Biomedical AI has to be secure by construction
In clinical data, privacy and provenance aren't features you add later — they're the foundation.
It is easy to build an impressive model on clinical data and only afterwards ask who was allowed to touch it. By then the answer is usually "too many people, in too many copies." In biomedical AI that is not a bug to patch — it's a design failure.
Secure by construction means the safe path is the only path. Access, consent and provenance are decided at the foundation, so a careless query simply cannot return what it shouldn't.
Three things I hold to
- The data layer enforces permission, not the application. Screens change; guarantees shouldn't depend on which screen you came through.
- Every read and write is accountable. If you can't say who saw a record and why, you don't really have consent — you have hope.
- Provenance travels with the data. A prediction you can't trace back to its inputs is not a result you can defend to a clinician.
Why it matters for the work I do
The Biomedical Intelligence Systems Lab works where biomedical data meets intelligent software. That phrase only means something if the software is trustworthy — and trust is cheaper to build in from the first table than to retrofit after a breach.
Models get the headlines. The scaffolding around them decides whether any of it is usable in a real clinic.