synthdata-no

Deterministic generator of synthetic Norwegian health data.

Combines Tenor-range synthetic fødselsnummer, gold-annotated bokmål clinical text, no-basis-aware FHIR R4B bundles, and tabular microdata with Norwegian kommune marginals — all in one pip-installable package.

Produced as a shared synthetic substrate for the open Norwegian health-AI portfolio: medspacy-no · fhir-safety-harness · omsorgsradar

No real persons. No re-identification risk by construction. NOT for statistical inference about the Norwegian population.

---

Installation

pip install synthdata-no          # PyPI (owner gate — not yet published)
# or from source:
git clone https://github.com/Alksalt/synthdata-no
cd synthdata-no
uv sync
uv run synthdata-no --help

---

Four data families

1. Persons — Tenor-range synthetic identifiers

Generates Person records with fødselsnummer or D-nummer in the Tenor/Test-Norge synthetic range (month+80, mod-11 checksum computed after the offset). Backed by Faker no_NO for names/addresses and the SSB KLASS 131 kommune registry.

CLI:

uv run synthdata-no persons --seed 42 --n 100 --out persons.csv

Python:

from synthdata_no.persons import generate_persons

persons = generate_persons(n=100, seed=42)
for p in persons[:3]:
    print(p.name, p.fnr, p.kommune)

Synthetic fnr convention: month digits are offset by 80 (month ∈ 81–92); D-nr additionally offset by 40 in the day digits. Example: a person born 17 November 2020 gets fnr 17912099997 and D-nr 57912075186 (verified Tenor pair). No generated identifier has month ≤ 12 — a guard test enforces this.

2. Tabular — Norwegian kommune microdata

Generates a synthetic microdata table driven by SSB 07459 population marginals (age × sex × kommune). Correlated columns via configurable conditional probability tables (CPTs). Optional planted Tenor-range fnr for anonymization pipeline testing.

CLI:

uv run synthdata-no table --seed 42 --n 600 --out table.csv

Python:

from synthdata_no.tabular import generate_table

df = generate_table({}, n=600, seed=42)
print(df.head())

Default columns: kommune (4-digit KLASS code), age, sex, tjeneste_bruk, diagnosekategori.

3. Clinical text — gold-annotated bokmål snippets

Generates synthetic bokmål clinical text records with character-offset entity spans compatible with medspacy-no. Covers all six ConText categories: NEGATED_EXISTENCE, POSSIBLE_EXISTENCE, HYPOTHETICAL, HISTORICAL, FAMILY, AFFIRMED. Includes pseudo-negation traps (kan ikke utelukkes → POSSIBLE, not NEGATED), conjunction-scope traps, abbreviation-bearing sentences, and 15 section-header families.

CLI:

uv run synthdata-no text --seed 42 --n 500 --out clinical.jsonl

Python:

from synthdata_no.export.medspacy import to_jsonl, to_spacy_examples
import spacy

path = to_jsonl("clinical.jsonl", n=500, seed=42)
nlp = spacy.blank("nb")
examples = to_spacy_examples(path, nlp)
print(examples[0].reference.ents)

JSONL shape: {"text", "spans": [{start, end, label, token_start, token_end}], "meta": {gold_context, trigger, section, template_id}}.

4. FHIR — no-basis-aware R4B bundles

Generates transaction Bundles (fhir.resources R4B) with Patient (no-basis-Patient profile), Medication (FEST+ATC dual coding), MedicationStatement, Condition (ICD-10-NO, code-only), and Observation (LOINC). All UUIDs are derived from the seeded RNG. Offline validation: pydantic model_validate round-trip + fhir-validator JSON-schema check.

CLI:

uv run synthdata-no fhir --seed 42 --n 5 --out fhir_out/

Python:

from synthdata_no.export.safety_harness import write_fixture_set
from pathlib import Path

write_fixture_set(Path("fhir_out"), n_patients=5, seed=42)
# produces fhir_out/patient_0000.json … patient_0004.json + index.json

---

Consumer quick-starts

medspacy-no — ConText gold evaluation

to_spacy_examples requires spaCy, which is an optional dependency:

pip install "synthdata-no[nlp]"   # installs spacy>=3.8 alongside synthdata-no
# or, from source:
uv add spacy

from synthdata_no.export.medspacy import to_jsonl, to_spacy_examples
import spacy

jsonl_path = to_jsonl("gold_eval.jsonl", n=500, seed=42)
nlp = spacy.blank("nb")
examples = to_spacy_examples(jsonl_path, nlp)
# examples is a list of spacy.training.Example with gold entity spans

Note: to_jsonl itself has no spaCy dependency and works with a bare pip install synthdata-no.

fhir-safety-harness — baseline fixture set

from synthdata_no.export.safety_harness import write_fixture_set
from pathlib import Path

write_fixture_set(Path("fixtures/"), n_patients=5, seed=42)
# Each bundle is the no-trap baseline. Trap content is owner-authored in fhir-safety-harness.

omsorgsradar — brfss-demo anonymize fixture

from synthdata_no.export.omsorgsradar import write_brfss_shaped_fixture
import pandas as pd

df = write_brfss_shaped_fixture("analyses/brfss-demo/microdata/brfss_sample.csv", seed=42)
# 600 rows, columns: state (kommune code), age, sex, diabetes, notes
# notes[0] and notes[1] carry planted Tenor-range fnr for the anonymize pipeline

# Always pass dtype={'state': str} when reading back — state is a zero-padded 4-digit code:
df = pd.read_csv("brfss_sample.csv", dtype={"state": str})

write_brfss_shaped_fixture also accepts top_n_kommuner=12 (default: 12 largest by population) to concentrate the 600 rows into fewer kommuner, ensuring enough rows per cell for k=10 analysis classes to survive. Increase to spread more thinly across the full KLASS snapshot.

All three at once

uv run synthdata-no fixtures --seed 42 --n 10 --out fixtures/
# fixtures/omsorgsradar/brfss_sample.csv
# fixtures/medspacy_no/clinical.jsonl
# fixtures/fhir_safety_harness/patient_0000.json … + index.json

---

Honesty framing

Synthetic data statement: All records generated by synthdata-no contain no real persons. Identifiers are constructed in the Tenor/Test-Norge synthetic range (month+80) and carry no re-identification risk by construction — they cannot be matched to any real register entry.

This package is NOT suitable for statistical inference about the Norwegian population or Norwegian health services. Default marginals (SSB 07459) and CPT values are approximate aggregate-informed defaults, not calibrated epidemiological models.

Clinical text templates were authored and reviewed by a physician (utdannet lege, master i medisin) for structural realism. No realism claim is made until the owner has completed the sign-off checklist in PHYSICIAN_REVIEW.md. The templates do not represent typical Norwegian EPJ/journal text — they are synthetic training examples for NLP evaluation.

GDPR anonymity is not self-certified. For advice on whether this data is personal data in your context, consult Datatilsynet.

---

Determinism scope

synthdata-no is deterministic at two levels:

• Structural fields (fnr, ICD-10 codes, ATC codes, CPT draws, character offsets, FHIR UUIDs) are always seed-deterministic — same seed + n → byte-identical output, regardless of Python version or dependency versions.
• Faker-derived text (names, addresses) is additionally Faker-version-dependent. faker>=25.0,<26 is pinned for this reason. Byte-identical output for name/address fields is only guaranteed within the same pinned Faker minor version. If you need cross-environment reproducibility for downstream tests that depend on name strings, pin Faker explicitly in your own requirements.txt.

In practice: fnr math, span offsets, and codes are always stable. Names may shift across Faker minor upgrades.

---

Synthetic fnr convention

The Tenor Test-Norge marker is month+80 applied before computing the mod-11 control digits:

Real date	Real fnr (illustrative)	Synthetic fnr	Synthetic D-nr
17.11.2020	17112099nnn	17912099997	57912075186

D-numbers additionally offset the day by +40. Both offsets are applied to the stem digits; the mod-11 control digits (k1, k2) are computed after the offsets. Values where k1=10 or k2=10 are discarded and regenerated (~17% of stems).

The is_synthetic_fnr() function returns True for any identifier in this range. The guard test in tests/test_persons.py asserts that no generated person ever has month ≤ 12.

---

Prior art

The following prior art is acknowledged and cited:

• Tenor / Test-Norge (Skatteetaten/NAV): the official Norwegian synthetic ID register; synthdata-no uses the same marker convention (month+80) and is compatible with real Tenor-issued identifiers.
• NorSynthClinical (ltgoslo / University of Oslo): 477 manually-authored bokmål family-history sentences with entity/relation gold; does not cover ConText-style negation/context annotation — the gap this package addresses (to our knowledge).
• Lund et al. 2024 (UNN-SPKI / Nor-DeID-SynthData): 1,200 GPT-4-generated Norwegian discharge summaries with PHI spans — a static de-identification benchmark, not a generator.
• Synthea (MITRE): active FHIR patient generator; synthea-international/no/ is a renamed Finnish skeleton with no no-basis profile, no bokmål text, and US names. A full Norwegian port would require months of Java work.
• fhir-kindling (PyPI): closest pip-installable FHIR generator; profile-agnostic, no Norwegian layer.
• NHN SyntPop (Testuniverset): synthetic persons portal, HelseID-gated, demographics only, no FHIR/clinical content.

To our knowledge, no existing open pip-installable Python package combines Tenor-range identifiers, gold-annotated bokmål ConText text, and no-basis-aware FHIR R4B bundles in a single deterministic library.

---

License and attribution

MIT License — see LICENSE.

Embedded data sources (see THIRD_PARTY_SOURCES.md for full attribution notices):

• FEST (DMP): NLOD 2.0. "Contains data from FEST (Forskrivnings- og ekspedisjonsstøtte), published by Direktoratet for medisinske produkter (DMP), made available under NLOD 2.0."
• LOINC: Regenstrief Institute. "This content includes LOINC codes … copyright © 1995 Regenstrief Institute, Inc. and the LOINC Committee … see https://loinc.org/license/."
• ICD-10 (WHO): CC BY-ND 3.0 IGO — codes only, no Norwegian display names embedded (adaptation would require WHO permission).
• SSB KLASS 131 (Statistics Norway): CC BY 4.0.

---

Author

Oleksandr Altukhov — utdannet lege (master i medisin), agentic-AI engineer.

---

Bokmål sammendrag

synthdata-no er et Python-bibliotek for deterministisk generering av syntetiske norske helsedata. Det produserer fire typer data:

1. Fødselsnummer og D-nummer i Tenor/Test-Norge-området (måned+80), med norske navn og adresser (Faker no_NO) og gyldige kommunekoder fra SSB KLASS 131.
2. Tabelldata med realistiske aldersfordelinger per kommune (SSB 07459), drevet av konfigurerbare betingede sannsynlighetstabeller.
3. Klinisk tekst på bokmål med gullmerkede entitetsspenn for ConText-evaluering — dekker negasjon, usikkerhet, historikk, familieanamnese og hypotetisk kontekst.
4. FHIR R4B-bunter med no-basis-Patient-profil, FEST+ATC-legemiddelkoding, ICD-10-NO-diagnoser (kun koder) og LOINC-observasjoner.

Biblioteket brukes som felles syntetisk grunnlag for medspacy-no, fhir-safety-harness og omsorgsradar.

Ingen reelle personer. Ingen gjenkjenningsrisiko. Ikke egnet for statistisk inferens om den norske befolkningen.