Evaluating A Deep Learning Model In Integrated Health Databases To Aid In The Prognosis of Tuberculosis

Tuberculosis is one of the main causes of death in the world, and Brazil has high rates of this disease. In this work, we evaluate a deep learning model, called DeepTub, for the prognosis of tuberculosis. With a well-defined methodology, we carried out two experiments to analyse whether there are significant improvements when using an integrated database. Then more cases of tuberculosis death were added, and the results were compared with the original model. Different record linkage techniques were applied to integrate the databases, with the databases resulting from each technique being used, first, as the input for the model, without modifying it, and, later, the grid search technique was applied to analyse the best configuration of the hyper-parameters. In both experiments, the model obtained better results in the AUC ROC metric using the original Information System for Notifiable Diseases (SINAN) database, with 74.15% and, in the second experiment, two integrated databases, SoundexBR and Metaphone_pt-BR, showed statistical similarity with the original SINAN database in the model results, with 73.92% and 73.92%, respectively. The DeepTub model, with the application of record linkage techniques, did not present relevant results regarding the improvement of evaluation metrics.