# A sample of what can be done

This sample illustrates clinical knowledge extracted from five DailyMed Structured
Product Labels (SPLs), with source-scope enrichment and RxNorm / SNOMED CT
terminology annotations. An extensive proprietary automated extraction pipeline
uses an LLM to build structured, source-grounded knowledge.

The broader goal is a rich, connected clinical knowledge graph across medicine,
built and improved at scale. Medication–disease and symptom relationships are
our starting point. This sample demonstrates representation and retrieval,
not comprehensive clinical coverage. Clinical validation is ongoing.

## Contents

- `sample.json`: UTF-8 JSON containing clinical assertions, product context,
  source XML and retained exact/related terminology annotations.
- `manifest.json`: SHA-256 checksums and counts for verification.
- `README.md`: this guide.

The ZIP uses LZMA compression for a smaller download. Use an extractor that
supports LZMA ZIP files. The JSON is supplied inside the archive.
Terminology annotations use the UMLS 2026AA snapshot.

## Reading the JSON

`results` contains one record per label: levothyroxine, metformin, lisinopril,
atorvastatin and amoxicillin. Original ingredient wording is preserved.
`label_id` is a local record key; `set_id` and `spl_version` identify the SPL.

Each record contains:
- `subject`: original ingredient name(s).
- `contexts`: route, dosage form, formulation and their source evidence.
- `assertions`: stable assertion ID, relationship, complete object wording,
  evidence phrases, source/section references and applicable context IDs.
- `sources`: source XML keyed by the IDs used in assertions/context evidence.
- `terminology.concepts`: concept ID, name and vocabulary namespace.
- `terminology.mappings`: selected code and label, exact/related match type,
  explanation and `umls_cuis`, the list of CUIs for that source/code/label.
- `terminology.links`: which concepts annotate each target, with association notes.
- `terminology.target_ids`: resolves target IDs to a subject (optionally with
  context IDs) or to an assertion ID.

To find an assertion's codes, use its `assertion_id` to locate its target in
`target_ids`, find that target in `links`, and join the listed `concept_ids` to
`concepts` and `mappings`. Subject targets provide ingredient/formulation codes.
Concept, target, context and source-unit IDs are local to each label record.
An ID in one namespace is not an ID in another. Codes are strings, not numbers.

## Interpretation

- An exact code matches the annotated concept, not necessarily the whole qualified
  assertion. Related mappings are not equivalent meanings.
- Component annotations are independent links, not formal SNOMED CT expressions.
- Original qualifiers, alternatives, evidence and scope notes remain attached
  to each assertion. Product source scope is not proof of route exclusivity.
- The extraction can contain incomplete, contradictory or incorrect clinical claims.
  Clinical validation is ongoing.
- Source snapshots may differ from current labels. This is a sample, not a
  clinical decision tool by itself.

Clinical assertions, context and evidence are preserved without clinical rewriting.
Internal work/result bookkeeping is omitted.

## Availability

A future scaled version of the knowledge graph will be published, free for
non-commercial use. Applicable source and terminology licenses remain separate;
this sample does not grant rights to third-party vocabularies. Terms for public
redistribution remain under review.