Protected Health Information (PHI)

The foundational federal regulation that establishes national standards for the protection of individually identifiable health information. It defines the 18 identifiers that constitute PHI and mandates that covered entities and business associates obtain written authorization before using or disclosing PHI for purposes beyond treatment, payment, and healthcare operations. The rule also grants patients rights over their own health information, including the right to access and request amendments to their records.

A definitive de-identification standard requiring the removal of 18 specific identifiers from a dataset before it is no longer considered PHI. These identifiers include:

• Names and geographic subdivisions smaller than a state
• All dates directly related to an individual (except year)
• Telephone numbers, fax numbers, and email addresses
• Social security numbers, medical record numbers, and health plan beneficiary numbers
• Full-face photographs and biometric identifiers Once stripped of these identifiers with no actual knowledge that the data could be re-identified, the dataset is considered de-identified.

An alternative de-identification pathway where a qualified statistician applies generally accepted statistical and scientific principles to render the risk of re-identification very small. The expert must document the methodology and justify the determination. This method is critical for AI model training because it allows for the retention of certain dates or geographic details—such as admission year or zip code—that may be essential features for predictive models, provided the overall re-identification risk is acceptably low.

A partially de-identified dataset that excludes direct identifiers but may retain dates, geographic information, and unique codes. An LDS is not fully de-identified and is still considered PHI, but it can be used for research, public health, and healthcare operations without individual authorization if the recipient signs a Data Use Agreement (DUA). This is a pragmatic middle ground for AI development, allowing temporal analysis and geospatial modeling that would be impossible with Safe Harbor-compliant data.

A legally binding contract between a HIPAA-covered entity and a vendor that creates, receives, maintains, or transmits PHI on its behalf. For AI vendors processing clinical data, a BAA is mandatory and extends HIPAA liability directly to the vendor. The agreement specifies:

• Permitted uses and disclosures of PHI
• Mandatory breach notification timelines
• Requirements for subcontractor compliance
• Data destruction or return obligations upon contract termination Without a signed BAA, any PHI exposure constitutes an unsecured breach.

The statistical probability that de-identified data can be linked back to a specific individual by cross-referencing with external datasets. Key attack vectors include:

• Linkage attacks: Matching quasi-identifiers like ZIP code, gender, and date of birth against public voter rolls
• Prosecutor risk: The probability that a known individual is present in the dataset
• Journalist risk: The probability that any individual in the dataset can be re-identified For AI training pipelines, ongoing re-identification risk assessments are essential, especially when combining multiple clinical datasets that may inadvertently create unique identifying combinations.