DT4H/CardioBERTa.ro_GP_enriched

DT4H_CardioBERTa_grandparents_ro_enriched

DT4H_CardioBERTa_grandparents_ro_enriched is a Romanian biomedical terminology encoder for clinical concept normalization and entity linking. It is initialized from [DT4H/CardioBERTa.ro] and specialized using CUI-supervised terminology pairs and metric learning.

Backbone

The backbone belongs to the CardioBERTa family from CardioLM - a multilingual suite of small language models for the cardiology domain. CardioBERTa comprises language-specific encoder models adapted to cardiology through continued pretraining on monolingual biomedical and cardiology-related corpora using Masked Language Modeling (MLM). The family covers Czech, Dutch, English, Italian, Romanian, Spanish and Swedish.

Training

Language	Romanian (ro)
Triplet collection	enriched
Strategy	grandparents
Objective	Multi-Similarity Loss
Mining	All triplets, margin 0.2
Pooling	CLS
Epochs	1
Batch size	256
Learning rate	2e-5
Max. length	25

CUI-supervised terminology pairs enriched with grandparent-level ontology relations.

Terminology statistics

Strategy	Triplets	CUIs	Unique terms	Unique positives	Terms/CUI	Δ terms
synonyms	70,817	70,817	139,248	70,439	2.00	0
parents	1,607,064	476,350	531,693	422,975	3.94	+392,445
grandparents	4,734,361	476,970	531,980	470,719	9.85	+392,732

This model uses 4,734,361 triplets, covering 476,970 CUIs and 531,980 unique normalized terms.

The training terminology is not distributed with this repository because it contains resources subject to UMLS licensing conditions. Only aggregate statistics are released.

Intended use

The model is intended for terminology embedding, biomedical candidate retrieval, concept normalization and entity linking, particularly in cardiology and clinical NLP pipelines. It is not intended for direct clinical decision-making.

Usage

import torch
import torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer

model_id = "DT4H/DT4H_CardioBERTa_grandparents_ro_enriched"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModel.from_pretrained(model_id)

inputs = tokenizer(
    "clinical concept",
    return_tensors="pt",
    truncation=True,
    max_length=25,
)

with torch.no_grad():
    output = model(**inputs)

embedding = F.normalize(
    output.last_hidden_state[:, 0, :],
    p=2,
    dim=1,
)

Reference

Danu et al. CardioLM - a multilingual suite of small language models for the cardiology domain.

Developed within the DataTools4Heart (DT4H) project, Grant Agreement 101057849.