clinical-notes-nlp

Overview

clinical-notes-nlp covers natural language processing of clinical documentation: discharge summaries, progress notes, radiology reports, nursing notes. It spans de-identification, medical concept extraction (UMLS/cTAKES/MedCAT), note-based phenotyping, and the special caution required when applying large language models to clinical text — where hallucinations can become chart entries.

Clinical notes are the richest and messiest EHR data: telegraphic, abbreviation-dense, copy-pasted, and full of implicit context. Standard NLP pipelines underperform until adapted.

When to use

• De-identifying clinical text for research (HIPAA Safe Harbor, surrogate replacement).
• Extracting concepts: diagnoses, medications, procedures, symptoms with negation/temporality.
• Note-based phenotyping: combining notes with structured data for case definitions.
• Information extraction: tumor staging from pathology reports, ejection fraction from echo reports, social determinants from notes.
• Evaluating LLMs for summarization or documentation assistance with safety guardrails.
• Measuring documentation quality: copy-paste detection, note bloat.

Core concepts

• De-identification. HIPAA Safe Harbor: remove 18 identifier categories (names, dates except year, locations, contact info, MRNs...). Automated tools (Philter, de-identification modules) miss edge cases — validate on samples, especially dates in narrative text and clinician names in signatures. Surrogate replacement (realistic fake names/dates) preserves readability better than redaction tags.
• Clinical sublanguage. Notes use heavy abbreviation ("pt c/o SOB, r/o PE"), telegraphic syntax, and institution-specific shorthand. Off-the-shelf NLP models trained on news/wikipedia fail; use clinical models (ClinicalBERT, GatorTron) or rule-based clinical pipelines, and build an abbreviation dictionary per site.
• Negation and uncertainty. "No evidence of pneumonia," "possible PE," "family history of MI" — concept extraction without negation/temporality/subject detection (NegEx/ConText algorithms) produces garbage phenotypes. Every extracted concept needs: negated? historical? about the patient or family? uncertain?
• Concept normalization. Map mentions to standard vocabularies (UMLS CUIs, SNOMED CT, RxNorm for meds). Tools: cTAKES, MetaMap, MedCAT (which learns site-specific embeddings). Ambiguity ("cold" = temperature vs viral illness) needs context-aware disambiguation.
• Copy-paste and note bloat. Cloned text propagates errors and inflates concept counts. Detect duplication (text similarity across notes); for phenotyping, deduplicate or down-weight repeated content — otherwise one copied problem list counts ten times.
• Phenotyping from notes. Notes capture what codes miss (symptoms, severity, social context). Best practice: combine structured + NLP features, validate against chart review, report PPV/sensitivity. Note-only phenotypes inherit documentation bias (what clinicians bother to write).
• LLMs on clinical text. Useful for summarization and first-draft documentation, dangerous as autonomous extractors. Guardrails: human-in-the-loop for anything entering the chart, grounded extraction (every claim traceable to source text spans), hallucination testing on adversarial cases, and no PHI in prompts to external APIs without a BAA. An LLM summarizing a chart must never invent allergies, doses, or code statuses.
• Evaluation. Intrinsic (precision/recall/F1 on annotated spans — annotate with dual reviewers and adjudication, report inter-annotator agreement) and extrinsic (does the phenotype predict what it should?). Report both.

Practical workflow

1. Governance. IRB, de-identification plan, BAA for any external compute. Notes are the most sensitive EHR data — treat them accordingly.
2. Corpus. Define note types and time windows; sample representatively; de-identify with validated tooling.
3. Annotate. Guidelines, dual annotation, adjudication, IAA reporting (F1 or kappa on spans). Budget more time than expected — clinical annotation is slow.
4. Extract. Pipeline: section detection → sentence splitting → concept extraction → negation/temporality/subject → normalization. Validate each stage.
5. Phenotype. Combine with structured data; validate against chart review; report PPV, sensitivity, and failure modes.
6. LLM use (if any). Grounded prompts, human review of outputs, hallucination audits, PHI controls. Document the human-in-the-loop design.
7. Monitor. Documentation practices drift (new templates, scribes, ambient AI); revalidate periodically.

Example (Python sketch):

from medcat import CAT
cat = CAT.load_model_pack("umls_model_pack.zip")
doc = cat(doc_text)  # entities with CUIs, negation, temporality metadata

Common pitfalls

• Off-the-shelf NLP without clinical adaptation (abbreviation soup defeats it).
• Concept extraction without negation detection ("no MI" counted as MI).
• Copy-pasted text inflating phenotype counts.
• De-identification validated on easy cases, failing on narrative dates/names.
• LLM hallucinations entering clinical documentation or research phenotypes unchecked.
• PHI sent to external LLM APIs without authorization.
• Single-annotator "gold standards" with no agreement reporting.