Identification of important predictors, using V-Cramer technique and Permutation feature importance, in predicting Tuberculosis treatment status

Abstract

In this study, we investigate the use of a fully connected neural network implemented with Keras to create prediction models for the treatment outcomes of tuberculosis patients. The data used were from the Sistema de Informação de Agravos de Notificação (SINAN) of the year 2023, integrated by iTB, a software developed by the Health Department of Manaus-AM. This work is still in progress. So far, we have achieved an accuracy of 92.80% and F1-Score values of 0.90 for cure, 0.89 for abandonment, 0.98 for TB-related death, and 0.99 for drug resistance. Additionally, applying the permutation-importance technique, we obtained an accuracy of 93.31%. Experiments with patient consultation data will still be added to the test data, along with a Cramer's V analysis.