cbx-clinical-entity-extractor-17M
medical entity extraction from text
2 min readupdated
This model automatically finds and classifies clinical entities in English texts like medical case reports. It identifies 41 entity types, including symptoms, diseases, diagnoses, medication dosages, and anatomical structures. Built on the compact Ettin encoder 17m base (around 17 million parameters), it outputs a flat list of structured spans with character positions. It is built for researchers and developers analyzing medical documents.
Key features
- 41 entity types: recognizes symptoms, diseases, procedures, medications, dosages, dates, and anatomy following the MACCROBAT2020 standard.
- Fast inference: processes a document in roughly 24 ms on CPU and 7 ms on GPU.
- Up to 2046-token context window: analyzes texts of around 9,000 characters in a single pass, using overlapping windows for longer records.
- Compact footprint: based on a 16.8-million-parameter encoder that runs smoothly even on modest hardware.
What you can do
- Structure medical reports: automatically pull symptoms, prescribed drugs, and procedures from unstructured patient histories.
- Build datasets: compile structured data samples for medical research and analytical databases.
- Analyze literature: rapidly scan batches of clinical publications for mentions of specific conditions or treatment regimens.
Example
from cbx_clinical import ClinicalEntityExtractor
ner = ClinicalEntityExtractor(".") # папка с файлами модели
ner.extract("A 47-year-old woman presented with fever and right upper quadrant pain.")
# [{'label': 'Age', 'start': 2, 'end': 13, 'text': '47-year-old'}, ...]
Read more
DetailsBenchmarks: 0.547 exact-match F1 and 0.685 overlap F1 on the MACCROBAT2020 test split.
Hardware requirements: lightweight model, runs on CPU (~24 ms) and GPU (~7 ms).
Limitations: non-commercial use only (CC BY-NC 4.0), not a medical device, processes texts one document at a time.
Training details: fine-tuned Ettin encoder with a linear BIO tagging layer (83 tags) on 140 documents.










Comments