Clinical NLP

---
name: 'clinical-nlp-extractor'
description: 'Extracts medical entities (Diseases, Medications, Procedures) and patient outcome mentions with temporal anchoring from unstructured clinical text using regex and simple rules (or LLM wrappers).'
measurable_outcome: Execute skill workflow successfully with valid output within 15 minutes.
allowed-tools:
  - read_file
  - run_shell_command
---

<!--
# 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

-->


# Clinical NLP Entity Extractor

The **Clinical NLP Skill** converts free-text clinical notes into structured data. It identifies key medical entities like problems/diagnoses, medications, and procedures.

## When to Use This Skill

*   When analyzing unstructured EHR notes.
*   To populate a patient's problem list or medication reconciliation.
*   To de-identify text (phi-removal) - *Basic version*.

## Core Capabilities

1.  **NER (Named Entity Recognition)**: Extracts Problems, Drugs, Procedures.
2.  **Negation Detection**: (Basic) Checks if a finding is denied ("No fever").
3.  **Structuring**: Returns JSON format compatible with FHIR/USDL.
4.  **Incidentaloma Identification**: Detect incidental imaging findings requiring follow-up across multiple anatomies in radiology reports. Park et al. (J Biomed Inform, Apr 2026, PMID 42061667) compare LLM-based vs supervised approaches — useful exemplar when scoping radiology-report NLP pipelines.
5.  **Outcome Identification with Temporal Anchoring**: Detect outcome mentions in clinician notes, normalize timing, distinguish documented events from speculation, and return evidence spans for audit.
6.  **Clinician-Note Outcome Timing Extraction**: Identify patient outcome mentions in clinician notes, normalize event timing, distinguish historical/current/future events, link extracted spans to structured variables, and validate against adjudicated chart-review labels.
7.  **Registry-Oriented Rare-Cancer RWD Extraction**: For low-prevalence cancer cohorts such as bone sarcoma, define the registry schema and field-level data dictionary before extracting and normalizing diagnoses, pathology type, tumor size, localization, grade, primary resection, treatment timelines, and outcomes; link each note-to-field value to its evidence span and source-note identifier; normalize documented dates and relative temporal expressions; distinguish absent documentation, unknown values, and explicit negative findings; route ambiguous or sampled values for human validation; monitor extraction quality with field-level and aggregate audit metrics such as precision and recall without assuming preset benchmarks; and preserve temporal anchors, missingness states, and validation status as provenance.
8.  **Oncology Consultation-Document Survival Modeling**: For prognostic modeling from initial oncology consultation documents, compare zero-shot prompting with task-specific fine-tuning before selecting an approach; use censoring-aware survival labels, leakage checks for post-baseline or outcome-revealing text, held-out calibration assessment, and clear limits that outputs are investigational decision-support signals rather than direct clinical-use recommendations.
9.  **Longitudinal Outcome-Event Identification and Timing**: Extract explicit and inferred outcome events with negation status; anchor onset and resolution as dates or intervals when timing is uncertain; separate note time from event time; reconcile, update, and deduplicate events across longitudinal notes; retain clinician-auditable evidence spans; and evaluate event-level precision and recall without assuming preset benchmarks.
10. **Rare-Cancer Longitudinal Registry Abstraction**: Extract registry fields from longitudinal rare-cancer notes using ontology-backed field schemas; resolve event chronology across note and event dates; encode explicit negatives, unknown values, and absent documentation as distinct missingness states; assign field-level confidence with supporting evidence spans; route uncertain records for clinician validation; and compare extraction results against manual abstraction without assuming preset performance benchmarks.
11. **Bone-Sarcoma Registry Extraction Pattern**: For rare-cancer registries, use bone sarcoma as a worked pattern: define the abstraction schema before extraction, link temporal events across records, normalize pathology and treatment fields, attach uncertainty flags, validate behavior for low-prevalence data, and manually audit a sample of extracted records.
12. **LLM Rare-Neoplasm RWD Extraction Workflow**: Use LLMs to structure oncology notes for rare-neoplasm real-world-data collection; for bone sarcoma, capture registry-defined disease-specific variables with source-note identifiers and evidence spans for each extracted field; preserve prompt, model, run, and reviewer provenance where available; validate extracted values against manual abstraction before registry loading; and quantify field-level extraction error so uncertain or high-error variables can be corrected or withheld from registry use.
13. **Rare-Neoplasm Registry RWD Abstraction Controls**: For bone sarcoma as a rare-neoplasm exemplar, define the cohort and registry variables before extraction; abstract clinician-note evidence into disease, treatment, outcome, and follow-up fields; build treatment and outcome timelines from dated and relative mentions; validate a documented sample of extracted records against manual review; and retain audit-ready provenance for each registry-style oncology dataset value, including source note, evidence span, extraction run, uncertainty, validation status, and missingness state.
14. **Outcome-Event Extraction with Temporal Anchoring**: For clinician-note outcome extraction, define outcome-specific event schemas before extraction; distinguish mention time from event time; encode negation, uncertainty, and speculative language; preserve source spans and temporal evidence; validate extracted events against chart review; and report precision and recall separately by outcome type.
15. **Schema-Driven Longitudinal Oncology RWD Abstraction**: Define a field-level schema before extracting variables from longitudinal oncology notes; normalize rare-disease terminology; resolve event timing across note dates, explicit dates, and relative expressions; encode explicit negatives, unknown values, and absent documentation as distinct missingness states; use targeted and random human-validation samples; measure inter-rater agreement for manual review; and retain audit-ready field provenance including the source-note identifier, evidence span, extraction run, reviewer status, and temporal evidence.
16. **Double-Reviewed Rare-Neoplasm Oncology RWD Abstraction**: Define a structured extraction schema for disease, treatment, outcome, and follow-up facts; normalize longitudinal events into a consistent timeline; retain source spans for every extracted value; distinguish missing, unknown, and explicitly negative facts; preserve conflicting values with their source provenance and route unresolved conflicts for review; and validate abstractions against double-reviewed records before registry use.
17. **Confidence-Tagged Rare-Neoplasm RWD Abstraction**: Use schema-driven extraction for rare-neoplasm records; normalize pathology terminology; resolve longitudinal event timing from note dates, explicit dates, and relative expressions; report missing, unknown, and undocumented values distinctly; attach field-level confidence tags and supporting evidence; and manually validate a documented sample before downstream use.
18. **Pre-Ingestion Rare-Neoplasm RWD Extraction**: Define an explicit field schema before extraction; normalize absolute dates, relative expressions, and note-versus-event time; retain source-span evidence for every value; distinguish explicit negatives, unknown values, and absent documentation; route ambiguous or conflicting fields for clinician adjudication; and complete field-level validation before registry ingestion while preserving validation status and provenance.
19. **Adjudicated Longitudinal Outcome Identity and Timing**: Extract outcome identity and temporally anchor each event to explicit dates, relative expressions, or bounded event intervals; encode negation, experiencer, and uncertainty separately; deduplicate repeated mentions across longitudinal notes while preserving conflicting evidence and provenance; prevent evaluation leakage from post-index or outcome-revealing text and patient overlap across data splits; and report adjudicated metrics separately for outcome identity and timing.
20. **Patient-Level Outcome Extraction and Event-Time Anchoring**: Extract outcome identity while distinguishing note time from occurrence time; represent exact dates, bounded intervals, and uncertain dates; classify negated and hypothetical outcomes separately from observed events; deduplicate repeated longitudinal mentions before patient-level aggregation while preserving evidence and conflicts; and evaluate precision and recall separately for outcome identity and timing without assuming preset benchmarks.
21. **LLM-Based Clinician-Note Outcome Extraction**: Use large language models to identify patient outcome mentions in clinician notes, normalize event timing, distinguish current, past, and future outcome status, preserve source spans, and validate against clinician-labeled notes for recall and timing accuracy.
22. **LLM Outcome Schema, Timing, and Chart-Review Validation**: For clinician-note outcome extraction with LLMs, define an outcome schema before extraction that captures outcome type, status, evidence span, note date, onset timing, resolution timing, temporal anchor, negation, uncertainty, and adjudication status; represent exact dates, relative expressions, bounded intervals, and unresolved timing separately; sample extractions for adjudication with emphasis on uncertain, negated, conflicting, and timing-sensitive cases; and validate outcome identity plus onset and resolution timing against chart-review labels.
23. **Clinician-Note Outcome Timelines**: Identify patient outcomes in clinician notes, normalize event time from explicit dates and relative temporal anchors, retain evidence spans, label uncertainty separately from observed outcomes, and validate extracted events against chart-reviewed outcome timelines.
24. **Sentence-Level Outcome Evidence and Timing**: Extract patient outcome mentions from clinician notes, normalize each outcome's event timing, preserve sentence-level evidence for the extracted outcome and temporal anchor, and label uncertain or inferred outcomes separately from directly stated outcomes.
25. **Clinician-Note Outcome Event Timing Guidance**: For LLM-based outcome extraction from clinician notes, define an outcome/event schema with outcome type, event status, note date, event time, temporal anchor, evidence span, uncertainty state, and adjudication status; anchor events to explicit dates, relative expressions, or bounded intervals while separating note time from event time; preserve evidence spans for both outcome identity and timing; route uncertain, negated, conflicting, or timing-sensitive cases through adjudication; and evaluate precision and recall separately for outcome identification and event timing.
26. **Clinician-Reviewed Outcome Timing from Notes**: Identify outcome mentions in clinician notes, classify timing relative to clinical events, capture uncertainty and negation, retain provenance spans for outcome and timing evidence, and validate extracted labels against clinician-reviewed annotations before downstream use.
27. **LLM-Assisted Patient Outcome Temporal Anchoring**: Use LLMs to extract patient outcome mentions from clinician notes, normalize each outcome to event dates or time windows, distinguish current outcomes from historical outcomes, store supporting evidence spans, and route ambiguous timing to human adjudication.
28. **Index-Anchored Outcome Status Extraction**: From clinician notes, separately capture the outcome mention, outcome status, timing relative to index events, evidence-span provenance for both outcome and temporal claims, and adjudication state for ambiguous or conflicting note evidence.
29. **Patient-Outcome Extraction and RWD Timeline Provenance**: Use LLMs to detect patient outcome mentions in clinician notes, anchor each outcome to explicit or relative timing, flag uncertain outcome and timing statements, assemble longitudinal event timelines across notes, validate outcome identity and timing against chart review, and retain downstream real-world-data provenance fields including source note, evidence span, temporal anchor, extraction run, uncertainty flag, and validation status.
30. **Clinician-Reviewed Outcome Event Anchoring**: For LLM-assisted clinician-note outcome extraction, extract patient outcome events, normalize event timing from note dates and temporal expressions, label uncertainty separately from observed outcomes, retain evidence spans for both outcome and timing claims, and validate extracted outcomes and timing against clinician-reviewed outcome annotations.
31. **Annotated Clinician-Note Outcome Snippets**: Identify patient outcome mentions in clinician notes; capture event timing, uncertainty, negation, and source snippets for outcome and temporal claims; validate extracted labels against annotated clinical notes before downstream analytics.
32. **Adjudication-Ready Outcome Status and Timing Extraction**: From clinician notes, distinguish the outcome mention, outcome status, onset timing, resolution timing, evidence span, uncertainty, negation, and adjudication-ready provenance so extracted outcome records can be reviewed against source-note evidence.
33. **LLM-Based Patient Outcome Extraction with Adjudicated Temporal Anchoring**: Use LLMs to identify patient outcome events in clinician notes, normalize event time from explicit dates and relative temporal cues, label uncertain outcomes separately from observed or negated outcomes, retain source evidence for outcome and timing claims, and validate extracted events and timing against clinician-adjudicated outcomes before downstream use.
34. **LLM Outcome Mention Extraction with Temporal Anchoring**: Define outcome schemas before extraction; normalize event timing from note dates, explicit dates, and relative temporal cues; preserve source spans for outcome and timing evidence; separate observed outcomes from inferred outcomes; and validate extracted labels against clinician-adjudicated outcome labels.
35. **LLM Outcome Normalization and Documentation-Time Separation**: Use LLM-based extraction to normalize patient outcome mentions from clinician notes to defined outcome labels; distinguish event time from documentation time; retain evidence spans for both outcome identity and temporal anchors; label uncertainty, negation, and speculation separately from observed outcomes; and report adjudication metrics for outcome identification and temporal anchoring without assuming preset benchmarks.
36. **Asserted, Negated, and Hypothetical Outcome Anchoring**: Identify patient outcome mentions in clinician notes; normalize event timing from note dates, explicit dates, and relative temporal cues; distinguish asserted outcomes from negated or hypothetical outcomes; and preserve exact text spans for audit and adjudication.
37. **LLM-Assisted Outcome Extraction with Temporal Adjudication**: Identify patient outcome mentions in clinician notes; normalize event timing from explicit dates, note dates, and relative expressions; distinguish observed outcomes from negated or hypothetical outcomes; attach evidence spans for outcome and temporal-anchor claims; and route uncertain cases for adjudication.
38. **Adjudicated LLM Clinician-Note Outcome Schemas**: For LLM-based clinician-note outcome extraction, define event, status, and time schemas before extraction; distinguish historical outcomes from current outcomes; capture uncertainty and negation separately from observed status; preserve note evidence spans for outcome and timing claims; and validate extracted labels against clinician-adjudicated labels before downstream use.
39. **Chart-Reviewed Outcome Event Extraction from Clinician Notes**: Use LLM-assisted event extraction to identify patient outcomes in clinician notes; normalize event timing from explicit dates, note dates, and relative temporal cues; retain evidence spans for outcome and timing claims; label uncertainty separately from observed, negated, or speculative events; and validate extracted outcomes and temporal anchors against chart-reviewed outcomes before downstream use.
40. **Present/Past/Future Outcome-Event Anchoring**: Define outcome-event schemas before clinician-note extraction; normalize onset and resolution timing from explicit dates, note dates, and relative temporal cues; preserve source spans for outcome and temporal evidence; separate present, past, and future outcomes from negated or uncertain mentions; and validate extracted events and timing against clinician-reviewed annotations.
41. **Clinician-Note Patient-Outcome Timeline Workflow**: Detect outcome mentions in clinician notes; anchor each mention to the note date, explicit dates, relative temporal cues, or bounded intervals; encode negation and uncertainty separately from observed outcomes; retain evidence spans for outcome and timing claims; assign adjudication labels for confirmed, rejected, uncertain, or conflicting cases; and evaluate patient-outcome timelines with metrics reported separately for outcome detection, temporal anchoring, and adjudication agreement without assuming preset benchmarks.
42. **Outcome Extraction with Onset and Resolution Timing**: For clinician-note outcome extraction, define outcome event schemas with onset timing, resolution timing, note date, event status, temporal anchor, evidence span, negation, uncertainty, and adjudication fields; route ambiguous, conflicting, negated, uncertain, or timing-sensitive records through adjudication; and validate note-level plus patient-level extraction with separately reported outcome-identification and temporal-anchoring metrics without assuming preset benchmarks.
43. **Occurrence-and-Time Outcome Labeling**: Assign separate labels for whether a patient-outcome event occurred and when it occurred; normalize temporal relations against both note and encounter dates; encode uncertainty and negation explicitly; deduplicate repeated events across longitudinal notes while preserving provenance; resolve annotation disagreements through inter-annotator adjudication; and report event-level extraction metrics plus timing metrics at stated tolerance windows without assuming preset benchmarks.
44. **Event-and-Time Outcome Modeling**: Model outcomes as events rather than simple mentions; distinguish observed outcomes from plans and hypothetical events; normalize explicit dates and relative temporal expressions; represent intervals and uncertain timing; link each outcome and temporal claim to supporting note spans; reconcile longitudinal contradictions while preserving provenance; and evaluate event identity and temporal accuracy separately through clinician adjudication.

## Workflow

1.  **Input**: A string of clinical text or a text file.
2.  **Process**: Tokenizes and matches against patterns/dictionaries.
3.  **Output**: JSON list of entities with spans and types.

## Example Usage

**User**: "Extract entities from this note."

**Agent Action**:
```bash
python3 Skills/Clinical/Clinical_NLP/entity_extractor.py \
    --text "Patient has diabetes type 2. Prescribed Metformin 500mg. No chest pain." \
    --output entities.json
```

## References

- Park N, Ahmed F, Sun Z, Lybarger K, Breinhorst E. Automated identification of incidentalomas requiring follow-up: A multi-anatomy evaluation of LLM-based and supervised approaches. J Biomed Inform, 2026 Apr 28. https://pubmed.ncbi.nlm.nih.gov/42061667/
- Abdullahi T, Hamzeh A, Sears I, Abadi N, Singh R. Identifying and timing patient outcomes in clinician notes using large language models. Artif Intell Med, 2026 Jun. https://pubmed.ncbi.nlm.nih.gov/41886942/
- 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 Digit Oncol, 2026 Jun. https://pubmed.ncbi.nlm.nih.gov/42021926/
- Phaterpekar T, Zeng Z, Mali Y, Leung B, Ho C. Investigating fine-tuning versus zero-shot learning for general large language models when predicting cancer survival from initial oncology consultation documents. ESMO Real World Data Digit Oncol, 2026 Jun. https://pubmed.ncbi.nlm.nih.gov/42004490/

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