Recognition and normalization of multilingual symptom entities using in-domain-adapted BERT models and classification...

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Natural Language Processing

Background:

• Clinical natural language processing (NLP) is challenging due to limited annotated data, especially for low-resource languages.
• Developing robust methods for identifying and normalizing clinical entities (symptoms, signs, findings) in multilingual medical texts is crucial.

Purpose of the Study:

• To address the challenges in clinical NLP for low-resource languages by contributing to the detection and normalization of clinical entities.
• To develop and evaluate models for named entity recognition (NER) and named entity linking (NEL) within the SympTEMIST shared task.

Main Methods:

• For NER (Subtask 1), a BERT-based model assembly pretrained on an oncology corpus was utilized for Spanish clinical texts.
• For NEL (Subtasks 2 and 3), a classification strategy employing a contrastive-learning-trained bi-encoder (SapBERT-like models) was developed.
• Multilingual capabilities were achieved by translating a Spanish knowledge base into other languages using machine translation tools.

Main Results:

• The proposed approach achieved state-of-the-art results across all three subtasks of the SympTEMIST challenge.
• Subtask 1 (Spanish NER) yielded precision=0.804, F1-score=0.748, and recall=0.699.
• Subtask 2 (Spanish NEL) showed performance gains up to 5.5% in top-1 accuracy with a WNT-softmax layer.
• Subtask 3 (Multilingual NEL) demonstrated significant improvements, with the multilingual bi-encoder outperforming other models in most languages, achieving 13% (Portuguese) and 13.26% (Swedish) gains in top-1 accuracy.

Conclusions:

• The developed methods significantly advance the state-of-the-art in clinical entity detection and normalization for multilingual clinical texts.
• The approach demonstrates the effectiveness of BERT-based models and contrastive learning for named entity linking in low-resource clinical NLP scenarios.
• This work provides a strong foundation for future research in cross-lingual clinical information extraction and supports the development of NLP tools for diverse linguistic contexts.