The archive is the part nobody is digitising.
Turning photographed paper medical records into filed, structured patient data, without the patient’s identity ever leaving the country.
NAME: KALTABANG TARILONGA Age: 4 Sex: M Nem blong mama: Serah Kalo Village: Ipota Island: Tanna NHN: VU4471203 Weight 12.4 kg Temp 38.4 C BP 90/60 Diagnosis: Rheumatic Heart Disease Treatment: Benzathine Penicillin 1.2 MU IM stat Apgar Score 9. Glasgow Coma Scale 15. Seen by: Sr Leisale
NAME: [PERSON_003] Age: 4 Sex: M Nem blong mama: [PERSON_001] Village: [VU_VILLAGE_001] Island: [VU_ISLAND_001] NHN: [VU_NHN_001] Weight 12.4 kg Temp 38.4 C BP 90/60 Diagnosis: Rheumatic Heart Disease Treatment: Benzathine Penicillin 1.2 MU IM stat Apgar Score 9. Glasgow Coma Scale 15. Seen by: [PERSON_002]
A synthetic record, and actual output. The identity is replaced with reversible tokens whose cleartext stays in Vanuatu; the medicine crosses intact, byte for byte. Note what was not taken: Apgar Score and Glasgow Coma Scale are people’s surnames as well as clinical instruments, and a system that cannot tell the difference hands back a record with the medicine removed.
This is already Ministry policy, not a proposed opportunity.
Pillar 1 of the Vanuatu Digital Health Strategy 2025–2030 commits the Ministry of Health to acquire a Document Management System and to scan and electronically archive patient records, which future clinical systems are to read from.
Before migrating to a new NEHR, it is essential to clean, cull, and reorganize the vast volume of paper-based and legacy electronic records currently in use. This step, scheduled for 2025, is critical… Vanuatu Digital Health Strategy 2025–2030
Scheduled for 2025, and a hard prerequisite for the National Electronic Health Record. It is therefore on the critical path and already slipping.
Scanning is not structuring.
A Document Management System gives you a searchable picture of a page. It does not give you a child’s immunisation history as data. Those are different problems, and only the first one is funded.
Vanuatu has no national electronic health record and no facility running one — one of three Pacific countries in that position out of fourteen surveyed. Paper dominates, and the Ministry’s own strategy records that some records cannot be read for age or rain damage, that others cannot be located, and that information held at one facility is not visible at any other.
The regional systems all capture what is written today. Tamanu is the Pacific electronic medical record, DHIS2 has carried aggregate reporting since 2014, mSupply covers pharmaceuticals. None of them ingests the archive. That gap sits beside those systems and feeds them rather than competing with them.
Three commitments, and what each one costs.
There is no Health Information Policy or Act in Vanuatu — the Strategy says so outright. So there is no legal instrument permitting offshore processing of Ni-Vanuatu patient identifiers, and no regulator to seek comfort from. In that vacuum the only defensible posture is the strict one: a single egress point that re-checks every payload and refuses to send rather than trusting the step before it.
A confidently transcribed 1 that was a 7 becomes a permanent error in a real person’s chart, quite possibly after the paper it came from has been destroyed. A gap is recoverable; a confident error is not. Unreadable fields are returned as unreadable, with what could be seen and why it could not be read.
Transcribe and structure; never diagnose, suggest, or judge whether a value is concerning. The MHRA confirmed in July 2026 that products intended solely for transcription are not regulated medical devices, and that products supporting diagnosis or treatment are. Staying on the transcription side of that line is free. Crossing it costs a regulatory programme.
Removing the name is not de-identification in a country this small. Village, age and diagnosis together can name one person with the name already stripped, so island, province and place are tokenised alongside it, and each facility extends the list with its own catchment villages.
A name is proven by where it sits, not by a dictionary.
The obvious way to find names is a list of names. On Vanuatu paper records that fails in both directions at once. A lexicon built elsewhere does not contain Ni-Vanuatu given names. A form completed in block capitals — the ordinary way a form is completed — gets read as acronyms and skipped entirely. And in the other direction, a rule that treats two capitalised words as a person will take Ferrous Sulphate and Benzathine Penicillin with it.
So the decision is made from the slot instead of the word. A value is a name because of where it sits: after a name-bearing field label, in English, Bislama or French. The set of name labels on a health form is small and can be written down. The set of drug names cannot. That also means a kastom name in a name field is caught with no lexicon at all, because the field does not need to know how the name is spelled.
| What stands around the value | On the page | Decision |
|---|---|---|
| A name label before it | Nem blong mama: Serah Kalo | Held back |
| A title before it | Seen by: Sr Leisale | Held back |
| A clinical label governing the line | Impression: Acute Otitis Media | Never a person |
| A dose or route after it | Ferrous Sulphate 200 mg od | Never a person |
| An eponym tail after or inside it | Wilson disease · Apgar Score | Never a person |
The last row is the hard one. Medicine names its diseases after physicians, so Wilson, Hughes, Graves, Kaposi, Apgar and Calmette are surnames and clinical vocabulary at the same time. Nothing about the word separates them. What separates them is the word after it.
The models are already good at the hard parts.
| Not built | Why |
|---|---|
| An OCR engine | Frontier vision models read handwriting better than anything we would train, and degrade honestly when asked to. |
| A medical ontology | SNOMED CT and ICD are named in the Strategy and belong to the national record, not to us. |
| An electronic medical record | Tamanu exists, is open source, and was built for Pacific settings. This feeds it. |
| Aggregate reporting | DHIS2 has been the source of truth for that since 2014. |
| A fine-tuned model | There is no training corpus, and the health information workforce is largely short-term and donor-funded. Anything bespoke outlives the person who understood it by about a month. |
What happens next, and in this order.
- Establish accuracy on aged handwriting. Nothing else matters until that is known, and it is judged on the errors the model did not flag, not on a headline percentage.
- Image masking. The clerk outlines the identity block once per form type and it is applied to every photograph of that form before sending. This is the difference between a demonstration and something that can touch a real record.
- Offline-first capture. Connectivity is largely addressed: satellite broadband has been installed at 118 nominated hospitals, clinics and dispensariesEST and Starlink is available nationwide. Power is the binding constraint, not bandwidth, so a design assuming a live round trip per page is the wrong shape for this deployment.
- Then a conversation. The Strategy was written with WHO technical support, and WHO holds the pen on it. Any approach is a contribution to Pillar 1 through that channel, not a vendor pitch.