HIPAA offers two ways to de-identify protected health information: Safe Harbor and Expert Determination. Safe Harbor removes specified identifiers. Expert Determination documents that identification risk is very small for a defined dataset and anticipated recipient. The useful choice depends on the data you need to preserve and the release you can govern.
Takeaways:
- Removing names is not enough; Safe Harbor addresses 18 identifier categories.
- Expert Determination is a documented risk analysis by a qualified expert, not an internal opinion.
- Dates, free text, rare events, geography, and outside datasets often decide which path fits.
- Keep a release record that ties the method, data version, recipient, controls, and approver together.
A date field forces the real decision
A quality team wants to study 30-day readmissions. The analyst needs admission and discharge dates. A privacy reviewer points out that Safe Harbor generally removes date elements more specific than year when they are directly related to an individual. Dropping the dates may break the analysis. Keeping them without a valid method may keep the dataset inside HIPAA.
That is the trade. The HHS de-identification guidance explains the two methods under 45 CFR 164.514(b). Safe Harbor follows the specified identifier-removal route and requires no actual knowledge that remaining information could identify the person. Expert Determination requires a person with suitable statistical and scientific knowledge to document that identification risk is very small.
Neither method begins with a tool. Begin with the intended data, recipient, purpose, linkable outside information, and the fields the work cannot lose.
The method decision record
Use this record before anyone transforms the dataset. A later reviewer should be able to tell what was approved without reconstructing a meeting.
| Decision field | Safe Harbor route | Expert Determination route |
|---|---|---|
| Data need | Analysis can work after specified identifiers are removed | Useful fields would be lost under the fixed removal rule |
| Main test | Remove required identifier categories and retain no actual knowledge of identifiability | Qualified expert documents that risk is very small |
| Context | Rule is applied to the dataset; actual knowledge still matters | Dataset, anticipated recipient, environment, controls, and outside data shape the analysis |
| Dates and geography | Detail is sharply limited by the rule | Detail may be retained if the expert’s documented result supports it |
| Evidence | Field inventory, transformation rule, validation result, release approval | Expert report, methods, results, conditions, data version, release approval |
| Change trigger | New columns, free text, data source, or known identification route | Any change that can alter the expert’s assumptions or risk result |
The record should name a versioned dataset. “Claims extract” is not enough if next month’s extract contains two new quasi-identifiers and an unreviewed notes column.
Safe Harbor is mechanical until it isn’t
The regulation at 45 CFR 164.514 lists the Safe Harbor identifiers and adds the actual-knowledge condition. That second part matters. A dataset can pass a field checklist while a covered entity knows that a rare occupation, event, or combination points to one person.
Free text deserves its own test. A clinical note can contain a name, phone number, exact date, workplace, location, record number, or a unique story. Removing structured columns does not clean the narrative. Run realistic samples through the transformation, then inspect misses and false confidence.
Codes need care too. The rule permits a code or other means of record identification in limited circumstances, but the code cannot be derived from information about the person, and the mechanism for re-identification cannot be disclosed. Tokenizing a medical record number with a reversible house key is not automatically a public de-identification method.
Expert Determination keeps utility under conditions
HHS says there is no single numerical threshold that universally defines “very small” identification risk. The expert considers the data and the environment in which it is expected to be used. HHS also requires the methods and results that justify the determination to be documented.
That flexibility can preserve dates, finer geography, or other fields needed for valid analysis. It also creates conditions. The conclusion may assume a named recipient, contract limits, access controls, no onward disclosure, a specific outside-data environment, or a fixed dataset. Change those facts and the old determination may no longer answer the new release.
Ask the expert report to state its dataset fingerprint, anticipated recipients, data-use conditions, transformations, risk measures, unresolved limits, and re-evaluation triggers. Store it with the release record, not in an inbox.
An AI dataset pressure test
Suppose a team wants 20,000 old notes to evaluate a summarization model. The easiest mistake is to remove patient names and call the rest anonymous. Notes can contain dates, clinicians, relatives, addresses, workplaces, facilities, rare diagnoses, and singular events. Model prompts, evaluation exports, and error logs can create new copies after the original transformation.
Write four boundaries before release:
- the source fields and note sections included;
- the de-identification method and evidence;
- every recipient and processing environment; and
- the outputs, logs, retention, deletion, and onward-use limits.
Then test sampled records. Record residual identifiers and alter the method. A pass rate invented for the project would be worse than no number, so use documented acceptance rules and human review proportionate to the risk.
De-identified is not the same as a limited data set
The HIPAA Privacy Rule also defines a limited data set under 45 CFR 164.514(e). It can retain some information that Safe Harbor would remove, but it remains protected health information and requires a data use agreement for permitted purposes. Do not use “limited” and “de-identified” as synonyms in architecture diagrams or contracts.
We think every release needs a one-page cover sheet: data version, method, evidence location, recipient, permitted purpose, controls, expiry, onward-disclosure rule, and owner. Boring paperwork is useful when someone asks six months later why exact dates left the source system.
Product scope stays separate
Patient Square is an AI clinical platform. Practice Copilot brings the whole practice under one AI copilot: an ambient AI Medical Scribe that hands back a structured SOAP note, ICD-10 suggestions, and a prescription draft minutes after the visit, plus a bundled AI EHR, scheduling, and messaging as you move up the plan. Hospitals get Hospital Copilot.
This article does not claim that Patient Square turns clinical notes into de-identified datasets or provides Expert Determination. For product privacy and security questions, use the Patient Square security page and the AI scribe data-ownership guide. If you are mapping an AI feature that may receive PHI, the HIPAA-compliant AI feature card covers the live-service review rather than secondary data release.