Work

Selected work

I did this work at KU Leuven and Hasselt University, as a doctoral researcher from 2020 to 2024 and as a postdoctoral researcher since.

For some studies, one accessible database is enough. Others are harder: the records cannot be pooled, each site has only a few patients, or patients differ from one place to the next. Those are the studies I work on.

Stories

Illustrative values. To balance the two groups, studies use a propensity score: a model of who is likely to get which medicine. The mandate tests whether my methods keep it reliable.

What I am working on now

Can a drug safety study stay fair when each site has few patients?

Once a medicine is on the market, safety studies compare the people who take it with people who take another one. That is only fair if the two groups are alike. With few patients per site, the standard way of making them alike can fail, and it fails most often right after approval, when few people have taken the new medicine yet.

My Innovation Mandate takes the methods from my multiple sclerosis work to this problem. The data are Optum's health records and insurance claims from the United States. In them, payers, providers and regional health systems act as the sites.

Not shown yet: Whether the methods carry over. In my multiple sclerosis study each patient was a few dozen clinical measures, while in claims data a person is a very long list of billing and diagnosis codes.
How I will judge it: Step by step against what is done today: whether the two groups end up alike, and whether that holds without adding bias.
Who this is for: Teams that run safety studies across sites, and anyone whose network is made of many small ones.

An Innovation Mandate: a personal grant from the Flemish government, with Johnson & Johnson Innovative Medicine as industry partner.

I lead it · Awarded 2026

How it came about

COVID-19 and multiple sclerosis

Keeping four registries in a worldwide study

Four registries could not join a worldwide study on multiple sclerosis and COVID-19, because they were not allowed to share patient records with anyone.

The route I built added 45.5% more records. The results were used for worldwide COVID-19 advice for people with multiple sclerosis.

Architect and first author · JMIR Medical Informatics, 2023 · With the MS International Federation and the MS Data Alliance · 2021 to 2023

Read the story

Three ways into one study

Records represented in the study11,284

All three routes. Records from each are represented in the study. Only one route kept the records where they were. Press a route to see what it carried.

Counts at the time the paper was published. JMIR Medical Informatics, 2023.

Disability progression

Two years ahead, will this person get worse?

Patients with multiple sclerosis differ from country to country, and one shared model can miss those differences. I designed an architecture, AdaptiveDualBranchNet, in which each country keeps part of the model as its own while the rest is learned from all of them.

Personalization improved AUC-PR, a measure of how well it finds the few who get worse, by 31% over the same method without it, in simulated federation. 26,246 patients, 146 centers.

Conceived and led, first author · npj Digital Medicine, 2025

Read the story

FLkit

An open guide for teams starting a study across centers

What a team needs to start is spread over framework documentation, legal templates and people's heads. We gathered it into one guide, in the order the decisions happen.

Public and maintained, with 29 contributors and support from ELIXIR Belgium.

I lead it · Live · 2024 to now

Read the story

All four centers send their records0.812

Two send, two keep theirs0.825

All four keep theirs0.846

The degree of federation on a public benchmark: 740 records across 4 centers, logistic regression, area under the curve. Bars start at 0.5. Not a clinical result.

A project I stopped

FL4E, a framework I retired

FL4E is a framework for federated studies in health, built so research teams would not have to start from zero each time. Its main idea is the degree of federation: some centers send their records and some keep them, in the same study.

My own next study, on disability progression, did not use it. I retired it, and wrote a paper about why.

What I would do differentlyBuild it as small pieces that a team can pick up one at a time.

The paper on why it went unused names five barriers to reuse. Medical Informatics Europe 2026.

Built it, first author · JMIR Formative Research, 2024 · Code · Retired

Read the note

In brief

Other work

FLoRankPrototype. Not yet run on a real data network.

Each site wants a model shaped to its own patients. A full private copy for each is expensive, and it stops them learning from one another. FLoRank keeps one large shared part and adds a small private part per site. For this work I received an ECTRIMS personal grant for scientific quality, in 2025.

ECTRIMS 2025 abstract

A national parcel network2017 to 2019. Simulation study.

A simulation of Iran Post's parcel network, to decide where the sorting hubs should go. The best layout was 15% lower on projected transport cost, in simulation.

Publications, talks and code

22 entries, each labeled as a journal paper, abstract, chapter, preprint or thesis. Talks at Roche, Flower, ECTRIMS and Medical Informatics Europe.

Publications Talks GitHub Google Scholar

What it runs on

Python, PyTorch and Flower for federated training. scikit-learn and pandas for the study code. R and SQL for registry work. OMOP, through OHDSI, for a common data model. Docker, Kubernetes and continuous integration so a site can install and run the code.

Let's talk.

If your question depends on real world health data, I probably want to hear about it.

Email me

A greeting in each language I speak, where it lives.