Compliance decisions
De-identifying patient notes before AI
Removing a patient’s name is not enough to establish HIPAA de-identification. HHS describes two methods: Safe Harbor, with its identifier-removal and actual-knowledge conditions, and Expert Determination. Review narrative clues and the recipient’s context as well as structured fields.
For Healthcare privacy teams and security engineers
An annotated synthetic note shows why ordinary masking can leave identifying information.
Entirely synthetic patient, events and identifiers. No expert determination or product test has been performed.01
Choose a method before editing
HIPAA de-identification is a defined standard, not a synonym for redaction. Safe Harbor requires removal of specified identifiers relating to the individual and their relatives, employers or household members, alongside the condition about actual knowledge of remaining identifying information. Expert Determination relies on a qualified person’s assessment of a very small identification risk.
Choose the route based on the information needed and the intended recipient. If the task can use invented material, do that first. If real patient details must be retained, masking a few fields is not a shortcut to declaring the record outside HIPAA.
Source context: HHS: Methods for de-identification of PHI
02
Read beyond the obvious fields
The Safe Harbor list includes names, detailed geography, most date elements, telephone and fax numbers, email addresses, Social Security and medical-record numbers, health-plan and account numbers, certificate or licence numbers, vehicle and device identifiers, web addresses, IP addresses, biometric identifiers, full-face photographs and other unique identifying characteristics or codes.
Dates directly related to an individual generally retain only the year. Ages over 89 and related date elements must be aggregated into a category of age 90 or older. Geographic rules include a conditional exception for some three-digit ZIP codes. Use the complete HHS guidance rather than treating this short explanation as the full removal specification.
Source context: HHS: Methods for de-identification of PHI
03
A note can identify someone without a name
The same rules apply to structured fields and free text. A narrative may mention a distinctive employer role, a publicised clinical event or a sequence of visits that a recipient can link to an individual. Copying a cleaned paragraph into an AI prompt does not remove the clues in an attached document, filename or image.
In the fictional note below, replacing “Mara Example” with “Patient A” leaves a full visit date, age 93 and an identifying occupation. The revised wording illustrates fewer details; it is not a certified Safe Harbor output. The reviewer still needs the complete identifier checks and the actual-knowledge assessment.
Source context: HHS: Methods for de-identification of PHI
04
When useful detail needs an expert assessment
Expert Determination can assess whether information that supports the task can be retained with sufficiently small identification risk. HHS describes a person with appropriate statistical and scientific knowledge applying generally accepted principles and documenting the methods and results.
The assessment concerns the proposed data and anticipated recipients. It is not an approval of every future dataset or destination. Keep the expert’s documented conditions with the release record and revisit changes that affect the analysis. A tool’s confidence score or a list of entities it detected is not an expert determination.
Source context: HHS: Methods for de-identification of PHI
05
Check the exported artifact, not just the screen
Review the actual text or file that will leave the organisation. Inspect all pages, worksheets, comments and relevant metadata; compare that output with the chosen method. Keep a restricted record of what was reviewed, how it was transformed and which release conditions apply. Do not attach the original identified note to the review email unnecessarily.
Use synthetic fixtures to evaluate a redaction or input-control tool before it handles real records. Record missed content as a test finding rather than quietly weakening the standard. Other confidentiality duties or contractual restrictions can remain relevant even where a HIPAA de-identification method is satisfied.
Source context: HHS: Methods for de-identification of PHI
Put it into practice
Patient-note annotation exercise
Compare three treatments of one invented note. The worksheet illustrates review decisions, not a legal determination about a real dataset.
Entirely synthetic patient, events and identifiers. No expert determination or product test has been performed.
Invented original
Mara Example, MRN TEST-4821, age 93, attended on 14 March 2026.
The worksheet also includes a distinctive occupation.
Name-only masking
Patient A, age 93, attended on 14 March 2026.
Date and age details remain. This is not de-identification.
Reduced-detail example
An adult aged 90 or older discussed a routine follow-up in 2026.
Teaching example only. Apply the full chosen method to a real record.
| Item in invented note | Why review it | Illustrative treatment |
|---|---|---|
| Mara Example; MRN TEST-4821 | Name and medical-record identifier. | Remove both, rather than replacing only the visible name. |
| Seen on 14 March 2026 | Date directly related to an individual. | Retain 2026 only for a Safe Harbor review. |
| Age 93 | Age above 89. | Use the category age 90 or older. |
| Only harbour master in a small town | Distinctive context may identify the person. | Omit occupation and locality where unnecessary; assess remaining context. |
| Patient A attended on 14 March, age 93 | Masking retains multiple identifying clues. | Do not label this version de-identified. |
| Older adult discussed a routine follow-up | More general wording loses clinical detail. | Check task utility and the complete chosen method before any real release. |
Work through your review
Use the checks to organise the evidence you need. Your selections stay in this tab.
0 of 3 reviewed
Example files for this task
Keep the source material and the instructions together. You can also download the complete worksheet or matrix as CSV.
patient-note-deidentification-exercise.mdInspect
# Patient-note annotation exercise
Entirely synthetic patient, events and identifiers. No expert determination or product test has been performed.
Compare three treatments of one invented note. The worksheet illustrates review decisions, not a legal determination about a real dataset.
| Item in invented note | Why review it | Illustrative treatment |
| --- | --- | --- |
| Mara Example; MRN TEST-4821 | Name and medical-record identifier. | Remove both, rather than replacing only the visible name. |
| Seen on 14 March 2026 | Date directly related to an individual. | Retain 2026 only for a Safe Harbor review. |
| Age 93 | Age above 89. | Use the category age 90 or older. |
| Only harbour master in a small town | Distinctive context may identify the person. | Omit occupation and locality where unnecessary; assess remaining context. |
| Patient A attended on 14 March, age 93 | Masking retains multiple identifying clues. | Do not label this version de-identified. |
| Older adult discussed a routine follow-up | More general wording loses clinical detail. | Check task utility and the complete chosen method before any real release. |
## Review steps
- Record the chosen legal method: Distinguish Safe Harbor review, documented Expert Determination and ordinary masking.
- Review the complete output: Apply identifier and contextual checks to free text, all file content and relevant metadata.
- Keep the release conditions: Record the recipient, remaining information and reason the selected method is satisfied; do not invent reviewer sign-off.
## Invented original
Mara Example, MRN TEST-4821, age 93, attended on 14 March 2026. The note describes the only harbour master in a small town seeking a routine follow-up.
## Ordinary masking, insufficient by itself
Patient A, age 93, attended on 14 March 2026. The note describes the only harbour master in a small town seeking a routine follow-up.
## Reduced-detail teaching example
An adult aged 90 or older discussed a routine follow-up in 2026.
The last version is a teaching example only. This exercise does not establish that a real record meets Safe Harbor or Expert Determination.
## Source and scope
Guide: https://aona.ai/resources/guides/hipaa-de-identification-ai-patient-notes/
Source check: 21 September 2026. General information; no professional approval or installed-product result is represented.
- HHS: Methods for de-identification of PHI: https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html
patient-note-deidentification-exercise.csvInspect
Item in invented note,Why review it,Illustrative treatment
Mara Example; MRN TEST-4821,Name and medical-record identifier.,"Remove both, rather than replacing only the visible name."
Seen on 14 March 2026,Date directly related to an individual.,Retain 2026 only for a Safe Harbor review.
Age 93,Age above 89.,Use the category age 90 or older.
Only harbour master in a small town,Distinctive context may identify the person.,Omit occupation and locality where unnecessary; assess remaining context.
"Patient A attended on 14 March, age 93",Masking retains multiple identifying clues.,Do not label this version de-identified.
Older adult discussed a routine follow-up,More general wording loses clinical detail.,Check task utility and the complete chosen method before any real release.
Before you proceed
Keep these distinctions clear
- Pseudonyms can retain the link
- Replacing a name with a stable label is not, by itself, HIPAA de-identification.
- A clean paragraph can hide a risky file
- Check the actual exported file, including other content that was not visible in the edited paragraph.
Apply it to employee AI use
Bring your actual data path.
Aona can support a scoped evaluation of sensitive-input detection and file redaction on supported installed paths.
Detected entities and redacted output do not establish HIPAA Safe Harbor or Expert Determination. Aona does not provide an expert’s legal de-identification decision.
Use the invented note to compare the intended transformation with the actual exported content and record any remaining identifiers.
Review your use caseFAQ
Questions for this decision
Is replacing names with Patient A enough?
Can exact dates stay in a Safe Harbor dataset?
Does a redaction tool certify de-identification?
Is a limited data set the same as de-identified information?
Evidence behind the guide
Sources and scope
Prepared by Aona. Sources checked 2026-09-21. The cited material supports the specific points below; it does not certify a product or your use case.
- HHS: Methods for de-identification of PHI
Safe Harbor and Expert Determination, identifiers in free text, actual knowledge and residual identification risk.
regulator · checked 2026-09-21