ADLC for Deterministic Clinical Content: Guidelines, Assessments, and Measures

The HL7 Connectathon is September 19–20 in Rockville, Maryland. That makes it a useful place to test a simple idea: if an agent helps turn clinical source material into computable knowledge, the output has to be more than plausible.

ADLC applies that idea to deterministic clinical content. Guidelines, assessments, and measures have to be inspectable. They have to run against a known input. And another implementation has to be able to reach the same result.

The proposed AI-Assisted Development of Clinical Reasoning Knowledge track starts with a bounded older-adult fall-risk screening use case. Participants receive the same source corpus, a shared FHIR R4 Questionnaire, synthetic patient bundles, and machine-readable assertions. They can use an agentic process, a conventional process, or both. The method is intentionally not the thing being standardized.

The contract is the result. A submission should preserve source traceability, produce valid FHIR and executable CQL where applicable, and disclose its tools, versions, instructions, human decisions, and unresolved issues. It should also show whether its artifacts can be stored, retrieved, translated, and evaluated by independent implementations.

That is a better test than asking whether a model can generate a convincing PlanDefinition. Generation is cheap. Trust is still expensive.

A fixture is a clinical-reasoning boundary

The shared fixtures make the expected behavior concrete. The example below is illustrative and synthetic; it reflects the proposed track’s input/output contract, not a clinical recommendation or a production decision service.

1Input
2  Patient: age 81, ambulatory encounter
3  QuestionnaireResponse: completed
4  Unsteady when walking: false
5  Worried about falling: false
6  Fell in the past year: true
7
8Expected output
9  In Screening Population: true
10  At Increased Fall Risk: true
11  SDC extraction: 3 final Observations, each derived from the response

The neighboring edge case matters at least as much. A response that is in-progress and missing a required answer must not quietly become a negative screen. Its expected risk result is unknown; baseline extraction is not invoked, and no final Observations are emitted. The distinction between false, unknown, and an error is part of the contract.

This is what we mean by testable clinical reasoning. The source text, structured artifact, executable logic, fixture input, expected result, and human review all remain connected. FHIR validation and CQL compilation are necessary checks. They are not clinical approval, and fixture agreement does not prove clinical correctness.

The useful Connectathon question is therefore not “which model won?” It is whether different teams can make their process and artifacts legible enough to compare, review, reproduce, and improve.

Where we put this to work

The same boundary shows up in our tooling. The proposed Connectathon track remains implementation-neutral; this section describes where we apply the same discipline in our own tooling. rh-skills is an open-source workflow toolset for turning clinical evidence into structured artifacts and FHIR/CQL outputs, with deterministic CLI work and auditable agent prompts kept separate. ReasonHealth Hub is the hosted side of that pattern: terminology search, validation, and expansion, plus FHIR artifact lifecycle support.

They are useful for Connectathon-style work for the same reason the shared fixtures are useful: they keep source material, intermediate artifacts, executable logic, terminology checks, and human review connected. They do not turn generated output into clinical approval, but they make that output easier to inspect and re-check.

Related: Oncology Medication Prior Authorization at the HL7 Connectathon applies the same discipline to a payer-facing workflow, but keeps the proposal and its clinical boundaries explicit.

Work with us

If you are developing ADLC-style clinical content, ReasonHealth Hub provides terminology search, validation, expansion, and artifact lifecycle support, while rh-skills provides the open-source authoring workflow. For consulting help designing fixtures, FHIR/CQL outputs, and review gates, contact us.