clinical-nlp

Clinical NLP

Information extraction from clinical free text. Covers note parsing and sectioning, biomedical named entity recognition, concept normalization to UMLS and ICD-10, assertion and negation detection, temporal relation extraction, adverse event identification, and de-identification. Written for public and credentialed research corpora, not for live patient records.

When to Use This Skill

Activate when the user requests:

• Entity extraction from clinical notes, discharge summaries, or pathology reports
• scispaCy, medspaCy, cTAKES, MedCAT, or QuickUMLS pipelines
• Concept normalization to UMLS CUIs, SNOMED CT, RxNorm, or ICD-10
• Negation, uncertainty, or family-history assertion detection
• Adverse event or medication extraction from narrative text
• Temporal relation extraction from notes
• De-identification of clinical text
• Choosing between a rule-based pipeline and a clinical transformer

Inputs

Before Any Code: The Access Question

Clinical text is not like expression data. The corpus decides what you can do.

  PubMed abstracts        open. No agreement. Use freely.
  BC5CDR, NCBI-Disease    open annotated corpora. Good for benchmarking NER.
  MedMentions             open, UMLS-linked. Good for entity linking.
  MIMIC-IV-Note           PhysioNet credentialed: CITI training + signed DUA.
                          Cannot be redistributed, posted, or pasted into a
                          hosted LLM API. The DUA prohibits it.
  n2c2 (formerly i2b2)    separate DUA through the hosting institution.
  UMLS Metathesaurus      free but licensed. Required by QuickUMLS, MedCAT
                          and the scispaCy UMLS linker. You must accept the
                          licence and download it yourself; no package ships it.

Two consequences that catch people:
  - Sending credentialed note text to an external API is a DUA breach, not a
    grey area. Run models locally on that text.
  - Never commit note text, even "de-identified" text, to a repository.
    De-identification is imperfect (below), and the DUA governs regardless.

Develop on the open corpora. Move to credentialed data only inside the
environment the DUA allows.