Qualidade de dados através de árvores de decisão

Abstract

After the implementation of a program based on artificial neural networks for the screening of patients for the diagnosis of tuberculosis in Health Units of the Brazil, it was identified that the data collected could be “biased”. This suspicion occurred when comparing the data and outputs obtained by the neural network with the diagnosis of doctors. Through interviews and investigations by those responsible for the software implemented, it was identified that users of the program, often under pressure by working conditions, they could be altering patient input data so that they could obtain, as a possible diagnosis by screening, an alternative other than tuberculosis. In this work, a methodology is proposed that seeks to identify the reliability of the information processed by the system, as well as applying techniques that delay and demotivate the user's tendentious completion of the questionnaire. Using data grouping techniques (decision trees) and visualization analysis (Principal Component Analysis - for categorical data - Analysis of Correspondence), it is intended not only to qualify the data, but also to quantify the reliability.