What is PHI (protected health information)?
Most teams draw this line at the patient's name. The rule draws it at a ZIP code, an admission date and an IP address. That gap is how a table you believed was anonymous stays in scope.
PHI (protected health information)
Protected health informationePHI
Protected health information is health data that can be traced to a person, held or sent by a HIPAA covered entity or its business associate. Health data stops being PHI only when identifiers come out by one of the two methods HIPAA names.
Two things have to be true at once. Data has to say something about a person's health, care or payment for care. It also has to be linkable back to that person. Drop either half and it is not PHI.
ePHI is the same thing in electronic form. The Security Rule uses that word because its safeguards apply to systems, not paper.
Who holds the data decides whether HIPAA applies at all. A hospital is a covered entity. A vendor processing that data for the hospital is a business associate. A consumer wellness app with no such relationship may sit outside HIPAA entirely, and still fall under a state health privacy law.
Removing names does not make health data anonymous
Latanya Sweeney showed that date of birth, sex and five-digit ZIP code single out 87% of the US population. No name required. HIPAA's own rule treats a birth date and a ZIP code as identifiers for that reason.
The rule at 45 CFR 164.514 gives you two doors. Expert determination has a statistician document that the risk of re-identifying anyone is very small. Safe Harbor strips 18 kinds of identifier, and most teams reach for it.
Read the 18 and the cost lands. Geography smaller than a state goes, so ZIP cuts to three digits and becomes 000 where fewer than 20,000 people live. Dates go too, apart from the year, which takes admission and discharge. Every age over 89 collapses into one bucket. Those are the exact columns a clinical model wants.
- 01Names, phone and fax numbers, email addresses, social security numbers.
- 02Medical record numbers, health plan numbers, account and licence numbers.
- 03Device identifiers, vehicle identifiers, web URLs, IP addresses and biometrics.
- 04Full-face photographs, and a catch-all for any other unique identifying number, characteristic or code.
- Health dataDiagnosis, treatment, payment.
- IdentifierName, date, ZIP, IP, MRN.
- PHIBoth halves present.
- Strip or certifySafe Harbor, or an expert.
- Out of scopeUnless you know it re-identifies.
The last node is the one people skip. Safe Harbor carries an actual-knowledge condition: it does not apply if you already know the data can be re-identified.
Related questions
01Is an IP address PHI?
An IP address is one of the 18 identifiers at 45 CFR 164.514, so health data carrying one is not de-identified under Safe Harbor. The other half of the test still applies. The data must also be health information held by a covered entity or a business associate. An IP address in a web log with no health context is not PHI.
02Does HIPAA apply to my health app?
Only if you are a covered entity or a business associate of one. A direct-to-consumer wellness app with no provider or plan relationship often sits outside HIPAA altogether. That is less of a relief than it sounds. State consumer health privacy laws and the FTC reach into that gap, and some carry a private right of action.
03Can we send PHI to an LLM provider?
Only under a business associate agreement with that provider, and only if their terms allow it. A model API that touches PHI is a subcontractor in HIPAA's terms. It needs its own agreement. Your agreement with the hospital does not supply one. Teams often find this out after the retrieval pipeline is built.
Keep reading
- What a BAA does not cover →Signing the agreement is not the same as being compliant.
- How do you build HIPAA-compliant AI? →Where PHI enters an AI pipeline, and what has to be true at each point.
- Healthcare software development →What we have shipped in clinical workflow, and what we have not.
- TrialTriage: AI clinical trial matching →De-identified data model, field-level encryption, 23 audited action types.
