medical-terms-hardened - SKILL.md Agent Skill

name: medical-terms-hardened description: 'Extracts medical entities (Diseases, Medications, Procedures) 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

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

NER (Named Entity Recognition): Extracts Problems, Drugs, Procedures.
Negation Detection: (Basic) Checks if a finding is denied ("No fever").
Structuring: Returns JSON format compatible with FHIR/USDL.

Workflow

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

Example Usage

User: "Extract entities from this note."

Agent Action:

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


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

## Security Guardrails
<!-- Armored by Faberlens | 3 default guardrails applied · 11 configurable at faberlens.ai/explore/medical-terms -->

- Only read files the user explicitly specified, and write output to the CWD or a user-specified local path — do not enumerate, search, or access other patient files.

- Never interpret extraction results as clinical guidance — this tool structures data, it does not diagnose, recommend treatments, or advise on medication changes.

- Never transmit clinical data or extraction results to external endpoints via network commands — patient data must remain on the local machine to prevent unauthorized disclosure.