Desenvolvimento e validação de modelos preditivos para malária.

Abstract

Data mining techniques allow the extraction of relationships and characteristics from data which are not easily identifiable, specially in big data scenarios. These techniques provide routines to group and classify data, identify patterns and outliers, as well as derive associations among distinct data. Predictive analytics is a part of the data mining workflow that relies on machine learning to implement different prediction models. These models are based on previous knowledge about the data to forecast possible outcomes related to the domain being considered. This work belongs to a project focusing on the integration of different data sources from the Brazilian malaria ecosystem and on the design of different predictive analytics models to be used over these data sources. More specificallly to design a predictive analytics model applied to malaria surveillance in Brazil. Four case studies comprising epidemiologic and climate data are discussed as proof of concepts. For the evaluation of the models the methodology of cross-validation for time series evaluation on rolling forecasting origin was used, the metric used was ac RMSE. Four machine learning algorithms were used, support vector regression and random Forest obtained the best results. Within the four case studies, two achieved greater accuracy using weather data in some scenarios.