Transformer-based relation extraction and concept normalization using an annotated clinical trials corpus.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Clinical Research

Background:

• Manual review of electronic health records (EHRs) for clinical trial patient selection is time-consuming.
• Limited availability of Spanish language resources hinders NLP applications in clinical research for Spanish-speaking populations.

Purpose of the Study:

• To present version 3 of the CT-EBM-SP corpus, a comprehensive Spanish clinical trial dataset.
• To facilitate automated patient eligibility screening using NLP techniques.
• To advance medical concept normalization and relation extraction in Spanish clinical text.

Main Methods:

• Developed and annotated the CT-EBM-SP corpus (1200 trials, 292,173 tokens) with 23 entity and 18 relation types.
• Included UMLS semantic groups, drug information, temporal data, negation, and speculation.
• Encoded 11 attributes and normalized entities to UMLS Concept Unique Identifiers (CUIs).
• Evaluated inter-annotator agreement (IAA) for entities, attributes, and relations.
• Benchmarked Transformer models for relation extraction (RE) and medical concept normalization (MCN).

Main Results:

• Achieved high IAA for entities (F1=0.861), attributes (F1=0.810), and relations (F1=0.791).
• Successfully normalized 81.75% of entities with high IAA (F1=0.966).
• Transformer models achieved strong performance: RE (avg. F1 0.858-0.879) and MCN (accuracy 0.896 at rank 1).

Conclusions:

• The CT-EBM-SP corpus is a valuable resource for NLP research in Spanish clinical trials.
• The developed models demonstrate the effectiveness of Transformer architectures for RE and MCN tasks.
• Public availability of the corpus and models supports further advancements in clinical trial recruitment and NLP in Spanish.