clinicaltrials-extraction

Clinical Trials Extraction Skill

NCT ID Capture

For any Trial entity, the name field MUST be the NCT identifier (format: NCT followed by 8 digits, e.g. NCT04303780) when one is present in the document. Study acronyms ("CodeBreaK 200"), sponsor-assigned codes ("AMG 510 study"), and descriptive titles ("Phase 3 study of sotorasib in NSCLC") go into attributes.trial_acronym or attributes.descriptive_title — not into name.

If no NCT ID is present, fall back to the study acronym. Only use a descriptive title as name as a last resort, and set confidence <= 0.6.

This rule exists because the post-build enrichment step resolves Trial nodes against ClinicalTrials.gov via their NCT ID. Nodes without an NCT ID cannot be enriched.

Output Format

Return a DocumentExtraction JSON envelope. The document_id field is REQUIRED. All entities and relations are lists of objects.

{
  "document_id": "nct04303780_protocol",
  "entities": [
    {
      "name": "NCT04303780",
      "entity_type": "Trial",
      "confidence": 0.95,
      "context": "Phase 3 CodeBreaK 200 trial evaluating sotorasib vs docetaxel",
      "attributes": {
        "trial_acronym": "CodeBreaK 200",
        "study_type": "Interventional"
      }
    },
    {
      "name": "sotorasib",
      "entity_type": "Compound",
      "confidence": 0.95,
      "context": "AMG 510; KRAS G12C inhibitor evaluated as experimental arm",
      "attributes": {
        "inn": "sotorasib",
        "sponsor_code": "AMG 510"
      }
    }
  ],
  "relations": [
    {
      "source_entity": "NCT04303780",
      "target_entity": "sotorasib",
      "relation_type": "tests",
      "confidence": 0.95,
      "evidence": "Trial NCT04303780 evaluates sotorasib in the experimental arm"
    }
  ]
}

Field Requirements

Entity fields

• name (required): Canonical identifier. For Trial, use NCT ID. For Compound, use INN. For Condition, use MeSH term.
• entity_type (required): Must match one of the 12 types in domain.yaml exactly (case-sensitive).
• confidence (required): Float 0.0–1.0. See Confidence Calibration section.
• context (required): Short excerpt or paraphrase from the source document justifying extraction.
• attributes (optional): Dict of domain-specific metadata (e.g. trial_acronym, sponsor_code, phase, dose).