Friday, November 20, 2026

Patients can be given control: Consent About AI

Patients can have different preferences about the use of their health information, especially that use of their data by AI. A patient might permit clinical decision support, deny model training, or allow training only when data is de-identified. Recording a preference is only the first step: the organization also needs a technical way to represent the decision and apply it when data is requested.

The Consent About AI Implementation Guide (IG) explores that step using FHIR Consent. It treats the Consent as the healthcare organization's actionable record of the patient's decision: it contains rules an access-control system can evaluate and enforce. The IG is available based on FHIR R4 and FHIR R6. The signed consent paperwork remains the legal record and is referenced from the FHIR Consent.

The IG supports two complementary ways to express AI-related rules. A provision can identify a purpose of use, such as MLTRAINING for AI training, TREATDS for clinical decision support, or PMTDS for payment decision support. This approach is independent of a particular vendor or model, but it depends on the system requesting data accurately declaring its purpose. The IG makes that an explicit expectation for both a Data Requester (the AI system accessing data directly) and an AI Agent (a gateway accessing data on the AI's behalf).

Alternatively, a provision can identify a specific AI system or model through Consent.provision.actor.reference, using a FHIR Device that conforms to the AI Device profile. That gives the consent a more specific target, but also makes it more sensitive to change: a new model or version may need a new Device and a corresponding consent rule, and data access must be attributed to that Device.

The guide formalizes these patterns in a Consent about AI profile. It also shows how a permit can carry a limit, such as an obligation to de-identify data, through the standard consent-provision-limit extension. A limit is useful only when it can actually be enforced. The IG assigns that responsibility to the Access Control actor, which must evaluate the declared purpose, any identified AI Device, and any limits on the permit.

The examples cover allowing and denying AI for training, clinical decision support, and payment decision support; allowing a specific AI for a purpose; and allowing training on de-identified data. They also show a combined Consent containing AI rules alongside ordinary Treatment, Payment, and Operations rules.

That combined example highlights an important FHIR Consent modeling constraint: Consent has provisions, and each nested provision is an exception to the rule above it. Nested provisions therefore alternate between permit and deny, which can require deeper nesting than a set of independent rules might suggest.

Consent About AI does not introduce a new consent resource. It profiles the existing FHIR Consent, provides examples for common patient choices, and defines requirements for the systems that author, request, and enforce those choices. The guide is experimental and intended to invite discussion and feedback. Explore the profile, examples, actors, and conformance requirements here:

- GitHub source repository - https://github.com/JohnMoehrke/ConsentAboutAI
- Current FHIR R4 CI build
- Current FHIR R6 CI build

Note that the CI build reflects the current development state. This has been a 100% volunteer project by me. I would very much like to be sponsored to take this through the HL7 process. See my Consulting business, Moehrke Research LLC if you would like to sponsor this project.

No comments:

Post a Comment