Clinical notes are finally usable as data
Most of what clinicians know about a patient lives in free text. Progress notes, radiology and pathology reports, discharge summaries, operative notes, nursing assessments. It’s all in your EHR, and very little of it reaches your analytics.
For years, we built healthcare data platforms around the structured part: diagnosis codes, lab values, medication lists, claims. That made sense. Structured data is reliable and easy to query. But it leaves a lot out. The reasoning behind a diagnosis, a detail a physician noticed, a housing problem mentioned in passing. Until recently, pulling that out of text at scale was either too expensive or too inaccurate.
That has changed.
Why older tools didn’t stick
Healthcare has worked on clinical text for more than a decade. Tools like cTAKES and MetaMap could find medications, diagnoses and procedures in notes, and they did fine in controlled settings. Production was harder.
The first problem was vocabulary. Rule-based systems needed exhaustive dictionaries. A note that says “pt c/o sob x 3d, worse w/ exertion” would stump a system looking for “patient complains of shortness of breath for three days.” Every abbreviation and local habit needed its own rule.
The second was context. Clinical text is full of negation and hedging. “No evidence of malignancy” contains the word malignancy and means the opposite. Tools like NegEx handled simple cases with patterns, but missed harder ones like “family history of diabetes, patient denies symptoms.”
The third was cost. Tuning a pipeline for one document type at one hospital took months. At the next hospital, with different templates and abbreviations, you started over.
Most health systems that tried ended up with a few fragile pipelines that needed constant care, and the return rarely justified the effort.
What language models change
Large language models change the cost of extraction. The same model can read a cardiology note, a radiology report and a psychiatric evaluation without retraining. It handles abbreviations and negation because it learned language, not a dictionary.
You can describe what you want and get usable structured output from text the model has never seen. “Heart attack,” “MI,” “STEMI” and “myocardial infarction” land on the same concept without a hand-built synonym list. The model understands that “ruled out for PE” means no pulmonary embolism, and that “mother had breast cancer” is about the mother.
That moves clinical text extraction from a custom project to a pipeline step. You configure it instead of building it from scratch for each use.
A pipeline you can trust
Dropping a model call into your ETL is not enough. You need a pipeline built for PHI protection, auditability, cost control and clinical review. The pattern I see working has four layers:
- De-identify. Before any text reaches a model, remove or tokenize PHI with a tool like Philter or Presidio. The text goes to the model. The mapping back to the patient stays inside your environment.
- Extract. The model reads the de-identified text and returns structured JSON: diagnoses, medications, procedures, social factors, findings. Each item comes with a confidence score and the exact text that supports it. Those source spans are required. Without them you can’t check or audit anything.
- Code. Map each extracted item to a standard vocabulary such as SNOMED CT, ICD-10, RxNorm or LOINC. That can be a second model call, a similarity search against an index of terms, or both. Every concept should end up with a standard code.
- Load. Coded results land in your warehouse, whether that’s Snowflake, BigQuery or Databricks, as tables that join to your claims, labs and administrative data. Analysts query them with ordinary SQL.
Cost is a design choice, too. The price per note looks small until you multiply it by every note your system writes in a year. Use smaller, cheaper models for simple work like medication lists, and save larger models for harder reasoning like rebuilding a clinical timeline. Batch the work, and reuse results where templated reports differ only in their findings.
Where it’s working
A few uses already working in production:
- Clinical trial matching. Eligibility often depends on details buried in notes, like prior treatments or lab trends. Reading notes automatically makes screening much faster.
- Social needs. Physicians often write about housing, food, transportation or substance use, but rarely enter structured codes for them. Extracting those mentions gives population health teams data they didn’t have.
- Quality abstraction. Instead of reading every chart, abstractors check cases that arrive pre-filled.
- Prior authorization. Gathering the supporting findings from across a record takes a nurse a long time per case. Extraction can assemble the evidence for review.
- Research cohorts. Researchers can find patients like “treatment-resistant depression with at least two SSRI trials” in narrative text, not only in codes that may be incomplete.
The HIPAA question
Can you send clinical notes to a language model? Yes, with care. There are three main paths.
De-identify first. Strip PHI before any text leaves your environment. You lose some context, such as names, dates and places, but you can meet HIPAA’s Safe Harbor standard.
Use a covered service. Major cloud providers and some model vendors will sign a business associate agreement. With a BAA in place you can send PHI, but check their logging, retention and data handling against your own policies.
Run the model yourself. Open-weight models can run entirely inside your own infrastructure, and for many extraction tasks they are getting close to the hosted ones.
Many health systems mix these, sending de-identified text to cloud models for complex reasoning and running their own models for high-volume routine work.
Where to start
If you’re building or updating a data platform this year, clinical text belongs in the design now rather than on a later roadmap.
- Take stock of your text. Which document types do you have, in what volume, and what clinical information exists only in free text?
- Pick one use case with a clear return, such as social needs, quality abstraction or prior authorization, and build it end to end.
- Design the tables now. Every extracted item should record its source document, the sentence it came from, the model version and a confidence score.
- Build review into the workflow. Have people check extractions, especially for anything close to clinical decisions. The checked examples become your test set for measuring accuracy over time.
Do step one this month. A count of note types and volumes, plus a short list of questions your analysts can’t answer from codes alone, will show you where the first pipeline belongs.
Discussion