Uso de aprendizado de máquina em análise preditiva na interrupção do tratamento da tuberculose em pessoas que vivem com HIV

• Ano de defesa: 2025
• Autor(a) principal: Soares, Karllian Kerlen Simonelli
• Tipo de documento: Tese
• Tipo de acesso: Acesso embargado
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Doutorado em Saúde Coletiva; Centro de Ciências da Saúde; UFES; Programa de Pós-Graduação em Saúde Coletiva
• Palavras-chave em Português: Epidemiologia; Tuberculose; Coinfecção; Saúde Coletiva
• Link de acesso: http://repositorio.ufes.br/handle/10/19589

Objective: To build a prediction model for interruption of tuberculosis treatment in people living with HIV. Methods: This is a cross-sectional study developed in three stages: first, the analysis of the quality of SINAN data using the Centers for Disease Control and Prevention (CDC) Guide, from 2016 to 2018, with five methodological stages that included quality analysis, standardization of records, duplication analysis, data completeness through linkage with the SINAN-HIV database, and data anonymization. In the second stage, in addition to the methodological process of preparing the database and descriptive data analysis, the STATA statistical package, version 16 (StataCorp LP, College Station, TX, USA) was used to perform descriptive analyses with identification of relative and absolute values, and tables were generated for data analysis. The third stage consists of building the predictive model through machine learning using Multilayer Perceptron (MLP) and Restricted Boltzmann Machines (RBM) artificial neural network algorithms and Random Forest and CatBoost decision trees of TB-HIV co-infection, from 2016 to 2021, in Brazil, implemented in Python version 3.10.3; with validation through accuracy, sensitivity, specificity, true positive values and true negative values. The study obtained ethical approval under opinion no. 4022892 on 05/12/2020. Results: In the first stage, the study showed that 89% of the mandatory variables and 91% of the essential variables presented satisfactory completeness. In the case of TB-HIV co-infection, 73% of the variables were completed, but essential variables related to monitoring of TB treatment presented unsatisfactory completeness. In the second stage, of a total of 4,428 cases, 325 cases were of TB-HIV co-infection, 322 cases were located in the SINAN-TB database and three cases were located after linkage with the SINAN-HIV database that presented a record of a negative result for the HIV diagnostic test in the SINAN-TB database. The vulnerability profile of coinfection was observed in men (71%), young (20 to 39 years) (52%), mixed race (59%), with up to 8 years of education (25%), alcoholics (29%) and smokers (37%) and who used drugs (26%), with 65% adherence to antiretroviral therapy and only 44% with a cure outcome and 20% interrupted treatment; approximately 61% did not undergo directly observed treatment and only 6.9% of cases reported receiving assistance from the government's income transfer program. In the third stage, a total of 12,556 cases of TB-HIV coinfection in Brazil were analyzed, and the Multilayer Perceptron neural network algorithms were sensitive in identifying potential cases of treatment interruption, and were validated by an accuracy of 0.73, sensitivity of 0.75, and specificity of 0.62; Positive Predictive Value (PPV) of 0.91 and Negative Predictive Value (NPV) of 0.31. Conclusion: Training and capacity building to improve data collection, integration and analysis are essential to promote data quality. As well as social support, in order to enable access to health services and timely treatment for the most vulnerable. And finally, the implementation of new technologies, which optimize the breaking of the chain of TB transmission in people living with HIV, favoring actions aimed at screening, treatment and monitoring of cases. To strengthen care networks and promote equity in access to health services.