RareNeoplasmRwdLlmExtraction Agent

<!--
# COPYRIGHT NOTICE
# This file is part of the "Universal AI Agentic Skills" project.
# Copyright (c) 2026 MD BABU MIA, PhD <md.babu.mia@mssm.edu>
# All Rights Reserved.
#
# This code is proprietary and confidential.
# Unauthorized copying of this file, via any medium is strictly prohibited.
#
# Provenance: Authenticated by MD BABU MIA
-->



<!-- AUTHOR_SIGNATURE: 9a7f3c2e-MD-BABU-MIA-2026-MSSM-SECURE -->

---
name: 'rare-neoplasm-rwd-llm-extraction'
description: 'Extract registry-ready rare-neoplasm variables from clinical text using schema-first LLM workflows with temporal normalization, ontology mapping, provenance, adjudication, privacy, and validation.'
measurable_outcome: 'Execute skill workflow successfully with valid output within 15 minutes.'
allowed-tools:
  - read_file
  - run_shell_command
  - web_fetch
---

# Rare Neoplasm Real-World Data LLM Extraction

## Overview

Extract structured real-world data from oncology notes, pathology reports, operative records, imaging reports, and treatment summaries for rare neoplasms. Use a schema-first workflow to handle sparse documentation and terminology variation while preserving evidence provenance, uncertainty, privacy, and a clear route to human review.

## When to Use This Skill

- Build or update a rare-cancer registry from unstructured clinical documents.
- Extract bone sarcoma or other rare-neoplasm diagnoses, pathology, staging, treatment, response, recurrence, and outcomes.
- Harmonize institution-specific tumor terminology with oncology ontologies.
- Reconstruct longitudinal cancer timelines from fragmented records.
- Audit an LLM extraction pipeline against manually curated registry data.
- Produce review queues for ambiguous, conflicting, or low-confidence variables.
- Do not use the workflow as a substitute for clinical diagnosis, treatment decisions, or source-document verification.

## Core Capabilities

1. **Define the extraction contract.** Specify the cohort, tumor family, source types, index event, field definitions, cardinality, allowed values, units, and required evidence before processing records. Version the schema and ontology releases so outputs remain reproducible.

2. **Apply privacy controls.** Process only authorized records in an approved environment, enforce least-privilege access, minimize retained identifiers, avoid placing protected health information in prompts or logs beyond operational need, and follow applicable institutional and legal requirements. De-identify or pseudonymize exports unless identifiable data are explicitly required and authorized.

3. **Prepare source documents.** Preserve document identifiers, document type, authoring service, encounter date, specimen date, and original text offsets. Remove duplicate documents, identify amended reports, and retain report versions so later corrections can supersede earlier findings without erasing provenance.

4. **Extract schema-bound facts.** Return only requested fields in a machine-valid structure such as JSON or a registry table. For each assertion, capture the raw value, normalized value, status, source document, evidence span, character or line offsets when available, and extraction timestamp.

5. **Link every value to evidence.** Use the shortest sufficient verbatim span supporting each extracted fact. Distinguish direct statements from inferred values; do not present an inference as explicitly documented. If a derived field is required, record the derivation rule and all supporting source fields.

6. **Normalize time and clinical episodes.** Convert dates to ISO 8601 when the source supports it, retain the original expression, and represent partial dates without inventing missing precision. Anchor relative expressions to a documented reference date, associate events with the correct tumor episode, and flag impossible or conflicting sequences.

7. **Map rare-tumor terminology.** Preserve the original disease wording and map it to the project-approved ontology, such as ICD-O, NCI Thesaurus, or SNOMED CT, including code, label, ontology version, and mapping confidence. Use morphology, primary site, molecular findings, and pathology context; abstain rather than force an unsupported code.

8. **Resolve negation, uncertainty, and context.** Separate confirmed findings from suspected, ruled-out, historical, family-history, planned, canceled, and hypothetical statements. Track whether a statement applies to the patient, a specific specimen, a metastatic site, or another tumor episode.

9. **Represent missingness explicitly.** Distinguish `not_documented`, `not_assessed`, `not_applicable`, `unknown`, `redacted`, `conflicting`, and `extraction_failed`. Never convert absent text into a negative clinical finding or silently impute a value.

10. **Calibrate confidence and abstention.** Assign field-level confidence using documented evidence quality, terminology fit, temporal consistency, and cross-document agreement. Configure thresholds by field risk: auto-accept high-confidence facts, route intermediate-confidence facts to review, and abstain on unsupported or contradictory facts.

11. **Reconcile longitudinal conflicts.** Prefer final pathology over preliminary interpretations when appropriate, amended reports over superseded versions, and more specific evidence over generic summaries. Preserve all conflicting assertions and record the deterministic precedence rule or human decision used to select the registry value.

12. **Require human adjudication where needed.** Create a review queue containing the candidate value, alternatives, evidence spans, source metadata, conflict reason, and proposed resolution. Require qualified review for low-confidence mappings, discordant pathology, uncertain primary site, duplicate tumor episodes, and safety-critical fields.

13. **Validate against curated data.** Compare outputs with an independently manually curated registry using a locked schema and adjudication guide. Report field-level completeness, exact or clinically meaningful agreement, precision, recall, error categories, abstention rate, and review burden as applicable; stratify by document type and tumor subtype without inventing performance claims.

14. **Use the bone sarcoma rare-neoplasm example as a validation case.** Validate schema-first variable extraction for rare-neoplasm real-world data collection with a bone sarcoma scenario that exercises sarcoma-specific schemas, temporality and event ordering, ontology normalization, evidence provenance capture, registry readiness, adjudication workflow routing, and field-level metrics for rare-neoplasm chart abstraction, aligned to the 2026 PubMed-indexed example without inventing unsupported benchmarks or claims.

15. **Use the bone sarcoma RWD extraction template.** For sparse rare-neoplasm records, define registry-ready variable schemas for diagnosis, pathology, staging, treatment, response, recurrence, outcome events, and missingness; normalize absolute, partial, and relative dates; map disease, treatment, site, and event terms to approved ontologies; capture provenance with source documents and evidence spans; route ambiguous or conflicting variables to adjudication queues; and validate outputs against manually curated records.

16. **Apply the PubMed 42021926 bone sarcoma RWD extraction pattern.** Treat the 2026 bone sarcoma rare-neoplasm paper as a concrete pattern for schema-first variable capture, temporal normalization, provenance-linked evidence extraction, ontology mapping, validation against manual abstraction, and registry-readiness checks without adding unsupported performance claims.

17. **Apply the bone sarcoma sparse-record validation pattern.** Use the bone sarcoma rare-neoplasm RWD example to validate schema-first abstraction, temporal normalization, ontology mapping, provenance capture, clinician adjudication, and error analysis for sparse oncology records without adding unsupported benchmarks or claims.

18. **Validate registry-field governance with the bone sarcoma example.** Use the 2026 bone sarcoma rare-neoplasm RWD example as a validation pattern for schema-first variable capture, temporal normalization, oncology ontology mapping, audit-trail completeness, and clinician adjudication of LLM-extracted registry fields.

19. **Operationalize the bone sarcoma fragmented-record workflow.** Use the 2026 bone sarcoma rare-neoplasm RWD workflow as a worked pattern for LLM-assisted variable extraction from fragmented oncology records: normalize dates and event order without inventing precision, map tumor and treatment terms to approved ontologies, attach registry-ready provenance to every value, route uncertain or discordant fields to adjudication, and validate registry outputs against curated reference records without adding unsupported performance claims.

20. **Use PubMed 42021926 as a registry-quality validation case.** Treat the 2026 bone sarcoma rare-neoplasm RWD example as a concrete validation case for schema-first LLM extraction, temporal normalization, oncology ontology mapping, provenance capture, manual adjudication, and registry-ready quality metrics, while limiting claims to the reported paper scope.

21. **Apply the bone sarcoma variable-definition pattern.** Use the 2026 bone sarcoma rare-neoplasm RWD example as a concrete extraction pattern: define registry variables schema-first, normalize temporality and event order, attach registry-grade provenance to each value, map diagnosis, morphology, site, treatment, response, recurrence, and outcome terms to approved oncology ontologies, route uncertain or conflicting fields to adjudication queues, and validate outputs against independent manual abstraction without adding unsupported benchmarks.

22. **Run the bone sarcoma worked use case.** Use the 2026 bone sarcoma rare-neoplasm RWD example as a worked case for schema-first variable capture, temporal normalization, oncology ontology mapping, provenance-linked extraction, privacy-safeguarded prompting and export, adjudication of ambiguous or conflicting fields, and registry-readiness validation without adding unsupported performance claims.

23. **Apply bone-sarcoma RWD extraction controls.** For bone sarcoma rare-neoplasm abstraction, define variables schema-first with registry-ready field rules, normalize temporal expressions and event order, map diagnosis, morphology, site, treatment, response, recurrence, and outcome terms to approved ontologies, capture source provenance for each value, route uncertain or conflicting fields to adjudication queues, and validate outputs against registry-ready fields without adding unsupported performance claims.

24. **Apply the bone sarcoma rare-neoplasm RWD example end to end.** Use PubMed 42021926 as a bone sarcoma example for schema-first variable capture, temporal normalization, source provenance, oncology ontology mapping, adjudication queues, privacy controls, and validation against manual abstraction without inventing unsupported performance claims.

25. **Strengthen fragmented-record extraction with the bone sarcoma example.** Use PubMed 42021926 to tighten rare-neoplasm RWD schemas around registry-ready variable definitions: normalize tumor entity, morphology, site, and treatment terms to approved ontologies; anchor treatment and outcome events to explicit source dates or documented relative references; preserve source-document and evidence-span provenance for fragmented clinical records; and route ambiguous, conflicting, or unsupported values through adjudication before registry export.

26. **Maintain an audit trail.** Record model and prompt versions, schema and ontology versions, preprocessing steps, source hashes or stable identifiers, extraction status, reviewer actions, and final disposition. Make each final value traceable to source evidence and any transformation or adjudication.

27. **Run release checks.** Reject outputs that fail schema validation, contain unsupported values, lack required provenance, expose unauthorized identifiers, or violate chronology constraints. Sample accepted, rejected, and abstained cases for periodic human quality review and monitor drift when data sources or models change.

28. **Apply the bone sarcoma validation pattern.** Use the 2026 PubMed-indexed bone sarcoma rare-neoplasm RWD extraction example as a validation pattern for schema-first variable definition, temporal normalization, oncology ontology mapping, source provenance capture, privacy review, and human adjudication before registry-ready output export, without adding unsupported benchmarks or claims.

29. **Extract bone sarcoma registry facts from clinician notes.** Use PubMed 42021926 as a concrete rare-neoplasm RWD pattern for schema-first extraction from clinician notes: normalize temporal events from documented dates or relative anchors, map disease and care terms to approved oncology ontologies, send ambiguous or conflicting fields to adjudication queues, and require registry-quality provenance checks linking every accepted value to source metadata and evidence spans.

30. **Apply the 2026 bone sarcoma RWD finding.** Use PubMed 42021926 as a rare-neoplasm real-world data collection example for schema-first variable extraction, temporal normalization, oncology ontology mapping, source-document and evidence-span provenance, adjudication workflow routing, privacy-controlled prompting and export, and registry-quality validation metrics without inventing unsupported benchmarks or claims.

31. **Harden registry-ready extraction with PubMed 42021926.** Use the bone sarcoma rare-neoplasm paper to require schema-first registry variables, normalize absolute, partial, and relative temporal expressions, map extracted disease and care terms to approved oncology ontologies, link every accepted value back to source notes and evidence spans, route ambiguous or conflicting outputs through adjudication, enforce privacy controls for prompting and export, and validate outputs against registry-ready variables without adding unsupported performance claims.

32. **Convert fragmented oncology records into registry-ready variables.** Use the PubMed 42021926 bone sarcoma example to guide schema-first extraction across notes, pathology, imaging, operative records, and treatment summaries: normalize treatment and outcome events from documented dates or relative anchors, map diagnosis, morphology, site, treatment, response, recurrence, and outcome terms to approved ontologies, retain source identifiers and evidence spans for every value, route ambiguous, conflicting, or unsupported fields to adjudication queues, and report registry-ready validation metrics only when measured against curated reference records.

33. **Validate bone sarcoma RWD collection controls.** Use PubMed 42021926 as a concrete bone sarcoma rare-neoplasm validation case for schema-first extraction, temporal normalization, registry variable mapping, missingness reporting, adjudication queues, and privacy-preserving provenance through source identifiers, evidence spans, and de-identified or pseudonymized exports without adding unsupported performance claims.

34. **Apply the bone sarcoma manual-abstraction validation example.** Use PubMed 42021926 as a concrete rare-neoplasm RWD example that starts with schema-first variable capture, normalizes temporal expressions and event order, maps tumor and care terms to approved ontologies, links every extracted value to source provenance, routes uncertain fields to clinician adjudication, and validates outputs against manually abstracted registry fields without adding unsupported benchmarks or claims.

35. **Use the bone sarcoma registry-readiness example.** Apply PubMed 42021926 as a rare-neoplasm real-world data extraction example for schema-first variable capture, temporal normalization, approved oncology ontology mapping, provenance fields for source documents and evidence spans, adjudication workflow routing for ambiguous or conflicting values, and registry-readiness checks before export without adding unsupported benchmarks or claims.

## Inputs / Outputs

### Inputs

- Authorized clinical text or document references, with stable document and patient or case identifiers.
- A versioned rare-neoplasm extraction schema and data dictionary.
- Cohort definition, index-date rules, tumor-episode rules, and document precedence rules.
- Approved ontology names, versions, code sets, and local terminology mappings.
- Field-specific confidence thresholds and human-review criteria.
- Privacy, retention, access-control, and export requirements.
- Optional manually curated registry records for validation.

### Outputs

- Schema-valid patient-, tumor-, specimen-, and event-level records.
- For each extracted field: raw text value, normalized value, evidence span, source identifier, temporal anchor, ontology mapping, confidence, and assertion status.
- Explicit missingness, abstention, conflict, and extraction-failure states.
- A chronological tumor and treatment event timeline with unresolved inconsistencies flagged.
- A human-adjudication queue with evidence and resolution history.
- Validation and error-analysis summaries against the curated reference set.
- A reproducibility and audit manifest containing model, prompt, schema, ontology, preprocessing, and review versions.

## References

- Teterycz P, Rynkun S, Szostakowski B, Wągrodzki M, Rutkowski P. “Accelerating real-world data collection using large language models in rare neoplasms: a bone sarcoma example.” *ESMO Real World Data and Digital Oncology*. 2026 Jun. PubMed: https://pubmed.ncbi.nlm.nih.gov/42021926/