NER y clasificación automática de documentos médicos con la aplicación de transformers

Abstract

Abstract: This work presents the application of Transformer-based models for the automation of medical text analysis in Spanish, focusing on Named Entity Recognition (NER) and the classification of clinical documents. Medical texts from public datasets and real documents were used, which were anonymized and subjected to minimal cleaning in order to preserve their original content. Throughout the development of this study, three pre-trained models focused on the clinical domain were evaluated. The evaluation was carried out without fine-tuning or the use of manual BIO tagging schemes, with the objective of observing how these models behave in real-world scenarios. Among the evaluated models, MEDSPANER stood out for its ability to accurately identify relevant clinical entities, allowing its direct use. The extracted entities were then used in the classification of medical documents. The results indicate that, in certain cases, it is possible to use specialized Spanish-language models for clinical tasks without the need for fine-tuning, which facilitates the analysis and organization of medical information.