AI Transparency on FHIR, published by HL7 International / Electronic Health Records. This guide is not an authorized publication; it is the continuous build for version 0.1.0 built by the FHIR (HL7® FHIR® Standard) CI Build. This version is based on the current content of https://github.com/HL7/aitransparency-ig/ and changes regularly. See the Directory of published versions

Use Cases

Page standards status: Informative

Use Case 1: AI Attribution in Clinical or Administrative Documentation Review

Clinicians reviewing previous treatment plans must distinguish between content inferred by AI, decisions assisted by AI, and those made without AI involvement. This parallels standard practice of identifying the human clinician responsible for treatment plans, their role in development, and their clinical background.

Moreover, the use case can be expanded to different roles and contexts, as shown in the table below:

Data Viewing Questions by Actor

When this Actor is Viewing Data	The key questions may be…
Clinician	What is happening (to modify)? Why is it happening?
Payor	What matches prior authorization criteria?
QI	What matches the desired outcomes, or desired approach to care?
Safety Board	Multiple questions for root cause analysis
Legal	Who is responsible?

Use Case 2: AI Training on Clinical Content

When AI trains on clinical content, it should avoid training on data that was generated by AI.

Use Case 3: Discovery of output from former AI model determined to be problematic

When in the course of using a given AI model, it may be come understood that that given AI model was making poor or dangerous outputs. Therefore one needs to discover all the instances where a Patient may have been impacted by this given AI model.

All outputs from a given AI model would have traceability back to a single Device resource that describes that given model. Thus one needs to trace to all resources with a Provenance.agent referencing that Device instance, and should trace to all resources otherwise referenceing that Device instance (e.g. Observation.device, Composition.author).

Use Case 4: Discovery of output from a poorly crafted Model-Card

When in the course of using AI, it is common to reuse Model-Card -prompts- that have been introduced as "good models", but at a later time realize that this given Model-Card has errors or improper formatting so as to cause poor AI output or dangerous output. Therefore one needs to discover all the instances where a Patient may have been impacted by this given Model-Card.