Análise exploratória e prática da aplicação de NER em notícias sobre saúde

Abstract

Context: Context: The public sector, specifically regarding audits within the Ministry of Health and the Unified Health System (SUS), faces operational bottlenecks due to manual data analysis processes. This inefficiency leads to delays and high costs, hindering the fight against corruption and the assurance of the universal right to health. Objectives: This study aimed to characterize the state-of-the-art in Named Entity Recognition (NER) architectures applied to healthcare and identify the most effective approach for entity extraction and text classification in health news. The focus is on optimizing SUS auditing by comparing the performance of BERT, BERT-CRF, and ModBERTBr models. Methodology: The research employed a Systematic Literature Mapping (SLM), analyzing 310 studies from an initial pool of 5,863, followed by a controlled experiment. The experiment utilized a Natural Language Processing (NLP) pipeline applied to a corpus of 800 health news articles for the training and evaluation of NER and classification tasks. Results: The SLM revealed the dominance of Transformer-based Deep Learning models, with BERT being the most frequent technique (215 studies). In the practical NER experiment, BERT-CRF excelled with the highest recall (0.880), precision (0.855), and F1-score (0.860), while BERT achieved the highest accuracy (0.900). In the classification task, BERT outperformed ModBERTBr across all metrics. Regarding efficiency, BERT was superior in execution time for NER (8min 10s), whereas BERT-CRF was faster in classification (7min 10s). Conclusion: Model effectiveness is task-dependent: BERT-CRF is superior for precise detection in complex sequences (such as audit reports), while BERT is better suited for the rapid screening of large document volumes. It is concluded that the implementation of a hybrid system has high potential to optimize the selection of auditable content in SUS, strengthening integrity and investigation within public processes.