Implement text analysis solutions 4.1 Apply language model text analysis
A hospital analytics team must process discharge summaries to extract diagnoses…
Study guide skill: Implement solutions to extract entities, topics, summaries, and structured JSON outputs by using generative prompting and Foundry Tools
Question B5-011
A hospital analytics team must process discharge summaries to extract diagnoses, medications, and dosages, link each to standard UMLS codes, keep negated findings such as 'no signs of pneumonia' from being counted as present, and return the results in FHIR format. Which capability should they use?
Select one answer.
Answer: A
Text Analytics for health performs named entity recognition, relation extraction, entity linking to UMLS vocabularies, and assertion detection in a single API call, and it can return the output in the FHIR structure. Assertion detection adds contextual modifiers such as certainty, which is what keeps negated findings from being treated as present. Prebuilt NER covers general entities such as people, places, and organizations, PII detection finds and redacts sensitive data, and custom NER requires you to label data and train a model without these healthcare-specific functions.
Deep dive
How it works
Text Analytics for health is a prebuilt Azure Language capability for medical text such as doctor's notes, clinical documents and discharge summaries. It runs four functions in a single API call, and each one covers part of this requirement:
- Named entity recognition extracts words and phrases of supported medical types, such as diagnosis, medication name, symptom or sign.
- Relation extraction connects related concepts, for example a medication to its dosage, or a condition to a time.
- Entity linking maps each entity to preferred names and codes from biomedical vocabularies in the UMLS Metathesaurus.
- Assertion detection adds contextual modifiers (certainty, conditionality, association, temporality). That's what marks "no signs of pneumonia" as negated rather than present.
It can also return output in the FHIR structure for integration with other electronic health systems.
Two governance points: it isn't a medical device or diagnostic tool, and decisions that affect individuals should be made with human oversight. Customers also license the UMLS source vocabularies they use.
The other options, and when you'd use them
- Prebuilt NER: It identifies general entities such as people, places, organizations and quantities. No UMLS linking, no assertion detection, no FHIR output.
- PII detection with redaction: It identifies, classifies and redacts sensitive data for privacy controls. It doesn't extract clinical concepts.
- A custom NER model: You'd label data and train a model to extract your own entity types. You'd still lack built-in UMLS linking, assertions and FHIR, which Text Analytics for health provides prebuilt.
Exam tip
Medical text + UMLS codes + negation/certainty + FHIR = Text Analytics for health. Assertion detection is the negation piece.
References
Last reviewed: