# ADR-005 — NCIt as the anchor ontology

Status: accepted (2026-09-08)

# Context

CancerIndex must cover every recognised malignant disease entity, not a list of common cancers, and must reconcile labels from registries (ICD-10/ICD-O site groups), clinical sequencing (OncoTree), curated evidence (CIViC uses DOID), trials (free text), literature (MeSH) and cohorts (GDC project/disease types). No single vocabulary covers all of these; a canonical spine is needed to which the others map.

Candidates: NCI Thesaurus (NCIt), Disease Ontology (DOID), MONDO, OncoTree, ICD-O-3, MeSH.

# Decision

The NCI Thesaurus is the anchor: canonical cancer entities are created from NCIt neoplasm concepts (roots in packages/ontology/src/top-level.ts: Neoplasm C3262, Malignant Neoplasm C9305, by-site C3263, by-morphology C4741, hematopoietic/lymphoid C27134, childhood malignant neoplasm C4005…), cancers.primary_ncit_code is unique, and the ncit hierarchy is the default tree for descendant aggregation. Reasons: NCIt is maintained monthly by the NCI, is the terminology of ClinicalTrials.gov condition coding and of CTRP/caDSR, carries rich synonyms with their source, includes cross-references to UMLS/MeSH/ICD-O/DOID (via EVS), is public domain, and its EVS REST API is open.

Other vocabularies are kept as additional dimensions and codes, never discarded:

OncoTree provides the tissue-based tree (hierarchy_type = oncotree) and cancer_codes.system = oncotree, mapped to NCIt through OncoTree's own NCI cross-references (EXACT_IDENTIFIER).
DOID (CIViC), MeSH (PubMed), ICD-10 / ICD-O (registries), UMLS CUIs, MONDO, GDC project codes are stored in cancer_codes with a match_type.
Registry site groups used for burden rankings are a curated overlay (TOP_LEVEL_CANCERS, 36 ICD-10 groups anchored to NCIt concepts), not taxonomy nodes.

Qualified disease states in NCIt (stage, recurrent, metastatic, laterality, treatment history) are recognised by deterministic rules (packages/ontology/src/qualifiers.ts) and kept as source records attached to their base disease rather than as canonical entities (spec §217-218).

# Consequences

One concept, one row, one CI-CAN id; aliases from every vocabulary hang off it, which is what makes deterministic reconciliation (CancerResolver) possible.
NCIt's own updates (retired concepts, merges) flow into status = deprecated|merged and entity_merges, with classification_version recorded on entities.
Where NCIt and another ontology disagree on granularity, the mapping is recorded as CURATED_BROADER / CURATED_NARROWER rather than forced to exact, and the difference is visible in the API.
The choice can be revisited if MONDO's harmonised cross-walks prove more complete for rare cancers; the model already stores MONDO codes, so the change would be a re-anchoring migration, not a schema change.