Clinical NLP
Last reviewed: 2026-06-28 — see the freshness policy.
Learning objectives
After this chapter you will be able to:
- Explain what clinical NLP extracts and why unstructured notes matter.
- Choose between managed clinical-NLP services and custom/LLM approaches.
- Design an NLP pipeline that maps text to standard terminologies under PHI constraints.
Why clinical NLP
Most clinical information lives in unstructured text — progress notes, discharge summaries, radiology reports, pathology reports. Structured fields capture a fraction of what the clinician knew. Clinical NLP turns that free text into structured, coded data: diagnoses, medications, procedures, anatomy, severity, negation ("no evidence of pneumonia"), and temporality.
The output feeds analytics, OMOP population, cohort selection, and RWE — extracting facts that were never in the structured record.
flowchart LR
Note["Clinical note (free text)"] --> NLP["Clinical NLP"]
NLP --> Ent["Entities: conditions, meds,<br/>procedures, anatomy"]
NLP --> Neg["Negation / uncertainty"]
NLP --> Codes["Mapped codes<br/>(SNOMED, RxNorm, ICD-10)"]
Codes --> FHIR["FHIR / OMOP"]
What makes it hard
- Negation & uncertainty — "no chest pain", "cannot rule out" invert meaning; naive keyword matching fails.
- Abbreviations & context — "MS" = multiple sclerosis, mitral stenosis, or morphine sulfate depending on context.
- De-identification — notes hide identifiers in prose (see De-identification & consent); scrubbing text is harder than dropping columns.
- Terminology mapping — extracted concepts must map to standard codes (see terminologies) to be useful.
Approaches
| Approach | When |
|---|---|
| Managed services — AWS Comprehend Medical, GCP Healthcare NL API, Azure Text Analytics for Health | Fast start, standard entities + code mapping, HIPAA-eligible. Default choice. |
| Open-source / framework — spaCy/medspaCy, John Snow Labs Spark NLP, cTAKES | Custom entities, on-prem control, large-batch processing. |
| LLMs (Bedrock, Vertex, Azure OpenAI, self-hosted NIM) | Flexible extraction, summarization, reasoning over notes — but require careful evaluation and PHI controls. |
The managed services (see the cloud platform chapters) already map toICD-10-CM/RxNorm/SNOMED and are HIPAA-eligible — usually the pragmatic starting point.
Design under PHI constraints
- Keep PHI in-boundary. Prefer HIPAA-eligible managed NLP, or self-hosted models (e.g. NVIDIA NIM) so notes are not sent to an un-covered API.
- Human-in-the-loop for clinical use. NLP output informing care needs review; extraction errors have clinical consequences.
- Evaluate against a gold standard. Measure precision/recall on annotated notes per entity type — especially negation.
- Map, then validate. Send extracted codes through a terminology service to confirm they are valid standard concepts.
Lab
Clinical NLP pairs naturally with the RAGonGCP lab (extract + index clinical text) and feeds the hls-lakehouse-rwd OMOP pipeline.
Check yourself
- Why is negation detection essential, and what goes wrong without it?
- When would you choose a managed clinical-NLP service over an LLM, and vice versa?
- How do you keep PHI in-boundary when running NLP over clinical notes?