Know what's missing before the visit.
MedLineage turns a patient's scattered records — PDFs, HL7, FHIR, lab files, in five languages — into one source-cited history for the specialist, and names the records still missing while there is time to request them.
Bring one referral. We run it de-identified and name the records missing from it before your next clinic day.
Prepares records for clinician review. Does not diagnose, does not recommend treatment, does not replace clinicians. Best-effort de-identification, not certified anonymization.
Check it yourself
100% provenance coverage · 1.00 extraction F1 on our published synthetic benchmark — raw JSON public, drift fails CI.
02What the clinic gets
The visit starts with the history already assembled.
Key findings
- Haemoglobin 9.8 g/dL on 2 Apr 2026, down from 12.4 g/dL on 9 Sep 2025[doc-3 p.2]machine-extracted
- Colonoscopy 14 Mar 2026: sessile lesion, 12 mm, ascending colon[doc-1 p.1]clinician-verified
Missing records
criticalHistology report for the biopsy taken on 14 Mar 2026
Not present in any uploaded document. It is the record a specialist asks for at this visit.
Every line above resolves to a document, a page and the language it was written in. Nothing that does not resolve reaches the packet.
A summary is not what the records desk is short of.
- What AI vendors promise
«AI summarizes the patient»
What the records desk is actually short ofWhich of the six records for Thursday's colonoscopy is not here
What MedLineage doesNames the missing record by type before the visit, split into critical, recommended and contextual
- What AI vendors promise
«Chat with your medical records»
What the records desk is actually short ofA number a clinician can check against the lab report itself
What MedLineage doesEvery value carries the document, page and source language it came from
- What AI vendors promise
«One-click patient summary»
What the records desk is actually short ofNot to re-read forty pages at the follow-up
What MedLineage doesCompares each new upload against what the case already knewLive
It reads what the clinic already has
- FSE / CDA2 (IT)
- HL7 v2
- FHIR R4
- SMART on FHIR
- Lab CSV
- PDF, 5 languages
The Italian Fascicolo Sanitario Elettronico is read directly: the signed CDA2 document inside an FSE PDF is parsed deterministically, with no model call. No data migration, no parallel system. Every supported format
03Verification that compounds
Every clinician sign-off makes the layer harder to copy.
When a clinician verifies a fact, the verdict is signed into a tamper-evident ledger and reused for every future read. Demand for a fact prioritizes its review; verified context accrues.
- 01
A specialist or an agent relies on a fact
- 02
A clinician verifies it
- 03
Every future agent reads verified context
- 04
And the state carries it forward: a verified fact stays verified as the record grows, with its knowledge age tracked over time.
Shipped behavior, demonstrated on synthetic cases — flag-gated per deployment.
Numbers that reproduce from a pinned benchmark.
291 synthetic gold facts · 5 languages
every recovered fact cites its source document
no synthetic case tripped the safety net
Synthetic multilingual gold set — not a clinical study. The scoreboard is committed to the repo and pinned in CI: any drift fails the build. gold_set.jsonl · scoreboard.json
Our first clinical pilot runs at Sanity Health — a cofounder-owned clinic, disclosed. Sanity Health is a small private clinic in Milan with around 500 patients, owned by a MedLineage cofounder — so read it as product development, not independent validation. The clinic pushes its inbound clinical records straight into MedLineage Connect and we turn them into a source-cited patient-context layer that its clinicians review and attest to, with every finding tracing back to the document, page, and source language it came from. Pilot in progress. MedLineage prepares records for clinician review — it does not diagnose, recommend treatment, or replace clinicians.
How the records are handled
- Records sent for extraction never become training data — the LLM sub-processor is contractually barred from training on them. Anthropic Commercial Terms, §B
- GDPR: article to code to test, with the gaps named
- Records flow under an opaque patient handle. De-identification is best-effort, not a certification.
- Each clinic's records sit in its own tenant partition — separation is by partition key, not by encryption.
- Running today at Sanity Health, Milan — cofounder-owned, disclosed
Start with one visit. Build the record every other tool can cite.
Every packet a clinician reviews leaves behind attested, source-cited patient context in the clinic's own tenant. Read it back through the app, an MCP tool call, or a FHIR-shaped export — the same facts, the same citations, no re-parsing.
Working with hospitals or second-opinion services? Bring 5 real cases — we run them de-identified and you sign off the packets. Explore the platform
Coordinates record-preparation work. Does not diagnose, does not recommend treatment, does not replace clinicians. No certified FHIR / GDPR / EHDS / MDR / CE conformance is claimed.