MEDIRA Health
An AI assistant for clinical documentation
An assistant that drafts discharge letters from structured findings — entirely inside the hospital's own infrastructure.
- saved per clinician per week
- 4.5 hsaved per clinician per week
- approved without correction
- 92%approved without correction
- patient records leaving the hospital
- 0patient records leaving the hospital
The situation
An internal survey showed clinicians spending several hours a week writing discharge letters — text largely assembled from findings that already existed in structured form. At the same time the constraints were hard: patient data must not leave the hospital, and a model that invents plausible-sounding diagnoses is worse than no tool at all.
How we worked
We built the evaluation harness before the product: 400 anonymised historical cases with clinician-approved reference letters as the benchmark. Only once it was clear which combination of model, context and prompting handled those cases reliably did we build the interface. The assistant only ever proposes — every letter is released by a human, and every statement traces back to its source finding.
- An evaluation set of 400 anonymised cases underpinning every decision
- Retrieval strictly limited to the findings of the case at hand
- Every generated sentence links back to its source in the record
- Runs in the hospital's own data centre; no data leaves the building
Outcome
The assistant has been in use across two departments for two quarters. Time saved runs to several hours per person per week — but what actually won acceptance was being able to demonstrate that it does not state anything absent from the record.
„The decisive thing was not that it works. It was that we could prove how well it works — without that, it would never have been approved.“