- De-identified data is no longer PHI, so the Privacy Rule stops restricting it. Data that is re-identified becomes PHI again.
- Safe Harbor removes 18 identifiers and cuts every patient date to the year. Expert Determination can keep more detail but needs a documented analysis by a qualified expert.
- A limited data set is still PHI: it needs a data use agreement and is limited to research, public health and health care operations.
- DICOM header profiles are one step. Burned-in text, faces in 3D renderings and free-text descriptions need their own checks.
What is HIPAA de-identification?
HIPAA de-identification is the process that turns protected health information (PHI) into information the Privacy Rule no longer covers. The standard is one sentence in 45 CFR 164.514(a): "Health information that does not identify an individual and with respect to which there is no reasonable basis to believe that the information can be used to identify an individual is not individually identifiable health information" (45 CFR 164.514).
There are two ways to meet it: Expert Determination, "a formal determination by a qualified expert", and Safe Harbor, the removal of listed identifiers with no actual knowledge that the rest could identify anyone (HHS OCR guidance). The rule binds covered entities (health care providers that bill electronically, health plans and clearinghouses) and their business associates. A business associate may de-identify records "only to the extent such activity is authorized by their business associate agreement".
Neither method makes identification impossible: "Although the risk is very small, it is not zero." For the general term outside health care, see the de-identification glossary entry.
Two edge cases: genomes, unsettled under Safe Harbor, and federated training, where data stays with its custodian.
Is de-identified data still PHI?
No. HHS says "the Privacy Rule does not restrict the use or disclosure of de-identified health information, as it is no longer considered protected health information." The regulation says the same in 45 CFR 164.502(d): "The requirements of this subpart do not apply to information that has been de-identified in accordance with the applicable requirements of" section 164.514 (45 CFR 164.502). Three points keep that answer honest:
- The key stays private. Disclosing a code designed to let the records be re-identified "constitutes disclosure of protected health information" (164.502(d)(2)(i)).
- Re-identified data is PHI again. If the holder links records back to people, HHS says the information "would again be protected by the Privacy Rule, as it would meet the definition of PHI."
- Other rules can still apply. HHS notes that "other federal protections also may apply, such as those found in Family Educational Rights and Privacy Act (FERPA) or the Common Rule", and state laws add their own. This page is general information, not legal advice.
Safe Harbor vs Expert Determination: which method fits?
Safe Harbor is a checklist an organization runs itself. Expert Determination is an analysis a qualified person signs. The practical differences show up in the detail a dataset can keep:
| Question | Safe Harbor, 164.514(b)(2) | Expert Determination, 164.514(b)(1) |
|---|---|---|
| Who decides | The covered entity or its business associate, by removing the 18 listed identifiers | A person with statistical and scientific knowledge and experience; HHS names no required degree |
| The test | Identifiers removed, and no actual knowledge that the rest could identify anyone | Very small risk that an anticipated recipient could identify anyone, alone or with other reasonably available information |
| Dates of birth, visits and lab tests | Year only; ages over 89 grouped as 90 or older | No fixed list; the expert decides what detail can stay for the recipient |
| Location | State, or the first three ZIP digits where that area has more than 20,000 people | No fixed list; finer geography can stay if other fields are generalized |
| Patient codes and hashes | Only a code that meets 164.514(c); a code derived from PHI generally has to go | Codes derived from PHI can stay if the expert approves and the keys are not disclosed |
| Paperwork | Not prescribed; keep a record of what was removed and how it was checked | Methods and results documented and available to OCR on request |
An expert trades detail between fields. HHS describes one data set with "detailed geocodes and generalized aged values (e.g., 5-year age ranges)" and another with coarse geography and age in days from birth. For business records outside health care, see the guide to de-identifying company data.
How does the HIPAA Safe Harbor method work?
Safe Harbor has two parts, and both must hold. The 18 listed identifiers of the patient and of the patient's relatives, employers and household members are removed, and the covered entity "does not have actual knowledge" that what remains could identify anyone (45 CFR 164.514(b)(2)). The full list, quoted word for word with where each item turns up, is on the 18 HIPAA identifiers page.
Three rules catch most teams out:
- Partial identifiers fail. HHS says "a data set that contained patient initials, or the last four digits of a Social Security number, would not meet the requirement of the Safe Harbor method for de-identification" (HHS OCR guidance).
- Every date tied to the patient goes to the year, test dates included: "Dates associated with test measures, such as those derived from a laboratory report, are directly related to a specific individual and relate to the provision of health care."
- Free text counts: "an identifier listed in the Safe Harbor standard must be removed regardless of its location in a record if it is recognizable as an identifier."
Safe Harbor needs no outside sign-off, but with every date cut to the year, a dataset can no longer show the days between a scan and surgery.
What is HIPAA Expert Determination, and who can do it?
Under 45 CFR 164.514(b)(1), "A person with appropriate knowledge of and experience with generally accepted statistical and scientific principles and methods for rendering information not individually identifiable" applies those methods, "determines that the risk is very small that the information could be used, alone or in combination with other reasonably available information, by an anticipated recipient to identify an individual who is a subject of the information", and "Documents the methods and results of the analysis that justify such determination" (45 CFR 164.514).
No license is required. HHS says "There is no specific professional degree or certification program for designating who is an expert at rendering health information de-identified." In an enforcement review, "OCR would review the relevant professional experience and academic or other training of the expert used by the covered entity, as well as actual experience of the expert using health information de-identification methodologies" (HHS OCR guidance).
"Very small" has no number: HHS says no numerical level "is deemed to universally meet the "very small" level". The expert sets it for the recipient and the setting, so one dataset can pass for one recipient and fail for another. Demographics weigh heavily: date of birth, sex and five-digit ZIP code together are, by one estimate HHS cites, "unique for over 50% of residents in the United States". The work usually runs like this:
- The expert assesses who will receive the data and what other sources they could link it to.
- The expert recommends changes: suppression (removing values or records), generalization (such as five-year age bands) or perturbation (close but different values). OCR sets no universal value for k in k-anonymity.
- The data team applies the changes and the expert tests again; HHS says this "may require several iterations".
- The expert writes up methods and results, which must be "documented and made available to OCR upon request."
No expiry date is required, though "certain de-identification practitioners use the approach of time-limited certifications", and later releases to the same recipient need a fresh look. An expert may also approve keyed hashes derived from PHI if the keys "are not disclosed, including to the recipients of the de-identified information."
What is a limited data set, and when do you need a data use agreement?
A limited data set sits between full PHI and de-identified data: it "is protected health information that excludes the following direct identifiers" of the patient and of relatives, employers and household members (45 CFR 164.514(e)). Sixteen direct identifiers go, but the set may keep "town or city, State, and zip code", and dates and ages are not on its list. Because it is still PHI, a covered entity may use or disclose one "only for the purposes of research, public health, or health care operations", and only under a data use agreement.
- A HIPAA data use agreement must "Establish the permitted uses and disclosures" and "Establish who is permitted to use or receive the limited data set".
- The recipient must not use or disclose the data beyond the agreement, must "Use appropriate safeguards", must report misuse it learns of, and must bind its agents to "the same restrictions and conditions".
- It must "Not identify the information or contact the individuals."
| Item | Safe Harbor de-identified data | Limited data set |
|---|---|---|
| Still PHI? | No | Yes |
| Dates of birth, admission, discharge and service | Year only | May stay |
| Town or city, state and ZIP code | State, and a three-digit ZIP only where allowed | May stay |
| Names, contact details, record and account numbers, full-face photos | Removed | Removed |
| Allowed purposes | Not restricted by the Privacy Rule | Research, public health or health care operations |
| Data use agreement | Optional | Required |
For de-identified data, a data use agreement is optional: "The Privacy Rule does not limit how a covered entity may disclose information that has been de-identified" (HHS). A covered entity may still ask for one. For a clause-by-clause outline, see the data use agreement template.
How does DICOM anonymization work?
Medical images hold identifiers in three places: the DICOM header, text burned into the pixels, and the anatomy itself. DICOM PS3.15 Annex E, the standard's attribute confidentiality profiles, covers "the removal and replacement of Attributes within a DICOM Data Set" that could leak identifying information (DICOM PS3.15 Annex E). Its Table E.1-1 gives each attribute an action code: the basic profile sets Patient's Name (0010,0010) and Patient's Birth Date (0010,0030) to Z, and Institution Name (0008,0080) to X/Z/D.
- Burned-in text. Unless an option is chosen, the basic profile "does not address information in the pixels" (section E.2). The Clean Pixel Data Option removes burned-in identifying text; DICOM notes "CT images do not normally contain such burned in annotation, whereas Ultrasound images routinely do", and that "it is not sufficient to superimpose an overlay or graphic annotation or shutter" (section E.3).
- Faces. For "high-resolution cross-sectional imaging of the entire head and neck", DICOM warns that "a 3D volume or surface rendering of the pixel data may be sufficient to allow identification" (section E.3.2).
- Descriptions. DICOM's own example of identifying free text is the study description "CT chest abdomen pelvis - 55F Dr. Smith" (section E.3.5).
- Dates. The modified-dates option keeps "the gross longitudinal temporal relationships between images obtained on different dates", and shifting all dates from an arbitrary event can keep exact intervals (section E.3.6).
| Code | Action, as DICOM PS3.15 defines it |
|---|---|
| D | replace with a non-zero length value that may be a dummy value and consistent with the VR |
| Z | replace with a zero length value, or a non-zero length value that may be a dummy value and consistent with the VR |
| X | remove Attribute, and if the Attribute is a Sequence, remove all Sequence Items and their contained Attributes |
| K | keep (unchanged for non-Sequence Attributes, cleaned for Sequences) |
| C | clean, that is replace with values of similar meaning known not to contain identifying information and consistent with the VR |
| U | replace with a non-zero length UID that is internally consistent within a set of Instances |
DICOM itself says the profiles do "not guarantee that all individually identifying information will be removed" and do "not replace a de-identification process, but should be part of it." Whether a study meets HIPAA still depends on Safe Harbor or an Expert Determination.
HIPAA de-identification example: a clinic visit export (illustrative)
An illustration written for this page: a fictional outpatient clinic prepares a visit export for an outside research team, once under each method. The field types are real; no patient, value or clinic is.
| Field in the export | Under Safe Harbor | Under an Expert Determination (one possible outcome) |
|---|---|---|
| Patient name, phone, email | Removed | Removed |
| Medical record number | Replaced by a random study code kept under 164.514(c) | Replaced by a random study code |
| Date of birth | Year only; ages over 89 grouped as 90 or older | Age in five-year bands |
| Visit and lab dates | Year only | Shifted by a per-patient offset, so intervals stay exact |
| Home ZIP code | First three digits, or 000 where the area has 20,000 people or fewer | First three digits, dropped for patients with rare diagnoses |
| Diagnosis and medication codes | Kept | Kept; rare codes rolled up to their parent code |
| Physician note text | Every listed identifier removed, then a sample read by a person | The same, plus occupations and family details the expert flags |
| Treating physician name | May stay; HHS does not require removing provider names | Removed at the expert's request |
The Safe Harbor version loses the time between visits. The expert version keeps it, but only for the recipient and conditions named in the expert's report.
Common mistakes with HIPAA de-identification
Most failures come from treating de-identification as one filter instead of a documented process.
- Calling records de-identified because names were removed. Safe Harbor needs all 18 categories gone, plus no actual knowledge.
- Keeping partial identifiers: initials, the last four digits of a Social Security number, or full dates kept "for internal use".
- Hashing record numbers without a secret key. DICOM warns that unencrypted hashes "are vulnerable to trivial dictionary based attacks."
- Calling a limited data set de-identified. It is still PHI and needs a data use agreement.
- Running a DICOM header tool and skipping burned-in text in ultrasound images and scanned documents.
- Reusing an old Expert Determination for a new recipient or a new release without asking the expert.
- Sending a contractor the whole chart when only some fields are needed. Under the minimum necessary rule in 164.514(d), a covered entity "may not use, disclose or request an entire medical record" unless it is "specifically justified".
Is de-identified health data also a data asset?
It can be. AI data buyers have asked Origin for longitudinal EHR records (patient records over time with diagnoses, medications, labs and outcomes), imaging with reports (CT, MRI, X-ray and pathology slides), behavioral health and specialty care records, and, as the newest request, doctor-patient conversation audio with transcripts. A buyer has also told Origin what these deals need to get through: de-identification signed off by an independent third party under HIPAA's Expert Determination method, not by the seller; IRB approval, or patient consent that covers commercial licensing; and a clear chain of rights showing who collected the data, who owns it and who can license it.
What must come out first: every identifier the chosen method requires, in fields, free text, image headers, pixels and audio, and any records the practice cannot show it has the right to share. A limited data set is not a shortcut: it is still PHI, a covered entity "may not sell protected health information" without the patient's authorization, and the research exception covers only "a reasonable cost-based fee" (45 CFR 164.502).
For the three requirements in turn, see health data licensing requirements. For charts, see licensing patient records; for studies, licensing medical imaging data.
Origin does not perform Expert Determination or IRB review. Origin is an independent data partner: it helps a practice work out what it holds, which of the three requirements it already meets and what is missing. Your practice decides what is shared and signs any license itself, and a buyer may pay Origin when a deal closes.
Does your company hold records like these? See what your data is worth.
Common questions
What is HIPAA de-identification in simple terms?
It is the process of removing or changing the details in health records that could identify a patient, so the records stop being protected health information. HIPAA allows two methods: Safe Harbor, which removes 18 listed identifiers, and Expert Determination, in which a qualified expert documents that the risk of identification is very small. Once either method is met, the Privacy Rule no longer restricts the data.
Is de-identified data considered PHI under HIPAA?
No. HHS says de-identified health information is no longer considered protected health information, so the Privacy Rule does not restrict its use or disclosure. Two limits apply: disclosing the code or key that could re-identify the records counts as disclosing PHI, and records that are re-identified become PHI again. State privacy laws and other federal rules may still apply to the de-identified data.
Who can perform a HIPAA Expert Determination?
A person with appropriate knowledge of and experience with generally accepted statistical and scientific methods for making information not individually identifiable. HHS says no specific degree or certification is required. Experts usually come from statistics, mathematics or other sciences, and OCR would look at their training and their actual experience de-identifying health information. The expert must document the methods and results of the analysis.
How long is a HIPAA Expert Determination valid?
The Privacy Rule sets no expiry date. HHS notes that some practitioners issue time-limited certifications because technology and available data change over time. Data already released under a lapsed certification is not automatically unprotected, but new releases to the same recipient, such as monthly extracts, need the expert to check whether the method still keeps the risk very small.
Is a limited data set the same as de-identified data?
No. A limited data set is still protected health information. It removes 16 direct identifiers, such as names, street addresses and record numbers, but may keep dates, town or city, state and ZIP code. It may be used only for research, public health or health care operations, and only under a data use agreement that bans re-identifying or contacting the people in it.
Can a limited data set be sold for a fee?
Generally not for profit. A limited data set is still PHI, and HIPAA bars a covered entity or business associate from selling PHI without the patient's authorization. The rule's exception for research disclosures, including limited data sets, applies only where the payment is a reasonable cost-based fee to prepare and transmit the data. This is general information; check the facts of a deal with counsel.
Does a DICOM de-identification profile make images HIPAA de-identified?
Not by itself. DICOM PS3.15 Annex E says its profiles do not guarantee that all identifying information is removed and should be part of a wider de-identification process. Burned-in text needs the Clean Pixel Data Option, head and neck scans may need facial features removed, and free-text descriptions need cleaning. The study must still meet Safe Harbor or an Expert Determination.
Is Expert Determination required to share health data with an AI company?
HIPAA does not require it: either Safe Harbor or Expert Determination de-identifies data. In practice, Safe Harbor strips the dates and detail many AI uses need, and buyers often ask for more. One buyer told Origin that health data deals need de-identification signed off by an independent third party under Expert Determination, not by the seller, plus IRB approval or consent and a clear chain of rights.
Sources and further reading
- eCFR: 45 CFR 164.514, other requirements relating to uses and disclosures of protected health information
- eCFR: 45 CFR 164.502, uses and disclosures of protected health information: general rules
- HHS Office for Civil Rights: Guidance Regarding Methods for De-identification of Protected Health Information
- DICOM PS3.15 Annex E: Attribute Confidentiality Profiles
- DICOM PS3.15 E.2: Basic Application Level Confidentiality Profile
- DICOM PS3.15 E.3: Basic Application Level Confidentiality Options (Clean Pixel Data)
- DICOM PS3.15 E.3.2: Clean Recognizable Visual Features Option
- DICOM PS3.15 E.3.5: Clean Descriptors Option
- DICOM PS3.15 E.3.6: Retain Longitudinal Temporal Information Options
