HIPAA de-identification before you use AI

HIPAA de-identification: the two methods

Under the HIPAA Privacy Rule, health information is de-identified when it meets one of two standards in 45 CFR §164.514(b): Safe Harbor or Expert Determination. De-identified information is no longer protected health information (PHI), so the Privacy Rule no longer restricts how it is used or disclosed.

That matters for AI. If you send PHI to an AI provider, the provider is handling PHI on your behalf and the usual HIPAA obligations follow. If you send properly de-identified text, the provider does not receive PHI at all.

The 18 Safe Harbor identifiers

Safe Harbor requires removing these identifiers of the individual and of the individual's relatives, employers and household members.

1. Names.
2. Geographic subdivisions smaller than a state, including street address, city, county, precinct and ZIP code (the first three digits of a ZIP code may be kept in some cases, if the area they cover has more than 20,000 people).
3. All elements of dates, except the year, directly related to the individual, such as birth, admission, discharge and death dates; and all ages over 89 (which may be grouped as 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/license numbers.
12. Vehicle identifiers and serial numbers, including license 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, except a permitted re-identification code.

Free-text clinical notes are where identifiers hide: a relative's name in a history, a date in a discharge summary, an employer mentioned in an occupational note. Detection needs to cover the whole text, not just the header fields.

Re-identification codes: why tokens differ from hashes

HIPAA allows you to keep a code that lets you re-identify de-identified data, under §164.514(c), provided two conditions are met. The code must not be derived from or related to information about the individual, and you must not use or disclose it for other purposes or disclose the mechanism for re-identification.

This is where the choice of replacement mode matters.

• Sequential tokens such as [PERSON_1] are assigned in order of appearance. They are not derived from the patient's name or record, so, kept with a protected key, they can fit the description of a permitted re-identification code.
• A hash of an identifier, such as an HMAC of a medical record number, is computed from that identifier. It is derived from information about the individual, so it would not qualify as a §164.514(c) code, even though it cannot be reversed mathematically.

The differences between tokens, hashes and other modes are explained in seven ways to replace personal data and summarised in the replacement modes section.

When you need a Business Associate Agreement

If a vendor creates, receives, maintains or transmits PHI on your behalf, it is generally a business associate and you need a Business Associate Agreement with it.

That applies at two points in an AI workflow. First, the de-identification tool: if it processes PHI on its servers, it handles PHI. Second, the AI provider: if what you send is not de-identified, the provider handles PHI too.

LEXARIS Cloud processes documents on our servers in St Petersburg, Russia (EU, UK, US and Swiss regions are planned and not yet available), and cloud processing of PHI is not currently offered. With Keys stay with you, detection and replacement run in your browser and LEXARIS does not receive the document content, so you can remove Safe Harbor identifiers before text leaves your device. You remain responsible for confirming all 18 categories, including dates other than the year, are removed. This mode is limited to text and documents with a text layer. The guide to client-side keys explains the trade-offs.

A practical workflow for clinics and health-tech teams

A workable routine combines automated detection, a human check and a clear rule about what may leave your systems.

1. Select the HIPAA profile. LEXARIS includes a HIPAA compliance profile alongside GDPR, UK GDPR, CCPA/CPRA and others, set out in the compliance profiles section.
2. De-identify. Replace identifiers with tokens. The tool detects more than 116 data types, including names, addresses, phone numbers, email addresses, account details and medical codes.
3. Check dates, places and free text. Safe Harbor is strict about dates other than the year and about geography smaller than a state. Review these carefully, along with any mention of relatives or employers.
4. Pass the GO/NO-GO check. When you send through LEXARIS, an unchecked document cannot be sent. The check reduces risk; review the result before sharing.
5. Send to the AI. The AI works on tokens only.
6. Restore inside your environment. Put the real details back into the answer where authorised staff can see them, and keep the key access-controlled.

If you also work under the GDPR

HIPAA de-identification and GDPR anonymisation are different tests. Data de-identified under Safe Harbor may still be personal data under the GDPR, particularly when a re-identification code and key exist.

If you handle patients or users in both the US and Europe, read pseudonymisation versus anonymisation under the GDPR alongside this guide, and involve both your HIPAA privacy officer and your data protection officer.