Glossary

De-identification vs anonymization

Two words the healthcare data world uses interchangeably and shouldn't — what separates de-identified data from truly anonymous data under US and EU law, and why the gap between them is where re-identification risk lives. As of September 2026.

By Jonas WeirReviewed by Jonas Weir · editorial reviewUpdated

The short version

  • De-identification removes or masks identifiers under a named standard; anonymization goes further, rendering data so a person can no longer be identified by any reasonably likely means.
  • HIPAA offers two de-identification routes — Safe Harbor (remove 18 identifier categories) and Expert Determination (a qualified person certifies the re-identification risk is very small).
  • Under GDPR, anonymous data falls outside data-protection law, but pseudonymised data remain personal data because a key can re-attribute them.
  • The gap matters: one study estimated 99.98% of Americans could be re-identified from any dataset using 15 demographic attributes.

De-identification and anonymization both strip identifying details from records, but set different bars. De-identification removes or masks identifiers under a named standard; anonymization goes further — rendering data so a person can no longer be identified by any reasonably likely means, which takes it outside privacy law. As of September 2026.

The two words get used as synonyms in healthcare, and the confusion has legal consequences. De-identification is a process with a defined outcome and a residual, bounded risk. Anonymization is a threshold: once truly met, the data stops being personal data at all. Almost everything called "anonymized" in practice is really de-identified, and the gap between the two is exactly where re-identification happens.

Why it matters in healthcare

Clinical datasets are dense with the very quantities that make people unique — dates, rare diagnoses, ZIP codes, admission timelines. The label you attach to a dataset decides what you may lawfully do with it. Treat de-identified data as anonymous and you may share, sell access to, or train on records that still carry obligations of consent, security, and subject rights. Getting the vocabulary right is the first control, before any technical safeguard.

How it works in practice

Under US law, the HIPAA Privacy Rule gives two routes. Safe Harbor requires removing 18 categories of identifiers of the individual and of their relatives, employers, or household members. Expert Determination asks a person with appropriate statistical knowledge to certify that the risk of re-identification is very small and to document how they reached that conclusion 1. Both produce de-identified data with low residual risk — a bounded risk, rather than a promise of anonymity.

Under EU law, the framing differs. GDPR says its principles do not apply to anonymous information — data rendered so the subject is no longer identifiable 2. But it defines pseudonymisation as processing so data can no longer be attributed to a person without separately-held additional information 2. The distinction is decisive: the European Data Protection Board's 2025 guidance states plainly that pseudonymised data, which could be attributed to a person using additional information, remain information on an identifiable person and are therefore personal 3. The guidance also frames a bounded pseudonymisation domain — the environment within which attribution is prevented — acknowledging that the protection holds only inside that boundary.

Where it appears today

Every health system that shares data for research, licenses records to a vendor, or trains a model on patient histories runs one of these processes first. NIST's foundational treatment defines de-identification as removing the association between data and identifiers, while warning that de-identified data can sometimes be re-identified 4. That warning has teeth. A generative-model study estimated that 99.98% of Americans could be correctly re-identified in any dataset using just 15 demographic attributes 5 — a result that reset how privacy regulators think about "incomplete" or sampled datasets.

Common misunderstandings

The costliest error is collapsing the two terms and assuming de-identified data carries no obligations. It usually does. A second error is treating anonymization as a fixed property of a file rather than a property of a file and its environment: the same table can be effectively anonymous inside a locked research enclave and re-identifiable the moment it is joined to an outside dataset. This is why the EDPB ties protection to a defined domain, and why HIPAA's Expert Determination weighs the data "reasonably available" to a would-be attacker. Anonymization is a claim about the world around the data, not the columns alone.

Related terms

When re-identification risk is too high to release real records at all, teams turn to synthetic patient data, which generates artificial records in place of masked ones. When the goal is to train across institutions without moving records anywhere, federated learning keeps the data local instead of de-identifying it for export. For the wider picture of where clinical AI has spread and what it runs on, see our AI in healthcare statistics.

We review this page yearly, and sooner when the EDPB finalises its pseudonymisation guidance, HHS updates its de-identification guidance, or a new study revises the re-identification risk estimates above.

Questions & answers

  • Is de-identified data the same as anonymous data?

    They describe different bars. De-identified data has had identifiers removed or masked under a standard, but often retains enough structure that re-identification stays possible. Anonymous data meets a higher threshold — the person can no longer be identified by any reasonably likely means — and only then does it fall outside privacy law.

  • Does HIPAA-compliant de-identification make data anonymous?

    Not automatically. HIPAA's Safe Harbor and Expert Determination methods produce de-identified data with low re-identification risk, but the standard is a defined, bounded risk rather than a guarantee of anonymity. The same dataset can still be personal under other regimes such as GDPR.

  • Why does the distinction matter for AI training?

    Because the legal obligations attach to the label. Anonymous data can be used freely; personal data — including pseudonymised and most de-identified data — carries consent, security, and rights obligations. Mislabelling de-identified data as anonymous is a common and costly compliance error.

Sources

  1. U.S. Code of Federal Regulations. 45 CFR 164.514 — Other requirements relating to uses and disclosures of protected health information. HIPAA Privacy Rule. www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.514
  2. Regulation (EU) 2016/679 (General Data Protection Regulation), Recital 26 and Article 4(5). Official Journal of the European Union. eur-lex.europa.eu/eli/reg/2016/679/oj/eng
  3. European Data Protection Board. Guidelines 01/2025 on Pseudonymisation. Adopted 16 January 2025 (version for public consultation). www.edpb.europa.eu/system/files/2025-01/edpb_guidelines_202501_pseudonymisation_en.pdf
  4. Garfinkel SL. De-Identification of Personal Information. NIST Internal Report (IR) 8053. National Institute of Standards and Technology. 2015. doi.org/10.6028/NIST.IR.8053
  5. Rocher L, Hendrickx JM, de Montjoye YA. Estimating the success of re-identifications in incomplete datasets using generative models. Nature Communications. 2019;10:3069. doi.org/10.1038/s41467-019-10933-3