Avaliação de modelos de aprendizado de máquina para classificação de gestantes e predição de gravidez de risco usando o histórico de consultas médicas

Abstract

Today, modern technology allows the creation and storage of ever-increasing amounts of data. Furthermore, the wide availability of machine learning methods aids in the analysis and interpretation of the large volume of data available, which brings clear benefits to society. Health systems, notoriously, generate a lot of data. There are important sensitive events, and that the analysis of information from the volume of stored data can contribute to early diagnoses. For example, pregnant women tend to be followed for a long period, which generates important information for decision-making. Therefore, the identification of a high-risk pregnancy or even the early detection of possible complications can guarantee the pregnant woman a pregnancy with more care and comfort. Thus, in this work we aim to evaluate machine learning models for classifying pregnant women at risk. Initially, we characterize the care provided to pregnant women, where the trajectories of care for patients in the basic health network are analyzed. Assessing the trajectory of SUS patients, whether between care units or service units, is an important activity for understanding and improving the system. Thus, we apply classification models to identify pregnant women at risk. Furthermore, based on existing data, we created techniques to automatically group pregnant women based on medical appointments. Both the classification model and the application of this model in temporal data had their effectiveness tested in a group of pregnant women registered in the public health system in the city of São Paulo-Brazil. The results demonstrate that the classification model has high performance, and that is able to classify a pregnancy at risk, with an accuracy of almost 74%, with only 4 weeks of pregnancy data. The groupings generated by the automatic model were validated by specialists in the area, and the information from these groups is consistent with the procedures performed by the pregnant women. Finally, we believe that the models created and the discussions in this dissertation can be used as a way to assist in decision making, both for physicians and health managers.