Postdoctoral researcher in AI for healthcare at the University of Twente, research affiliate at UHasselt. PhD in computational neuroscience, KU Leuven, 2023. Before the neuroscience there was a lot of compilers, type systems and web performance, and some of that is still lying around here.
My research line is the self-explaining hospital: clinical models that are interpretable by design, and that travel between institutions instead of forcing patient data to travel. Three things run through most of it:
- Structured tensor methods. Block-term decompositions as the model itself, so the fitted factors are the explanation rather than a post-hoc story about it.
- Federated and privacy-preserving learning. Horizontal and vertical federation, differential privacy, and what you can actually prove about the gap to centralised training.
- Clinical signals and imaging. Longitudinal MRI, ECG, ECoG and EEG, multi-omics, and the boring-but-decisive part: honest validation.
Research code. One repository per project, each the public deposit for a paper. Same standard throughout: raw inputs in, every reported number, table and figure out, from one documented command. Block-term neural operators and their theory, vertical federated tensor regression, differential privacy for federated coupling, interpretable ECG models for atrial fibrillation and mortality, cross-subject decoding from intracranial recordings, tensor methods for unaligned MRI.
Tools and teaching material. Things built to be used by other people: the ELIXIR Federated Learning Kit, a hands-on federated learning tutorial, a CV generator that builds LaTeX from YAML, a course calendar exporter, my website.
Things I build because I like building. A small dynamic scripting language, a WebGPU shooter that fits in one HTML file, a visual event-based programming IDE. Also the theses and honours projects from a previous life in programming languages, kept as-is.