. That is where my work starts.

Every project below has a figure you can press. They are all on one page

Global Data Sharing Initiative

Ran for real, across registries Architected. 2023

This one produced evidence that was used. At the start of covid, nobody knew whether multiple sclerosis drugs made the illness worse. No single country had enough patients to answer it. So the MS International Federation and the MS Data Alliance opened a worldwide collection. The answers that came out of it went into the global covid advice given to people with multiple sclerosis. The MS International Federation says so itself.

The decision I owned. Two ways in existed. Send your data, or fill in a form. Any registry that could do neither was simply out. So I built the third, and those registries run the analysis themselves and send back only the result. It brought 45.5% more records than the other two routes had collected.

The hard part was not the code. Eighteen registries, in different countries, under different laws, each with its own custodian and its own IT people. All of them had to agree on what counted as the same thing. Then each had to run it on their own machine. Getting that to happen was most of the work. Whether a study like this can run at all is decided there, long before anyone fits a model.

Switch the three routes off and on

18 registries and 11,284 records at publication. 3,527 of those records came through my route. Five registries use it now. The paper counts four, because the fifth came later. First author on the paper that describes how it was built. JMIR Medical Informatics 11(1):e48030, 2023.

Every number above came out of the paper cited beside it.

Three more, and the decision in each

The two above have their own sections. These three do not. So here is the constraint, what I decided, the paper to check it against, and the figure you can press.

ProjectThe decision I ownedWhat changedReceipt
FL4EPrototype, run on Azure Let every center choose for itself how much it takes part, instead of making it all or nothing. Half the sites can stay centralized and you keep 97.5% of the score Paper Turn the dial
FLoRankPrototype, first run Q4 2026 Fit each site too hard and there is nothing left for the next site to inherit. Keep the shared backbone strong and make the private piece small. A site that joins later starts from something, instead of from nothing Paper Run the rounds
FLkitPublic and live We kept watching teams rebuild the same first three months. Give them a template to start from, and a way in for each kind of person on the team. A public resource, supported by ELIXIR Belgium, that other teams run for themselves Visit FLkit Preprint See inside it

Publications

03Networks

None of this works alone

Five people crowded around a laptop at Hack4Health in Flanders Expo, Ghent. Ashkan Pirmani is seated at the left, watching the screen.
Hack4Health, Flanders Expo, Ghent. This is what the work actually looks like most of the time. Five people, one laptop, and nobody in the picture shares a first language.

Who I build with

The people I am about to do this with, and the ones who got me here. At Johnson & Johnson Innovative Medicine I will be working with Patrick Ryan, who leads Global Epidemiology there and chairs OHDSI. On the academic side, Liesbet Peeters at Hasselt and Yves Moreau at KU Leuven.

The networks I work in

The MSBase registry, where the records behind the study above came from. 146 centers, grouped into 32 countries for that study. OHDSI Belgium, and the organizing team for OHDSI Europe 2025. ELIXIR Belgium supports FLkit. It all runs on Flower.

The people I sit between

Clinicians, epidemiologists, registry custodians, engineers and lawyers, across more than thirty countries. None of them share a vocabulary. The industry has names for the work of holding them together: harmonization, phenotype definition, and data quality. It is where a cohort either becomes trustworthy or does not.

04Contact

Get in touch

If you have data that cannot move, or a question that needs more patients than one place holds, I would like to hear about it.

Email
AlsoLinkedIn BackgroundFull CV CodeGitHub