Context
Preventive health has a delivery problem, not a knowledge problem. People do the screening, receive a PDF of reference ranges, and take no action — because a page of numbers is not a decision. Meanwhile the same person’s results from two years ago sit in a different system, so the one thing that carries real signal, the trend, is invisible to everybody.
What I did
- Built the platform around longitudinal data rather than single reports. A value inside the reference range that has moved sharply since last year is more interesting than a value slightly outside it that has not moved at all.
- Wrote for two audiences from one data model: a clinician view that keeps the full picture, and a patient view that says what changed, what it likely means, and what to do next.
- Treated data protection as an architectural constraint rather than a compliance checkbox — least-privilege access, separation of identifiable data from analytics, and audit trails on anything clinical.
- Kept the model conservative on purpose. In preventive health, a system that cries wolf gets switched off, and a system that gets switched off prevents nothing.
- Grew the engineering team and the delivery practice around it, since a health platform lives or dies on whether it can be safely changed for years.
Outcome
Routine screening stopped being a one-off report and became a tracked trend. Clinicians get the full record; patients get a plain-language version of the same thing, which is the version that actually changes behaviour.
What I’d do differently
I would bring clinicians into the design earlier and more often. We built technically sound features that clinicians did not trust, and trust turned out to be a design problem — how a finding is worded, what uncertainty is shown, what the system declines to say — not an accuracy problem. Retrofitting that costs more than designing for it.