metadata
language: uk
license: apache-2.0
tags:
- spacy
- token-classification
- ner
- medical
- cardiology
- ukrainian
model-index:
- name: uk-cardio-ner
results:
- task:
type: token-classification
name: NER
metrics:
- type: f1
value: 0.9021
name: F1
- type: precision
value: 0.9102
name: Precision
- type: recall
value: 0.8942
name: Recall
Ukrainian Cardiology NER (uk-cardio-ner)
Named Entity Recognition (NER) model for Ukrainian medical discharge summaries Trained on real cardiology cases from the cardiovascular surgery clinic. The dataset used for training underwent rigorous cleaning, thorough depersonalization, and complete anonymization to ensure strict compliance with medical data privacy standards.
Model Description
This model extracts medication-related entities from Ukrainian clinical texts.
It was trained using spaCy with the youscan/ukr-roberta-base transformer
as the backbone.
Entities
| Label | Description | Example |
|---|---|---|
| DRUG | Drug name (trade or INN) | Бісопролол, Ксарелто |
| DOSE | Dosage | 5 мг, 10/2,5/5 мг |
| FREQ | Frequency | щоранку, 2 р/д |
| DUR | Duration | тривало, 14 діб |
Training Data
- Source: Discharge summaries from cardiovascular surgery clinic
- Size: 5,855 annotated medication prescription strings
- Language: Ukrainian
- Domain: Cardiology, cardiac surgery
- Annotation: Auto-labeled via DeepSeek LLM + manual validation
Performance
Evaluated on 1,171 held-out examples (20% split):
| Metric | Score |
|---|---|
| F1 | 90.21 |
| Precision | 91.02 |
| Recall | 89.42 |
Usage
import spacy
nlp = spacy.load("uk-cardio-ner")
doc = nlp("Бісопролол 5мг щоранку, тривало.")
for ent in doc.ents:
print(ent.text, ent.label_)
# Бісопролол DRUG
# 5мг DOSE
# щоранку FREQ
# тривало DUR
Citation
If you use this model, please cite:
TAIS DSS — Trusted AI for Intelligent Specialists IntracomUA OÜ, 2026 https://tais.esemi.org