The pipeline
Guideline to deployable rule, in four steps
We split the authoring pipeline into stages where AI accelerates pattern-matching and clinicians govern the decisions. Every stage has a clear handoff and a clear owner.
01
Extract intent
LLM parses the published guideline, registry spec, or measure into structured intent — populations, criteria, exceptions — with citations back to the source paragraph.
LLM-led
02
Clinical review
A clinician validates the structured intent against the source guideline, resolves ambiguity, and signs off on what the rule should and shouldn't fire on.
Clinician-led
03
Draft CQL & valuesets
LLM drafts CQL logic and proposes valuesets from VSAC and authoritative sources. Output is structured for diff review, not free-form code.
LLM-led
04
Informatics QA
Informaticists test against synthetic and real-world data, tune valuesets, and validate that the deployed rule behaves as the clinician specified.
Informaticist-led
How we approach it
Principles we don't compromise on
AI-accelerated authoring is only useful if the result is at least as safe and maintainable as hand-authored CDS. These are non-negotiable.
01
Clinicians stay in the loop
No rule reaches production without explicit clinical review of its intent and behavior. The model accelerates the pipeline; it doesn't sign off on logic.
02
Every claim cites its source
When the LLM extracts intent or proposes a valueset, it includes the paragraph or code system it pulled from. Reviewers see the evidence, not just the conclusion.
03
Outputs are testable artifacts
CQL, FHIR resources, structured valuesets — everything the model produces is something we can run a test suite against. No prose-only deliverables.
04
Lifecycle, not one-shot
Guidelines change. Valuesets change. The same pipeline is built to re-run when sources update — with diff review, not a from-scratch rewrite each time.