A comparative analysis of Spanish Clinical encoder-based models on NER and classification tasks

Area of Science:

• Natural Language Processing (NLP)
• Computational Linguistics
• Health Informatics

Background:

• A significant gap exists in NLP resources for the Spanish language, particularly within the clinical domain.
• Effective Spanish language models are crucial for clinical research and healthcare delivery due to the large Spanish-speaking population and increasing use of electronic health records.

Purpose of the Study:

• To assess the efficacy of encoder Language Models for clinical tasks in Spanish.
• To identify the most effective Spanish language and clinical language models for these tasks.

Main Methods:

• Evaluation of 17 distinct corpora focusing on clinical tasks.
• Benchmarking of Spanish Language Models and Spanish Clinical Language Models (encoder-based).
• Fine-tuning of over 3000 models.

Main Results:

• General-purpose encoder models performed better than specialized clinical models.
• Larger models were not consistently superior; RigoBERTa 2, a general-purpose model, achieved the highest average F1 score of 0.880.
• The study identified RigoBERTa 2 as the top-performing model.

Conclusions:

• Dedicated encoder-based Spanish Clinical Language models show advantages over generative models.
• A scarcity of diverse corpora, primarily focused on Named Entity Recognition (NER), necessitates further research.
• The limited availability of high-performing models underscores the urgent need for continued development in Spanish clinical NLP.