HIPAA Compliance Ontology

A comprehensive formal verification ruleset for HIPAA Privacy Rule, Security Rule (including 2025 updates), and AI governance requirements. Version 2.0.0 with 76 constraints.

Formal Constraints

104

Data Extractors

1. Overview

The HIPAA ontology provides formal verification of AI/LLM outputs against healthcare compliance requirements. Unlike pattern-matching or ML-based approaches, aare.ai uses the Z3 SMT solver to mathematically prove whether outputs comply with HIPAA regulations.

Why Automated Reasoning? Traditional guardrails use probabilistic detection (e.g., "98% accuracy"). For healthcare compliance with penalties up to $2.1M per violation, you need mathematical certainty. Z3 provides provable guarantees.

Coverage

HIPAA Privacy Rule - PHI disclosure, de-identification (18 identifiers), minimum necessary
HIPAA Security Rule - Access controls, encryption, audit trails
2025 Security Rule Updates - MFA, asset inventory, 72-hour recovery, vulnerability scanning
AI Governance - Training data consent, human review, algorithm transparency, BAA for AI vendors
Special Protections - Psychotherapy notes, substance abuse (42 CFR Part 2), genetic info (GINA)
Business Associates - Third-party BAA, subcontractor chains, cloud provider compliance

2. What Problem Does This Solve?

Healthcare organizations deploying LLMs face significant compliance risks:

$9.77 million - Average cost of a healthcare data breach in 2024
$2.1 million - Maximum HIPAA penalty per violation
259 million - Individuals affected by healthcare breaches in 2024 (up from 27M in 2020)

The Challenge with AI in Healthcare

LLMs can inadvertently disclose PHI in responses
AI-generated recommendations may lack required disclosures
External AI vendors (ChatGPT, Claude) require Business Associate Agreements
The 2025 HIPAA Security Rule makes many safeguards mandatory (no longer "addressable")
OCR now explicitly states AI systems processing ePHI are subject to the Security Rule

How aare.ai Helps

This ontology enables real-time verification of LLM outputs before they reach users. When integrated via the MCP server, Claude can self-verify its healthcare responses against 76 formal constraints.

3. Constraint Categories

PHI Detection (15 constraints)

Detects and blocks disclosure of the 18 HIPAA identifiers:

Patient names, dates of birth, Social Security numbers
Street addresses, phone numbers, email addresses
Medical record numbers, health plan IDs, account numbers
Device identifiers, IP addresses, URLs
Biometric data, photographs, vehicle identifiers

AI Governance (6 constraints)

New requirements for AI systems processing PHI:

AI_TRAINING_DATA_CONSENT - PHI used for AI training requires explicit patient consent
AI_MODEL_BIAS_DISCLOSURE - AI recommendations must disclose model limitations
AI_HUMAN_REVIEW - AI clinical decisions require qualified human review
AI_ALGORITHM_TRANSPARENCY - AI outputs must reference algorithm version/source
AI_VENDOR_BAA - External AI vendors (OpenAI, Anthropic) require BAA
AI_RISK_ASSESSMENT - AI systems require documented risk assessment

2025 Security Rule (6 constraints)

Based on the December 2024 HHS proposed rule updates:

SECURITY_MFA_REQUIRED - Multi-factor authentication for PHI access
SECURITY_ENCRYPTION_AT_REST - PHI must be encrypted at rest
SECURITY_ASSET_INVENTORY - PHI systems must be in documented inventory
SECURITY_VULN_SCAN - 6-month vulnerability scanning requirement
SECURITY_72HR_RECOVERY - Critical systems need 72-hour recovery plan
SECURITY_NETWORK_MAP - PHI data flows must be documented

Special Protections (5 constraints)

Enhanced protections for sensitive categories:

MENTAL_HEALTH_PSYCHOTHERAPY - Psychotherapy notes require separate authorization
SUBSTANCE_ABUSE_42CFR - Substance abuse records require 42 CFR Part 2 compliance
GENETIC_GINA_COMPLIANCE - Genetic information requires GINA compliance
RESEARCH_IRB_APPROVAL - Research use of PHI requires IRB approval
RESEARCH_LIMITED_DATA_SET - Limited data sets require data use agreement

4. All Constraints (English)

Below is every constraint in the ontology, organized by category. Each includes the formal logical formula and regulatory citation.

PHI Detection

15 constraints

PHI_NAME_DISCLOSURE

Patient name must not be disclosed to unauthorized recipients

NOT(has_patient_name AND NOT recipient_authorized)

HIPAA Section 164.502(a)(1)

PHI_DOB_DISCLOSURE

Date of birth must be de-identified or authorized

NOT(has_dob AND NOT is_deidentified)

HIPAA Section 164.514(b)

PHI_SSN_ZERO_TOLERANCE

Social Security Numbers are strictly prohibited in outputs

NOT has_ssn

HIPAA Section 164.514(b)(2)(i)(L)

PHI_MRN_ZERO_TOLERANCE

Medical Record Numbers must not be disclosed externally

NOT(has_mrn AND recipient_external)

HIPAA Section 164.514(b)(2)(i)(N)

De-Identification

9 constraints

DEIDENTIFY_ALL_18

All 18 HIPAA identifiers must be removed for Safe Harbor

is_deidentified IMPLIES phi_count = 0

HIPAA Section 164.514(b)(2)

DEIDENTIFY_AGE_THRESHOLD

Ages over 89 must be aggregated to "90+"

patient_age > 89 IMPLIES age_aggregated

HIPAA Section 164.514(b)(2)(i)(C)

DEIDENTIFY_ZIP_CODE

ZIP codes must be truncated to first 3 digits

NOT has_full_zip

HIPAA Section 164.514(b)(2)(i)(B)

AI Governance

6 constraints

AI_TRAINING_DATA_CONSENT

PHI used for AI/ML training requires explicit patient consent

NOT(mentions_ai_training AND has_phi AND NOT has_consent_mention)

OCR AI Guidance December 2024

AI_MODEL_BIAS_DISCLOSURE

AI-generated clinical recommendations must disclose model limitations

ai_generated_recommendation IMPLIES has_limitation_disclosure

HHS AI Governance Framework 2024

AI_HUMAN_REVIEW

AI outputs affecting treatment require qualified human review

ai_clinical_decision IMPLIES has_human_review_mention

OCR Section 1557 Guidance December 2024

AI_VENDOR_BAA

AI vendors processing PHI must have Business Associate Agreement

uses_external_ai IMPLIES has_baa_mention

HIPAA Section 164.502(e)

2025 Security Rule

6 constraints

SECURITY_MFA_REQUIRED

Multi-factor authentication required for PHI access

accesses_phi IMPLIES has_mfa_mention

2025 HIPAA Security Rule CPG-1

SECURITY_72HR_RECOVERY

Critical systems must have 72-hour recovery plan

is_critical_system IMPLIES has_recovery_plan

2025 HIPAA Security Rule - Disaster Recovery

SECURITY_ASSET_INVENTORY

Systems with PHI must be in documented asset inventory

has_phi IMPLIES asset_inventoried

2025 HIPAA Security Rule Proposed Updates

Access Control

9 constraints

ACCESS_MINIMUM_NECESSARY

Only minimum necessary information should be disclosed

phi_count <= minimum_necessary_threshold

HIPAA Section 164.502(b)(1)

ACCESS_ROLE_PHYSICIAN

Treatment details require physician-level access

has_treatment_details IMPLIES (recipient_role_physician OR is_deidentified)

HIPAA Section 164.502(b)

Business Associate

3 constraints

BAA_THIRD_PARTY

Third-party PHI access requires Business Associate Agreement

third_party_access IMPLIES has_baa_mention

HIPAA Section 164.502(e)(1)

BAA_CLOUD_PROVIDER

Cloud services storing PHI require BAA

uses_cloud_storage IMPLIES cloud_baa_confirmed

OCR Cloud Computing Guidance

See All 76 Constraints: The full list is available in the JSON section below, or explore the interactive HIPAA demo.

5. Usage

Python (aare-core)

from aare_core import SMTVerifier, LLMParser, OntologyLoader

loader = OntologyLoader()
ontology = loader.load('hipaa-v1')
parser = LLMParser()
verifier = SMTVerifier()

llm_output = """
Based on the patient's medical record, the AI system recommends
considering metformin. This recommendation was reviewed by Dr. Smith
and has limitations - consult your physician for clinical judgment.
"""

values = parser.parse(llm_output, ontology)
result = verifier.verify(values, ontology)

print(f"Compliant: {result['verified']}")
print(f"Violations: {len(result['violations'])}")

MCP Server (Claude Desktop / Claude Code)

See the MCP Server documentation for integration.

REST API

curl -X POST https://api.aare.ai/verify \
  -H "Content-Type: application/json" \
  -H "X-Api-Key: your-api-key" \
  -d '{
    "llm_output": "Patient John Doe (SSN: 123-45-6789) ...",
    "ontology": "hipaa-v1"
  }'

6. Regulatory References

This ontology is based on the following regulatory sources:

HIPAA Privacy Rule HIPAA Security Rule 2025 Security Rule Updates 42 CFR Part 2 GINA

Key Citations

164.502(a) - Uses and disclosures of PHI
164.514(b) - De-identification Safe Harbor method
164.502(e) - Business Associate requirements
164.312 - Technical safeguards (encryption, access controls)
164.308(a)(1) - Security management process
164.404-408 - Breach notification requirements
164.524-526 - Patient access and amendment rights
42 CFR Part 2 - Substance abuse records protection

7. Full Ontology JSON

Below is the complete ontology JSON file. This can be used directly with aare-core or customized for your organization's specific requirements.

Note: This is a large file (~1800 lines). The JSON is collapsed by default. Click "Show JSON" to expand.

Loading...