Multi-head CRF classifier for biomedical multi-class named entity recognition on Spanish clinical notes.

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Biomedical Named Entity Recognition (NER) is vital for extracting information from clinical texts, aiding in treatment improvements and drug development.
• Traditional NER methods often struggle with multi-class scenarios and overlapping entities in complex biomedical domains.
• Existing approaches for Spanish clinical documents are limited, hindering large-scale analysis.

Purpose of the Study:

• To propose and evaluate a multi-head Conditional Random Field (CRF) model for multi-class NER in Spanish clinical documents.
• To address the challenge of overlapping entity instances in biomedical text.
• To create the largest Spanish multi-class biomedical NER dataset by combining SympTEMIST, MedProcNER, DisTEMIST, and PharmaCoNER.

Main Methods:

• Implemented a multi-head CRF classifier to handle multiple entity classes simultaneously.
• Integrated four Spanish biomedical datasets (SympTEMIST, MedProcNER, DisTEMIST, PharmaCoNER) to form a comprehensive dataset.
• Performed entity linking to the Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT) vocabulary.

Main Results:

• The multi-head CRF model achieved a combined micro-averaged F1-score of 78.73% for NER.
• Achieved an end-to-end F1-score of 54.51% for clinical mentions normalized to SNOMED CT.
• Demonstrated competitive performance compared to single-class NER models.

Conclusions:

• The proposed multi-head CRF strategy effectively handles multi-class NER and overlapping entities in Spanish clinical texts.
• The combined dataset represents the largest Spanish multi-class resource for biomedical entity recognition and linking.
• The methodology offers a scalable and efficient solution for clinical information extraction, paving the way for relation extraction.