Predição de dados em inquéritos populacionais utilizando técnicas de aprendizado de máquina

Abstract

Monitoring the development of countries’ health and nutrition indicators rely on accurate and complete data. Low-and middle-income countries depend on populational surveys for decision-making since their information systems are still unreliable. These indicators are health outcomes which point out the status of a maternal and child health intervention, behavior or care. Despite being great sources of information, the populational surveys also lack information which could prevent the creation of these standardized indicators. Machine learning arise as an alternative to fill the gaps found in the surveys based on sociodemographic characteristics and related predictors. This work evaluates the performance of four machine learning techniques (neural networks, decision trees, KNN and XGB) in this task by observing the behavior of two health outcomes in four different samples. The results indicate that a great variation can be found depending on the selected outcome. The availability of those indicators as predictors is also a challenge. The studied machine learning techniques presented a small performance variation between themselves which suggests that the nature of the data is determinant to this kind of application. Still, the first outcomes showed acceptable results for a practical application while the second outcome reached lower estimates compared to those which specialists consider as a reasonable performance.