What Is HIPAA Safe Harbor De-Identification?

A hospital analyst wants to send 5,000 discharge notes to a vendor for a language model pilot. Names are gone, and so are phone numbers. The notes still say "admitted 14 March", a five-digit ZIP code and a note about a rare condition in a town of 900 people. Is that de-identified under HIPAA? Under Safe Harbor, no: dates and ZIP codes are on the list, and the rare condition raises the knowledge question. The details of the rule explain why.

What are the two HIPAA de-identification methods?

45 CFR 164.514(b) lets a covered entity treat health information as not individually identifiable in two ways. Under Expert Determination (paragraph (b)(1)), a person with appropriate knowledge of and experience with generally accepted statistical and scientific principles determines that the risk is very small that the information could be used, alone or with other reasonably available information, by an anticipated recipient to identify an individual, and documents the methods and results. Under Safe Harbor (paragraph (b)(2)), you remove specified identifiers and have no actual knowledge that the rest could identify someone. HHS explains both in its de-identification guidance.

What are the 18 Safe Harbor identifiers?

Paragraph (b)(2)(i) lists identifiers of the individual or of relatives, employers or household members of the individual. All 18 must be removed:

1. Names.
2. Geographic subdivisions smaller than a state, including street address, city, county and ZIP code. The first three digits of a ZIP code may stay only if the area formed by all ZIP codes with those digits has more than 20,000 people; otherwise they become 000.
3. All elements of dates except the year for dates directly related to the individual (birth, admission, discharge, death), and all ages over 89. Ages over 89 may be aggregated into one category of 90 or older.
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 license numbers.
12. Vehicle identifiers and serial numbers, including license plates.
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, except as permitted by paragraph (c).

What does "no actual knowledge" mean?

Paragraph (b)(2)(ii) adds a second condition: the covered entity must not have actual knowledge that the information could be used, alone or in combination with other information, to identify an individual. Removing the 18 items isn't enough if you know that the remainder points to one person. If a research team knows that the only patient with a particular transplant in a small town is in the file, the file fails the condition even if every listed identifier is gone.

The last item, "any other unique identifying number, characteristic or code", is the open-ended one. A patient number from your own system, a study ID that maps back to a record, a rare job title in free text and a distinctive handwritten note can all fit it. Paragraph (c) sets conditions for re-identification codes, so read it before assuming a token scheme qualifies. Those are judgment calls, and they are why many organizations choose Expert Determination for complex data.

How does masking relate to Safe Harbor?

Masking can remove many of the 18 identifiers from text before a prompt leaves your environment. It doesn't make the text de-identified by itself, for three reasons. Detectors have limits: a record number with no label next to it can pass, and a date outside clinical context can pass. Free-text facts need judgment, which a detector doesn't supply. And if you hold a mapping that restores the originals, the data you hold is still identifiable to you.

Maskflare's HIPAA Safe Harbor profile switches on the healthcare, US, global, person-name and credentials packs in one change, and spoken and obfuscated forms of values are included. The coverage report then reads your live policy against the 18 identifiers (A to R) and shows each as covered, partial, not covered or out of scope, with exports for your records. Biometric identifiers and photographs are not text, so they are out of scope for text masking. Where the report says partial, it also names the gap, such as a license number that needs a keyword beside it. The Detection Engine holds originals behind tokens encrypted for 30 days so answers can be restored.

What stays your job?

Maskflare doesn't make data HIPAA compliant, and there's no such thing as HIPAA certification. Whether a dataset meets Safe Harbor is your decision, and it depends on testing against your real documents and on what you know about them. If a vendor receives protected health information on your behalf, you still need a business associate agreement. Our HIPAA page covers that, and PII vs PHI vs PCI explains the vocabulary.