The 18 HIPAA identifiers, in plain English

Published 2026-07-29

If you work with health data in the United States and you want to use it without patient authorisation, you generally have to de-identify it. The HIPAA Privacy Rule gives you two routes, and the popular one is Safe Harbor: remove eighteen categories of identifier and the data stops being protected health information.

The list lives at 45 CFR 164.514(b)(2). Here it is in ordinary language, followed by the condition that most summaries leave out.

The eighteen

1. Names. Patients, relatives, employers, household members — not just the patient.
2. Geographic subdivisions smaller than a state. Street address, city, county, precinct. ZIP codes go too, except that the first three digits may be kept when the area they cover contains more than 20,000 people; where it does not, those three digits become 000.
3. All elements of dates except the year — birth date, admission, discharge, death, and all ages over 89 (which are aggregated into a single “90 or older” category).
4. Telephone numbers.
5. Fax numbers.
6. Email addresses.
7. Social Security numbers.
8. Medical record numbers.
9. Health plan beneficiary numbers.
10. Account numbers.
11. Certificate and licence numbers.
12. Vehicle identifiers and serial numbers, including licence plate numbers.
13. Device identifiers and serial numbers.
14. Web URLs.
15. IP addresses.
16. Biometric identifiers, including finger and voice prints.
17. Full-face photographs and comparable images.
18. Any other unique identifying number, characteristic or code.

Number 18 is the one that changes the character of the list. It is not a pattern; it is a catch-all, and it means the previous seventeen are a floor rather than a ceiling.

The condition people skip

Safe Harbor has a second requirement. After removing all eighteen identifier types, the covered entity must also have no actual knowledge that the remaining information could be used, alone or in combination with other information, to identify an individual.

That sentence does a lot of work. A note that reads “the only patient in the practice who has received this therapy” contains no identifier from the list and identifies one person to everyone in the building. So does a record describing an injury that was covered in the local newspaper. If you know the remaining data points at someone, the data is not de-identified, whatever your tooling says.

This is also why HIPAA offers the second route: expert determination, in which a qualified statistician applies accepted methods and documents that the re-identification risk is very small. Organisations use it precisely when the Safe Harbor list is either too blunt (it destroys the dates their research needs) or not enough (small populations, rare conditions).

Why dates are treated so harshly

Dates are the identifier people most resent losing, because they carry the clinical narrative: sequence, intervals, response to treatment. Safe Harbor keeps only the year.

The reason is linkage. An admission date plus a discharge date plus a hospital narrows a population to a handful of people, and hospital discharge data is often obtainable. Research on re-identification has shown for decades that a small number of quasi-identifiers — date of birth, sex, ZIP code — can single out a large fraction of a population. Dates are among the strongest.

If your work genuinely needs dates, that is an expert-determination conversation, not a tool-configuration one.

What a text scrubber can and cannot do here

The de-identify medical notes preset on this site switches on the detectors that map onto the list: names, addresses and postcodes, all dates, phone numbers, email addresses, SSNs, medical record numbers, account and licence numbers, IP addresses and URLs. It runs entirely in your browser, so the note never leaves the machine you are reading it on.

What it cannot do:

• Ages over 89. It does not detect an age as an age, so the “90 or older” aggregation is manual.
• Item 18. By definition, “any other unique characteristic” has no pattern. That is what the custom-terms box is for — a rare procedure name, a unique job title, a study code.
• Device, vehicle and biometric identifiers without a label near them.
• Free-text clinical narrative that identifies by circumstance. No detector sees that.
• The actual-knowledge condition. That is a judgement about what you know, which is not something software can hold an opinion about.

A practical sequence

1. Run the note through the Safe Harbor preset and read every highlight.
2. Fix what the machine missed: unusual names, labelless IDs, item-18 characteristics via custom terms.
3. Reread the scrubbed version asking one question — could someone who knows this patient recognise them here?
4. Handle ages over 89 by hand.
5. Record what you did. De-identification is a determination you may have to defend; a two-line note about the method and the review is cheap now and valuable later.

And the bit that has nothing to do with the list

De-identified data is outside HIPAA. That does not mean an AI vendor’s terms are irrelevant to your organisation, or that your institution permits the tool, or that the model’s output can go back into a chart without review. Those are separate decisions, and they usually have separate owners.

If the data is not de-identified and you still want to use AI on it, the route is a vendor relationship with a business associate agreement — which consumer chatbot plans do not offer. That is a procurement question, and no amount of careful pasting substitutes for it.