John Snow Labs

08/21/2026

Type "heart attack." Get SNOMED concept 22298006 and ICD-10 code I21. Type "MI." Same result. Type "myocardial infarction." Same result - because they are the same clinical fact.

The John Snow Labs Medical Terminology Server sits between the free-text your clinicians write and the standardized codes your downstream systems require. It understands clinical meaning across abbreviations, shorthand, typos, and patient-friendly language.

16+ medical vocabularies. On-premise. Deterministic. Full video overview in the comments.
https://hubs.li/Q04twkpQ0

Complete guide to using John Snow Labs' Medical Terminology Server ...

08/20/2026

Your NLP model extracts "metformin 500mg" from a discharge summary. The extraction is correct. Without standardization, the entity cannot connect to any downstream system.

- Your medication reconciliation system needs NDC codes
- Insurance claims require HCPCS
- Research databases expect RxNorm
- EHR integration demands standardized codes

Generative AI Lab's Medical Terminologies solve the translation step directly in the annotation workflow:
- ICD-10, LOINC, CPT, SNOMED CT, RxNorm, and MeSH integrated into annotation
- Automated code resolution during pre-annotation
- Manual lookup available in the interface for human review
- Entity, label, and code travel together through the workflow

Medical codes are the shared language that makes clinical NLP outputs connect to the systems that actually use them.

Learn more: https://hubs.ly/Q04twpV-0

08/19/2026

Annotated clinical datasets stay reusable even when your label schema changes. A concept labeled "Diagnosis" in 2019 and "Primary_Diagnosis" in 2023 no longer blocks the import.

Healthcare AI programs run parallel annotation teams, integrate vendor pre-annotations, and bootstrap new models on earlier training data. Annotations move between projects and systems constantly, and schema drift comes with it.

Generative AI Lab reconciles it at import with no-code label mapping:
- Proposes matches, you approve, it applies them as data is imported
- Human-in-the-loop, every mapping auditable
- Works across NER, assertion, classification, and relation labels
- Source data untouched

Carry years of annotation investment forward instead of rebuilding it.

Link: https://hubs.ly/Q04twkYc0

08/15/2026

Reuse 5,000 already-validated clinical notes in a new project without writing a single script to rename labels.

The blocker is usually naming. One project labeled a concept "diagnosis," another "Diagnosis," a third "Primary_Diagnosis." The clinical meaning is identical, the import fails anyway, and teams lose days transforming data that was already annotated and validated.

Generative AI Lab handles naming at import with no-code label mapping: case-insensitive matches like "diagnosis" to "Diagnosis" resolve automatically, while more complex variations like "Medication_Name" to "Medication" surface as suggestions for human review. Administrators approve the final mappings, every mapping is auditable, and the source annotation files are never modified.

The hard part is reusing annotations. Import-time mapping is what makes it practical.

https://hubs.li/Q04sJDq90

08/14/2026

Your EHR contains more data than your analytics can see.

The gap is in how the coding layer reads clinical language. A system built on keyword matching recognizes one spelling, one abbreviation, one phrasing. "HTN" and "hypertension" describe the same condition. "T2DM" and "E11" point to the same ICD-10 code. "MI" and "myocardial infarction" map to the same SNOMED concept.

Keyword-dependent systems fail on all of these - silently. No error message. No flag. No alert. The code does not get assigned. The record disappears from the denominator. The cohort shrinks. Nobody finds out until a downstream audit, if they find out at all.

The Snow Labs Medical Terminology Server was built to close this gap. 16+ medical vocabularies. Semantic search, spelling correction, abbreviation handling. On-premise. Deterministic: the same input always returns the same code.

2-minute explainer on why traditional approaches break down (link in comments).
https://hubs.li/Q04sJGkJ0

When a clinician types "GERD" and the system returns nothing becaus...

08/13/2026

De-identify DICOM images and their metadata inside Generative AI Lab, and verify every change before export.

- Masks image content and DICOM metadata in one project
- Per-field strategy: Masked with Characters, Masked with Fixed-Length Characters, Obfuscation, or None
- Comparison view shows original vs de-identified, image and metadata side by side, before anything leaves
- Runs on-premises; never send PHI to a third-party API to get it de-identified

Upload MRI and CT scans, set masking rules per field, review the diff, and export research-ready DICOM.

Link: https://hubs.li/Q04sHF270

08/12/2026

PDF annotation on clinical documents should not require deploying an OCR server first.

Most NER teams annotating clinical documents face the same barrier: deploy a licensed OCR server - cost, setup, infrastructure - or skip PDF input entirely. Neither works for a team getting started.

Generative AI Lab's built-in PDF text extraction removes the barrier:

- Extract text from PDFs during import, no licensed-server deployment required
- Preserve document structure: reading order, paragraphs, sections
- Zero licensing cost, zero setup - works for standard clinical PDF workflows out of the box

When your requirements evolve - domain-specific accuracy, complex layouts, strict precision thresholds - licensed OCR from John Snow Labs upgrades the pipeline without changing the workflow.

For teams building NER annotation workflows on clinical documents, this is the difference between "we can't afford to start" and "we're annotating PDFs today."

Learn more: https://hubs.li/Q04sJKF40

08/12/2026

Annotate DICOM files directly in Generative AI Lab, with no conversion script and no PNG export before the work begins.

Your 5,000 chest X-rays sit in PACS as DICOM. The platform accepts .dcm through the UI or API and renders them at original image quality, so annotation starts on upload. Multi-frame studies come in as a single task with each frame on its own page, and the DICOM metadata stays attached throughout. Everything runs inside your environment; patient images never leave your network.

This does not replace PACS viewers. Radiologists keep their clinical viewers, and annotation teams start with the data they already have.

Link: https://hubs.li/Q04sJzs00

08/07/2026

A 2024 study in npj Digital Medicine found that NLP on clinical notes identified adverse social determinants of health in 93.8% of patients. ICD-10 Z-codes found 2.0% of the same patients.

That 91-point gap is a structural problem. The clinical signal that determines outcomes
lives in text, not in codes. For any regulatory submission asking whether outcomes
correlated with housing, food, or transportation security, structured data is absent.

The FDA's real-world evidence guidance, effective February 2026, makes this a compliance issue. Every clinical fact in a submission must now be defensible individually - accurate, complete, and traceable. Sponsors built on structured-only
pipelines are about to find out what they've been missing, and retrofitting NLP into a pipeline designed around codes is not the same as building for completeness from the start.

Full piece in CIO: https://hubs.li/Q04rw6hB0

08/06/2026

Processing 1 million patient records with a locally deployed Medical LLM costs about
$1M. The same workload through frontier APIs costs $13M to $30M.

That gap isn't about API vendor choice. It's structural. Per-token pricing grows linearly
with data volume. Per-server licensing doesn't: a 100x increase in patients required 15x the infrastructure in this model. Two curves with those shapes always diverge.

At pilot scale - 10,000 patients - the premium is 1.35x to 3.17x depending on the API.
Manageable. At a million patients, it's a different budget category entirely.

https://hubs.li/Q04rwcY90