- name
- pseudonymizing-for-gdpr
- description
- Apply GDPR-grade pseudonymization to clinical or personal text with OpenMed, keeping a separately-held re-linkage key so the data can be controlled-re-linked later. Use when the user must process EU personal/health data under GDPR, asks for pseudonymization vs anonymization, needs Art. 4(5) / Art. 9 / Recital 26 alignment, wants a reversible mapping/key vault held apart from the data, or needs controlled re-linkage. Covers openmed.deidentify(policy="gdpr_pseudonymization", keep_mapping=True), storing the mapping in a separate key vault, reidentify() for authorized re-linkage, and retention. Pairs after extracting-pii-entities and configuring-privacy-policies.
- license
- Apache-2.0
- metadata
- {"project":"OpenMed","category":"de-identification","pairs":"after","version":"1.0"}
# Pseudonymizing for GDPR
Pseudonymization under the GDPR (Art. 4(5)) means processing personal data so it
"can no longer be attributed to a specific data subject without the use of
**additional information**" — provided that additional information (the
re-linkage key) is "kept separately and is subject to technical and
organisational measures." Crucially, **pseudonymized data is still personal
data** (Recital 26): re-linkage is possible, so GDPR still applies. This is the
opposite of anonymization, where re-identification is irreversibly prevented and
the data falls outside the GDPR.
OpenMed implements this with a single reversible de-identification pass plus a
mapping you store **away from the data**. This skill covers producing that
mapping, vaulting the key separately, and re-linking under authorization.
## When to use
- You process EU residents' personal or special-category health data (Art. 9)
and need a lawful, reversible safeguard rather than full anonymization.
- You need to keep a record-linkage capability (e.g. to recontact a patient,
reconcile longitudinal records, or honor a Subject Access Request) but must
separate the linkage key from the working dataset.
- A reviewer asks for the pseudonymization-vs-anonymization distinction in
writing, or for the ENISA-style "additional information kept separately"
control to be demonstrable.
Do **not** use this when the goal is irreversible anonymization for open release
— there, drop the mapping entirely and gate residual risk with
`reviewing-reidentification-risk`. Pseudonymization keeps a key; anonymization
must not.
## Quick start
```python
import openmed
# Synthetic record — never run this skill's examples on real PHI.
note = "Patient Maria Schmidt (ID 4471) seen 2024-03-02; contact maria@example.de."
result = openmed.deidentify(
note,
method="replace", # realistic surrogates, not [LABEL] holes
policy="gdpr_pseudonymization", # bundled GDPR profile
keep_mapping=True, # produce the reversible re-linkage map
consistent=True, # same input -> same surrogate in the doc
seed=20240302, # cross-run reproducibility of surrogates
)
pseudonymized_text = result.deidentified_text # safe to process / analyze
relink_key = result.mapping # surrogate -> original; SECRET
```
`result.deidentified_text` is the pseudonymized payload. `result.mapping` is the
"additional information" GDPR Art. 4(5) requires be kept separately — it is the
key that makes re-linkage possible, and therefore the most sensitive artifact in
the whole flow.
## Workflow
1. **Choose reversible pseudonymization, not masking.** Use `method="replace"`
with `policy="gdpr_pseudonymization"` and `keep_mapping=True`. Replacement
surrogates keep the text usable for downstream NLP while remaining
non-identifying. `consistent=True` (optionally with `seed=`) makes repeated
mentions resolve to one stable surrogate so intra-document linkage survives.
2. **Split the data from the key immediately.** The moment `deidentify` returns,
route `result.deidentified_text` to your working store and `result.mapping`
to a **separate, access-controlled key vault** — different system, different
credentials, different backups. Never persist them in the same row, file,
bucket, or log line. This separation is the technical-and-organisational
measure that makes the data pseudonymized rather than just "personal data
with PII in it."
3. **Process the pseudonymized text freely.** Run `analyze_text`, analytics,
model training, or transfer on `deidentified_text`. The key never leaves the
vault during ordinary processing.
4. **Re-link only under authorization.** When a lawful basis exists (e.g. an
authorized SAR or recontact), fetch the mapping from the vault and call
`openmed.reidentify(deidentified_text, mapping)`. Log *that* a re-linkage
happened (who, when, why, record id) — but never log the restored plaintext.
5. **Apply retention to the key.** The mapping has its own retention clock. When
the lawful basis for re-linkage ends, **destroy the mapping**. Once the key
is irreversibly gone and no other re-identification path remains, the
remaining text approaches anonymization and GDPR obligations shrink
accordingly. Verify that claim with `reviewing-reidentification-risk` before
relying on it.
## Hand-off to / from OpenMed
- **From** `extracting-pii-entities` / `configuring-privacy-policies`: confirm
the detector recall and the active policy profile before pseudonymizing, since
any identifier the detector misses leaks into `deidentified_text`.
- **OpenMed call:** Python `from openmed import deidentify, reidentify`; the same
capability is exposed as MCP tool `openmed_deidentify` and REST `/deidentify`.
Pass `policy="gdpr_pseudonymization"`, `keep_mapping=True`.
- **To** `auditing-deid-leakage`: scan `result.deidentified_text` for residual
identifiers before it leaves the boundary — pseudonymization is only as strong
as detection.
- **To** `reviewing-reidentification-risk`: quasi-identifier (age, ZIP, dates)
re-identification still applies to pseudonymized data; score k-anonymity on the
output and document residual risk.
## Edge cases & gotchas
- **Pseudonymized ≠ anonymized.** As long as `mapping` exists anywhere, the data
is personal data under Recital 26. Do not market a `keep_mapping=True` output
as "anonymous."
- **The mapping is the crown jewel.** A leaked mapping re-identifies everything
at once. Treat it as the highest-sensitivity secret: encrypt at rest, restrict
access, audit reads.
- **Surrogates can still carry quasi-identifiers.** `method="replace"` swaps the
identifier text, but free-text age, rare diagnosis, ZIP, or admission dates
remain. Pseudonymization does not address singling-out; pair with QI risk
scoring.
- **Reproducibility cuts both ways.** A fixed `seed` makes surrogates stable
across runs (good for linkage) but means an attacker who learns the seed and
algorithm can reproduce surrogates — keep the seed with the key, not the data.
- **Special-category data (Art. 9).** Health data needs a lawful basis *before*
processing; pseudonymization is a safeguard, not a lawful basis on its own.
- **Local-first.** Run entirely on-device. Do not send EU personal data to a
cloud de-identification service to satisfy GDPR — that may itself be a transfer.
## Standards & references
- GDPR Art. 4(5) — definition of pseudonymization:
https://gdpr-info.eu/art-4-gdpr/
- GDPR Art. 9 — processing of special categories (health) data:
https://gdpr-info.eu/art-9-gdpr/
- GDPR Recital 26 — pseudonymous data is personal data; anonymization test:
https://gdpr-info.eu/recitals/no-26/
- ENISA, *Pseudonymisation techniques and best practices* (2019):
https://www.enisa.europa.eu/publications/pseudonymisation-techniques-and-best-practices
- EDPB Guidelines on pseudonymisation (01/2025):
https://www.edpb.europa.eu/
GitHubで見る