ClinLinker-KB-GP is a state-of-the-art bi-encoder model for medical entity linking (MEL) in Spanish, optimized for clinical domain tasks. It enriches concept representations by incorporating not only synonyms but also hierarchical relationships (parents and grandparents) from the UMLS and SNOMED-CT ontologies. The model was trained with a contrastive-learning strategy using hard negative mining and multi-similarity loss.

💡 Intended Use

Domain: Spanish Clinical NLP
Tasks: Entity linking (diseases, symptoms, procedures) to SNOMED-CT
Evaluated On: DisTEMIST, MedProcNER, SympTEMIST
Users: Researchers and practitioners working in clinical NLP

📈 Performance Summary (Top-25 Accuracy)

Model	DisTEMIST	MedProcNER	SympTEMIST
ClinLinker	0.845	0.898	0.909
ClinLinker-KB-P	0.853	0.891	0.918
ClinLinker-KB-GP	0.864	0.901	0.922
SapBERT-XLM-R-large	0.800	0.850	0.827
RoBERTa biomedical	0.600	0.668	0.609

Results correspond to the cleaned gold-standard version (no "NO CODE" or "COMPOSITE").

🧪 Usage

python

from transformers import AutoModel, AutoTokenizer
import torch

model = AutoModel.from_pretrained("ICB-UMA/ClinLinker-KB-GP")
tokenizer = AutoTokenizer.from_pretrained("ICB-UMA/ClinLinker-KB-GP")

mention = "insuficiencia renal aguda"
inputs = tokenizer(mention, return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
embedding = outputs.last_hidden_state[:, 0, :]
print(embedding.shape)

For scalable retrieval, use Faiss or the FaissEncoder class.

⚠️ Limitations

The model is optimized for Spanish clinical data and may underperform outside this domain.
Expert validation is advised in critical applications.

📚 Citation

Gallego, Fernando and López-García, Guillermo and Gasco, Luis and Krallinger, Martin and Veredas, Francisco J., Clinlinker-Kb: Clinical Entity Linking in Spanish with Knowledge-Graph Enhanced Biencoders. Available at SSRN:http://dx.doi.org/10.2139/ssrn.4939986.

Authors

Fernando Gallego, Guillermo López-García, Luis Gasco-Sánchez, Martin Krallinger, Francisco J Veredas