NLP

Clinical text is everywhere and structure isn't. Legacy EHR exports carry CodeableConcepts with a free-text text and no coding[] — CQL retrieves against [Code: <valueset>] can't match them. Progress notes hold the richest clinical detail in the record, as prose. This page runs the real NLP machinery against both: the AutoCoder turns text-only concepts into coded Codings with a full audit trail, and the annotations engine reads clinical notes by meaning and derives structured Condition resources from them.

Both run the real thing: fhir4ds-v2 executes in your browser (Pyodide), the AutoCoder is the shipped implementation, and the terminology work — hybrid search (lexical BM25 + semantic SapBERT, fused, with match grades certain / probable / possible / broader) and the NER cascade — is served live by the public medterm4ds terminology Space.

How it works

1. Detection — the AutoCoder walks the resource for CodeableConcepts that have a text and no existing coding[] (never re-coding manual codes).
2. Search — each text goes to CodeSystem/$search on the terminology service (hybrid mode, system chosen by clinical category).
3. Filter — results below the configured minimum match grade are dropped; the top k survive. Re-filtering is instant — no new search.
4. Augmentation — each surviving result becomes a Coding with userSelected: false and the autocoding audit extension: engine, engine version, search mode, score, match grade.

The same flow in Python is three lines — see the Auto-Coding guide:

from fhir4ds.cql.loader import AutoCoder

coder = AutoCoder(endpoint, con)          # any terminology endpoint
coder.augment_resource(resource)          # in place, audited

Try it

Edit the resource (any code.text works — try a colloquial phrase and watch semantic search find the clinical concept), set the policy, and run.

Good to know

The terminology Space sleeps when idle — the first search after a wake can take up to ~2 minutes (the demo tells you when this happens; retry shortly after). The demo allows 30 requests/minute per IP; for anything heavier, self-host the terminology service. Nothing you type leaves your browser except the search query itself.

Auto-Coding

Loading demo…

Annotations

The same terminology service also reads free text. The example below runs the NER cascade (CodeSystem/$extract) on a clinical note — live, against the public Space — and wraps every affirmed concept as a derived FHIR Condition resource: coded, audit-trailed, and queryable.

Loading demo…

Where to go next

• Auto-Coding guide — the Python API, batch mode, caching, and config reference.
• Terminology service — endpoints: in-process medterm4ds or any FHIR R4 sidecar.
• Notes pipeline — free-text notes to derived Condition resources (NER cascade).
• Terminology search demo — the terminology operations themselves, live.