Prediction of indicators through machine learning and anomaly detection : a case study in the supplementary health system in Brazil

Abstract

The research aimed to investigate the stages of a MachineLearning model process creation inordertopredict the indicator over the number of medical appointments per day done in the areaof supplementary health inthe region ofPorto Alegre /RS - Brazil and to propose a metric for anomalies detection. Literature reviewand applied case study wasusedas a methodology inthis paper,besides wasused the statistical software calledR, in order toprepare the data and create the model. Thestages ofthecase study was: database extraction, division of the base in training andtesting, creation of functions and feature engineering,variables selection and correlationanalysis, choiceof the algorithms with cross-validationand tuning, training of models, application of the models in the test data, selection of the best model and proposal of the metric for anomalies detection. At the end of these stages, it was possible to select the best modelin terms ofMAE (MeanAbsolute Error), the Random Forest, which was the algorithm withbetter performance when compared to Linear Regression and Neural Network. It also makes possible to identified nine anomaly points and thirty-eight warning points using the standard deviation metric. It was concluded, through the proposed methodology and the results obtained, that the steps of feature engineering and variables selection were essential for the creation and selection of the model, in addition, the proposed metric achieved the objective of generates alerts in the indicator, showing cases with possible problems or opportunities.