PHI is any piece of health data that can be used to figure out who the patient is. It’s not just the medical diagnosis; it’s the diagnosis plus the patient's name, birth date, address, or even their IP address. If you can link the health information back to a specific person, it’s PHI, and it is heavily protected by law.
PHI is any piece of health data that can be used to figure out who the patient is. It’s not just the medical diagnosis; it’s the diagnosis plus the patient's name, birth date, address, or even their IP address. If you can link the health information back to a specific person, it’s PHI, and it is heavily protected by law.
Under HIPAA, there are 18 specific identifiers that, when linked with health information, constitute PHI: Names Geographic subdivisions smaller than a state (e.g., street address, city, ZIP code) All elements of dates (except year) directly related to an individual (birth date, admission date, etc.) Telephone numbers Fax numbers Email addresses Social Security numbers Medical record numbers Health plan beneficiary numbers Account numbers Certificate/license numbers Vehicle identifiers and serial numbers Device identifiers and serial numbers Web Universal Resource Locators (URLs) Internet Protocol (IP) address numbers Biometric identifiers (fingerprints, voiceprints) Full-face photographic images Any other unique identifying number, characteristic, or code De-identification: To use health data for AI training or research without patient consent, it must be de-identified. HIPAA provides two methods: Safe Harbor: Removal of all 18 identifiers listed above. Expert Determination: A qualified statistician certifies that the risk of re-identification is very small.
# De-identifying text using Microsoft Presidio (Open Source PII/PHI detection)
# pip install presidio-analyzer presidio-anonymizer
from presidio_analyzer import AnalyzerEngine
from presidio_anonymizer import AnonymizerEngine
# Sample clinical note containing PHI
clinical_note = """
Patient John Doe (SSN: 123-45-6789, DOB: 1980-05-15)
presented to Mount Sinai Hospital on 2023-10-25 with acute chest pain.
Contact: john.doe@email.com or 555-0198.
"""
# Initialize engines
analyzer = AnalyzerEngine()
anonymizer = AnonymizerEngine()
# Analyze the text to find PHI entities
analyzer_results = analyzer.analyze(text=clinical_note, language='en')
# Anonymize (redact or replace) the identified PHI
anonymized_result = anonymizer.anonymize(
text=clinical_note,
analyzer_results=analyzer_results,
operators={"DEFAULT": "replace", "PERSON": "mask"} # Custom masking rules
)
print("--- Original ---")
print(clinical_note)
print("\n--- De-identified (Safe for AI Training) ---")
print(anonymized_result.text)
# Output will replace names, SSNs, dates, and emails with tags like <PERSON>, <US_SSN>, etc.Handling PHI correctly is the single biggest compliance risk for Healthcare AI companies: BAA Requirement: Any AI vendor processing PHI for a healthcare provider must sign a Business Associate Agreement (BAA), accepting legal liability for data breaches. Cloud Architecture: AI infrastructure must be HIPAA-compliant (e.g., encrypted at rest and in transit, strict access controls, audit logging). Model Inversion Risk: There is emerging research showing that some AI models can inadvertently "memorize" and regurgitate PHI from their training data, making de-identification a critical pre-processing step. Global Equivalents: Outside the US, similar concepts apply (e.g., "Special Category Data" under GDPR in Europe).
A sealed, confidential personnel file. The file itself isn't dangerous, but if it contains your name, salary, and performance reviews, it must be kept in a locked cabinet, and only authorized people can view it.
PHI is any piece of health data that can be used to figure out who the patient is. It’s not just the medical diagnosis; it’s the diagnosis plus the patient's name, birth date, address, or even their IP address. If you can link the health information back to a specific person, it’s PHI, and it is heavily protected by law.
Under HIPAA, there are 18 specific identifiers that, when linked with health information, constitute PHI: Names Geographic subdivisions smaller than a state (e.g., street address, city, ZIP code) All elements of dates (except year) directly related to an individual (birth date, admission date, etc.) Telephone numbers Fax numbers Email addresses Social Security numbers Medical record numbers Health plan beneficiary numbers Account numbers Certificate/license numbers Vehicle identifiers and serial numbers Device identifiers and serial numbers Web Universal Resource Locators (URLs) Internet Protocol (IP) address numbers Biometric identifiers (fingerprints, voiceprints) Full-face photographic images Any other unique identifying number, characteristic, or code De-identification: To use health data for AI training or research without patient consent, it must be de-identified. HIPAA provides two methods: Safe Harbor: Removal of all 18 identifiers listed above. Expert Determination: A qualified statistician certifies that the risk of re-identification is very small.
Handling PHI correctly is the single biggest compliance risk for Healthcare AI companies: BAA Requirement: Any AI vendor processing PHI for a healthcare provider must sign a Business Associate Agreement (BAA), accepting legal liability for data breaches. Cloud Architecture: AI infrastructure must be HIPAA-compliant (e.g., encrypted at rest and in transit, strict access controls, audit logging). Model Inversion Risk: There is emerging research showing that some AI models can inadvertently "memorize" and regurgitate PHI from their training data, making de-identification a critical pre-processing step. Global Equivalents: Outside the US, similar concepts apply (e.g., "Special Category Data" under GDPR in Europe).