Home / Blog / De-identifying clinical notes for AI training: what HIPAA Safe Harbor requires
De-identifying clinical notes for AI training: what HIPAA Safe Harbor requires
Under HIPAA's Safe Harbor method, health data counts as de-identified only after 18 types of identifier are removed (including all date elements except the year) and the covered entity has no actual knowledge that what remains could identify a person. Free-text clinical notes are the hard case, because identifiers hide inside sentences rather than in labelled fields.
Clinical notes are some of the most useful text for medical AI and some of the hardest to share safely. In the United States, the HIPAA Privacy Rule gives two routes to de-identification: Safe Harbor and Expert Determination (45 CFR 164.514(b); HHS guidance). This post covers what Safe Harbor requires and why notes make it harder than structured records.
The 18 Safe Harbor identifiers
Safe Harbor requires removing these identifiers of the individual and of their relatives, employers and household members:
- Names
- Geographic subdivisions smaller than a state, including street address, city, county and ZIP code (the first three ZIP digits may stay if that area has more than 20,000 people)
- All elements of dates except the year (birth, admission, discharge, death), and all ages over 89
- Telephone numbers
- Fax numbers
- Email addresses
- Social Security numbers
- Medical record numbers
- Health plan beneficiary numbers
- Account numbers
- Certificate and license numbers
- Vehicle identifiers and serial numbers, including license plates
- Device identifiers and serial numbers
- Web URLs
- IP addresses
- Biometric identifiers, including finger and voice prints
- Full-face photographs and comparable images
- Any other unique identifying number, characteristic or code
On top of the list, the covered entity must have no actual knowledge that the remaining information could identify the individual.
Why free-text notes are harder
In a structured record, the phone number lives in a phone-number field. In a note, it can appear anywhere: "pt's daughter (Maria, 555-0134) will collect reports". Identifiers in notes show up as:
- Names of relatives, referring doctors and employers mentioned in passing
- Relative dates that reveal exact dates ("discharged two days after her 60th birthday")
- Rare conditions, occupations or events that single someone out in a small clinic
- Handwritten notes, where text has to be transcribed before anything can be removed
Automated tools catch most of these, but not all. A practical pipeline combines automated detection, human review of a sample, and rules for rare details.
Dates: year only under Safe Harbor
Safe Harbor allows only the year to remain. Many AI uses need finer timing, such as the gap between visits. Common approaches are converting dates to intervals ("day 0, day 14") or shifting all dates for one patient by the same random offset. Whether a shifted date still counts as a date element under Safe Harbor is a question for your privacy counsel. Expert Determination is the usual route when month-level or interval data is needed.
When Expert Determination fits better
Under Expert Determination, a qualified expert applies statistical or scientific methods and documents that the risk of re-identification is very small, for the specific data and the specific recipients. It can permit more detail than Safe Harbor (for example month-level dates or three-digit geography) when the expert's analysis supports it. It costs more and takes longer, but it is often the better fit for training data where timing or detail matters.
Outside the U.S.
HIPAA applies to U.S. covered entities and their business associates. Under the EU GDPR, health data is a special category (Article 9), pseudonymised data is still personal data, and only data that is truly anonymous falls outside the regulation (Recital 26). Other countries have their own rules. A sensible baseline is to use the Safe Harbor list as a checklist everywhere, then apply whatever the local law adds.
How we handle clinical data
At Datadhara, de-identification happens before data leaves the clinic, using the Safe Harbor list as a baseline checklist plus review for rare details. Every record keeps a reference to the clinic's written agreement. See the healthcare data page and our methodology, or look at the structure in the free synthetic clinic-note samples.
This is general information, not legal advice.
Questions
How many identifiers does HIPAA Safe Harbor require removing?
Eighteen, ranging from names and geographic subdivisions smaller than a state to all date elements except the year, plus any other unique identifying number, characteristic or code.
Can de-identified clinical notes keep the month of a visit?
Not under Safe Harbor, which allows only the year. Month-level dates may be possible under Expert Determination if a qualified expert documents that re-identification risk is very small.
Is pseudonymised data anonymous under GDPR?
No. Under the GDPR, pseudonymised data is still personal data. Only data that can no longer be attributed to a person by reasonably likely means is outside the regulation.
Tell us what your models are missing
A few lines is enough: the domain, the format, and roughly how much.
Request data →