What I actually work on
Two fields, and neither one is useful without the other. This page is the explanation. If you want what I have built with them instead, that is on the home page.
01Federated learning
The model moves, the records stay
This is my main field. One model, trained across many separate places at once. The data never leaves home. The name for it is federated learning.
- No two places hold the same kind of patient.
- One shared model fits none of them well. I let it adapt to each place, while it still learns from all of them. That is what closed the gap above.
- Sharing is treated as all or nothing.
- I introduced the degree of federation. A group picks a point in between, and moves along it as trust grows. Nobody has to walk away.
- Federation alone is not privacy.
- The records stay home, but gradients can leak, and membership inference is real. So I treat the governance agreement, and whatever you add on top, as part of the method rather than paperwork around it. Anyone who says federated means safe has not read the attack literature.
- Most teams have no systems engineer.
- A method that needs one does not get used. So the tooling has to run without one. I treat that as a design constraint, not an afterthought.
02Real world data
From records to evidence
And this is what it runs on. What ordinary care leaves behind, not what a trial sets out to collect. The field calls it real world data.
Visits, laboratory results, imaging and prescriptions arrive on their own schedule. Which is to say, on no schedule. Each source alone gives a weak answer, and no two of them agree.
Put them together and one answer holds across all of them. Getting from the top of this picture to the bottom is what the field calls real world evidence. That is the part I spend my time on.
- Every source records things its own way.
- Two registries can disagree on what a visit is. Making them line up is most of the work.
- It was never collected for research.
- Gaps and bias are built in. So the quality checks belong inside the pipeline, not in a caveat at the end of a paper.
- Rare conditions leave every cohort short.
- No single place has enough patients to see a pattern. The only route is to combine what each of them knows separately.
03Approach
What I hold to
Four things I do not bend on. They decide what I build, and what I refuse to build.
Privacy rules are the specification, not the obstacle.
If a method only works once the rules are relaxed, it is not a method. It is a wish. I start from what is allowed, and build inside it.
If it needs an engineer, it will not get used.
The groups who need this most have no systems team. I judge my own work by whether a small group can run it without me in the room.
Evidence that stays in a paper has helped nobody.
Patients, clinicians and funders belong in the work, not in the audience at the end. That is why nearly everything I build is open.
Messy data is the only data most people leave.
Ordinary care records are incomplete and biased. They are still the only trace most people leave behind. Throwing them away is easy. Handling them well is the job.