4nt11/avk-datasets

AVK Spec Lexicons

Domain vocabularies harvested from healthcare standards' own published definitions, for resolving terms that general lexicons (WordNet, UMLS) don't contain. Built for AVK, an ambiguity-triage tool that flags where standards prose is ambiguous — but usable standalone as a term → concept resource.

The idea: DiagnosticReport, CodeableConcept, PixelData, StudyInstanceUID are defined precisely by FHIR and DICOM themselves, yet are absent from UMLS (they're HL7/DICOM model classes and data elements, not clinical findings). A standard is its own best lexicon — so these are harvested straight from what each standard publishes.

Subsets

config	terms	source
fhir_r5	5,095	FHIR R5 definitions.json — StructureDefinitions (resource + datatype names) and CodeSystems
dicom	23,316	DICOM PS3.6 via pydicom (data-element keywords + UID/SOP registry) + PS3.16 Code Meanings

Schema

One JSON object per line:

field	meaning
norm	normalized lookup key (camelCase split, lowercased) — DiagnosticReport → diagnostic report
term	the term as published
concept_id	stable id (FHIR StructureDefinition URL, DICOM-DE:<keyword>, DICOM-UID:<uid>, or <scheme>:<code>)
kind	resource / complex-type / primitive-type / attribute / uid / code
source	FHIR-R5 or DICOM
definition	short description where the source provides one

Two terms sharing a concept_id are the same concept (synonyms); distinct concept_ids are distinct concepts.

Usage

from datasets import load_dataset

fhir = load_dataset("4nt11/avk-datasets", "fhir_r5", split="train")
dcm  = load_dataset("4nt11/avk-datasets", "dicom",   split="train")

Each subset also ships as a ready-to-query SQLite index (*_lexicon.sqlite, table terms, indexed on norm) — what AVK's spec oracle backend reads directly.

Notes & caveats

• kind='code' is lower-precision. FHIR CodeSystem displays and DICOM Code Meanings include generic words (the, request) that collide with ordinary prose. AVK excludes code from its similarity scoring by default and trusts only the named concepts; they're kept here for completeness.
• DICOM Code Meanings are best-effort. Extracted from PS3.16 (part16.html) tables, which use rowspans — column alignment is approximate, so some code rows may be imperfect. The 6,336 attribute + uid terms (from pydicom) are authoritative and parse-risk-free.

Provenance & licensing

This is a derived vocabulary index, and licensing applies in two layers:

• The database (this compilation — the selection, normalization, and concept_id scheme) is released under the Open Database License (ODbL) v1.0.
• The contents (the individual terms) retain their source licenses:
  • FHIR R5 © HL7, published under CC0 — freely redistributable.
  • DICOM © NEMA, reproduced under the DICOM Standard's permissive reproduction terms; pydicom (MIT) supplies the PS3.6 data.

ODbL governs the compilation only; it does not (and cannot) relicense the underlying HL7/NEMA content. Consult the primary publications (hl7.org/fhir, dicomstandard.org) for authoritative, current definitions. No UMLS-licensed content is included, by design — that's what keeps this shareable.