Cohort matching at a regional hospital (MedTech)
A digital twin of patient cohorts, with HIPAA-grade provenance and a decision engine that refuses silent writes.
- ·94% precision on cohort matching against clinician-annotated holdout
- ·7× speed-up on RWE feasibility studies vs the previous manual flow
- ·Full chain-of-custody for every cohort revision
The partner’s analysts assembled RWE cohorts by hand, then defended them for weeks. The twin ingested structured EHR, lab, and claims data, resolved patients into typed cohorts, and returned a decision with provenance on every proposed write.
What NR changed
Every mutation became a first-class event — PatientCohort v9 → v10 with actor, executor, policy, and citations — instead of a SQL delta. Clinician reviews moved from email to the trace itself. The decision engine refused silent overwrites and escalated differences to the reviewer.
Caveats
The TRL claim here is preliminary. The partner is pseudonymised. Data shown on the site is synthetic and does not come from the production environment.