Paper health records · Vanuatu · Updated 13 September 2026

The archive is the part nobody is digitising.

Photographs of paper medical records, turned into structured patient data. The patient’s identity never leaves the country.

  1. 01Photograph
  2. 02Cover identity
  3. 03Hold as codes
  4. 04Structure
Child health card · synthetic
NAME: KALTABANG TARILONGA[PERSON_003]Age: 4Sex: M
Nem blong mama: Serah Kalo[PERSON_001]Village: Ipota[VU_VILLAGE_001]
Island: Tanna[VU_ISLAND_001]NHN: VU4471203[VU_NHN_001]
Weight 12.4 kgTemp 38.4 CBP 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[PERSON_002]
Structured record
Form
Child health card
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

Patient [PERSON_003]. The name stays on the device.

Try it yourself → Clinic check-in demo → Built-in sample card. Nothing is sent anywhere.
Synthetic card. The codes are the tokeniser’s real output; the record shows the output format. Apgar and Glasgow are surnames as well as clinical scales, and both come through intact.
The gap

Scanning is not structuring.

A scanned page is a searchable picture, not a child’s immunisation history as data. Vanuatu has no national electronic health record, paper dominates, and some records can no longer be read for age or rain damage.

Tamanu, DHIS2 and mSupply capture what is recorded today. None of them reads the archive. This feeds them rather than competing with them.

The rules

Four rules everything else serves.

Identity stays

One exit point re-checks every image and every line of text before anything is sent, and refuses rather than trusting the step before it.

Illegible beats plausible

A confident wrong digit becomes a permanent error in someone’s chart. Unreadable fields come back marked unreadable.

No interpretation

Transcribe and structure only: no diagnosis, no suggestion, no “this value is high”. Under MHRA guidance, transcription-only tools are not medical devices; diagnostic tools are.

Place is identity

In a small country, village, age and diagnosis can identify someone without a name, so island, village and ID numbers are held back too.

Reading a name off a form

A name is found by where it sits, not by a dictionary.

A list of names misses Ni-Vanuatu names and mistakes Benzathine Penicillin for a person. So a value is a name only when it sits in a name field, in English, Bislama or French.

Around the valueOn the pageDecision
A name label before itNem blong mama: Serah KaloHeld back
A title before itSeen by: Sr LeisaleHeld back
A clinical label on the lineImpression: Acute Otitis MediaNever a person
A dose or route after itFerrous Sulphate 200 mg odNever a person
An eponym after or inside itWilson disease · Apgar ScoreNever a person
Deliberately not built

The models already do the hard parts.

Not builtWhy
An OCR engineCurrent vision models already read handwriting; we use the best available.
A medical ontologySNOMED CT and ICD belong to the national record.
An electronic medical recordTamanu exists and was built for the Pacific. This feeds it.
Aggregate reportingDHIS2 already does this.
A fine-tuned modelNo training corpus, and nothing bespoke to maintain.
Next
  1. Measure accuracy on aged handwriting, counting the errors the model did not flag.
  2. Offline-first capture. At health facilities, power is the constraint, not bandwidth.
  3. Output straight into Tamanu from the structured record.