Modelos baseados em redes neurais artificiais para o diagnóstico em triagem de tuberculose resistente e multirresistente no Brasil

Abstract

Drug-Resistant TB (DR-TB) implicates in more complex treatments and leads to higher deceased and morbidity numbers. In this context, we propose the use of screening tests that early identi es patients with higher probability of having DR-TB and prioritize them. Arti cial Neural Networks and Classi cation And Regression Tree models are generated, and a boosting algorithm is applied, considering as input the patient's symptoms and social-demographic variables. Speci c scores by each State are produced and the results are compared to a national-wide approach. Models with di erent complexity levels were developed in order to t the available resources in each site, being guided by variable relevance and data quality. Models developed by each State achieved an average sensitivity higher than 85% when screening RJ patients considering DR-TB from non DR-TB, against 82.7% using the national approach, indicating that the local clinical scores can better capture operational di erences present in the health system.