Sectoral Profile Draft v0.1 Healthcare

r2-health-v1 — Profile for healthcare AI deployments

Extends R+2 for clinical decision support, EHR processing, telemedicine triage, and patient-facing agents. Aligns with HIPAA Security Rule, GDPR Article 9 (special category data), and DPDP §9 (sensitive personal data).

STATUS Public draft. Aimed at hospital systems, EHR vendors, telemedicine platforms, clinical AI vendors, medical device manufacturers. Comments to [email protected].

1. Purpose

Healthcare AI presents specific accountability challenges: patient safety, sensitive-data category restrictions, audit-by-regulator (FDA, EMA, CDSCO, MHRA), and informed-consent traceability. This profile extends R+2 to make all of these demonstrable rather than asserted.

2. Extension fields

Extension key	Type	Description
health.patient_id_hash	string	SHA-256 of the patient identifier salted per-deployment. Never the raw MRN. Required when patient data is involved.
health.consent_artifact_uri	string	URI to the consent artifact under which the action was authorized. Typically `fhir://Consent/{id}` or equivalent.
health.clinical_role	string	The clinical role for whom the agent was acting (e.g., `physician`, `nurse`, `radiologist`, `patient-self-service`).
health.decision_class	string	One of: `diagnostic-assist`, `treatment-recommendation`, `triage`, `documentation`, `scheduling`, `billing`. Determines downstream regulatory scrutiny.
health.human_in_loop	boolean	True if a licensed clinician reviewed and authorized the action before it had effect. Required for diagnostic-assist and treatment-recommendation classes.
health.model_version	string	Version identifier of the model/policy used. Enables retrospective audit if a model is later recalled.
health.training_data_attestation	string	Hash linking to a published statement of the model's training data sources. Supports regulator inquiries into bias and provenance.
health.jurisdiction	string	Jurisdiction governing the action (e.g., `US-CA`, `EU-DE`, `IN-MH`). Determines applicable regulatory framework.

3. action_type namespace

Healthcare actions use the health/* namespace:

health/diagnostic/* — AI-assisted diagnostic actions
health/triage/* — symptom assessment and routing
health/treatment/* — treatment recommendation, dosing
health/documentation/* — clinical note generation, summarization
health/scheduling/* — appointment booking, follow-up
health/billing/* — coding assistance, claims processing
health/research/* — de-identified research operations

4. Compliance mapping

Regulation	R+2 field satisfying it	Notes
HIPAA §164.312(b) Audit controls	Full receipt chain + R+3 export	Tamper-evident audit trail across all PHI processing.
HIPAA §164.502(b) Minimum necessary	extensions.health.decision_class	Class restricts the scope of PHI accessed for each action type.
HIPAA §164.508 Authorization	extensions.health.consent_artifact_uri	Every action ties to a specific authorization or consent.
GDPR Art. 9(2)(a) Explicit consent	extensions.health.consent_artifact_uri	Sensitive category processing requires linkable consent.
GDPR Art. 22 Automated decision-making	extensions.health.human_in_loop	Demonstrates human oversight for significant clinical decisions.
DPDP §9 Children's data + sensitive data	(combined with r2-gov-v1.gov.in.minor_data)	Healthcare receipts targeting Indian minors should also use the gov-in profile.
FDA Software as a Medical Device	extensions.health.model_version + .training_data_attestation	Audit-ready record of which model produced which output.
EU AI Act Annex III (high-risk)	Full receipt chain + R+3 logging	Healthcare AI is typically high-risk; full logging is mandatory.

5. Air-gapped operation

Many hospital deployments require on-premise operation with no outbound connectivity. Conforming implementations MUST support:

Local Postgres + pgvector storage of all receipts.
Local key custody (HSM-backed for production, software-backed for development).
Local embedding generation (fallback to local sentence-transformers when OpenAI is not permitted).
Optional federation with a network-connected deployment for non-sensitive operations.

6. Special considerations for healthcare

De-identification: When the action operates on de-identified data, set extensions.health.patient_id_hash to the literal string "de-identified" rather than omitting it. This preserves auditability of the de-identification claim.

Emergency override: Some clinical workflows require break-glass access to PHI. The action_type health/emergency-override/* documents these events with full justification in action_data. All emergency overrides flow through the same R+2 chain and become permanent records.

Pediatric data: When the patient is a minor, set extensions.health.patient_id_hash with an additional "minor:" prefix in the salt to distinguish in downstream audits. Combine with the r2-gov-v1 profile if operating in India.

7. Adoption pathway

Begin with documentation-class actions (lowest regulatory risk) — clinical note summarization, scheduling assistance.
Expand to triage-class with human-in-loop.
Add diagnostic-assist with full clinical validation studies.
Only consider treatment-recommendation class after regulatory clearance pathway is established.

DCS Labs offers an "r2-health Implementation Partner" program for hospital systems considering adoption — pre-MSA technical consultation at no cost. Contact [email protected].

Contact

Editorial: [email protected]
Healthcare partnerships: same address, prefix subject with "[Health]"