BERT-Based Approach for Tuberculosis Diagnosis from Electronic Medical Records

Tuberculosis (TB) is an infectious disease and remains one of the leading causes of death worldwide. Its diagnosis poses significant challenges, often depending on the resources and capacities available within healthcare institutions. The analysis of data from various sources has enabled the development of models based on artificial and computational intelligence. In this study, Bidirectional Encoder Representations from Transformers (BERT) models were applied to classify electronic medical reports for TB detection. Their performance was compared with that of traditional natural language processing (NLP) approaches to evaluate improvements in classification accuracy. Results indicate that BERT is a promising strategy for this specific task, achieving an area under the ROC curve (AUC) of 71.81% in the best scenario, which depends on computational resources and data availability.