Wiki · Translation · concept
De-identification
Stripping the identifiers that tie a health record to a person — necessary to share clinical data, but not enough to make it anonymous.
De-identification removes the pieces of a health record that point to a specific person so the rest can be shared, pooled, or trained on. It is the front door of clinical data privacy — the first thing you do before data leaves a hospital — and it is legally load-bearing: under HIPAA, “de-identified” data falls outside the privacy rule entirely, so how you strip identifiers decides what you are even allowed to do.
How it works
HIPAA defines two routes. Safe Harbor is a checklist: remove 18 specified identifiers — names, geographic detail finer than state, every date more precise than a year, record and account numbers, and so on — and the data is deemed de-identified. Expert Determination is the statistical route: a qualified expert certifies that the re-identification risk is “very small” (45 CFR §164.514). For free-text clinical notes, finding the identifiers is itself a machine-learning problem — NLP systems like Philter learn to scrub protected health information out of the prose before a downstream model ever sees it (Norgeot et al., 2020).
Why it matters (for bio × AI)
De-identification is what makes clinical data usable at all — it is the precondition for sharing notes, building cohorts, and pretraining on records. But the honest limitation is structural: removing the obvious identifiers does not make a record anonymous. The leftover quasi-identifiers still fingerprint people. Latanya Sweeney showed that 87% of Americans are uniquely identified by just ZIP code, birth date, and sex — none of which Safe Harbor requires you to remove (Sweeney, 2002) — and later work estimates 99.98% could be re-identified from 15 demographic attributes (Rocher et al., 2019). So de-identification is necessary but not sufficient: it is one layer of a privacy stack, not the whole thing.
Related
See differential privacy for a mathematical bound on leakage that de-identification can’t provide, and federated learning for training on data that stays inside each institution — both layers that cover what de-identification alone leaves exposed.