PREDICTIVE MODELS FOR INFANT MORTALITY IN THE STATE OF PARANÁ

Resumen

The prediction of infant mortality guides preventive measures regarding deaths. The objective of this article is to evaluate models that could predict the early, late and post  neonatal childhood mortality rate in the state of Paraná. Public available data were adopted for the period between 1996 and 2014. Predictive models for the period from 2015 to 2018 were compared, taking into account the mean absolute percentage error rate (MAPE). We also evaluated models for the period 2011 - 2014, given the possibility of comparing the expected results with the real value. We used three algorithms, namely: Linear Regression, Support Vector Machine Algorithm (SVM) and MultiLayer Perceptron (MLP). From the results, it was possible to identify greater precision in the models generated by MLPs, with a MAPE of 1% in 2015, 6% in 2016, 6% in 2017 and 8% in 2018, this model had an average accuracy of 30% higher if compared to the model that makes use of SVM (SMOreg), and in average it is 60% more accurate than the linear regression model. It is believed that increasingly refined adjustments in the MLP algorithm may further increase its accuracy and that the usage of other non-absolute quality measurements could demonstrate to which direction the prediction error is moving to.