IMPACT OF THE COVID-19 PANDEMIC AND MACHINE LEARNING MODELS FOR PREDICTING PREMATURE BIRTHS IN THE CAPITALS OF THE NORTHEAST REGION OF BRAZIL, 2018-2021

Abstract

Premature birth is a global problem due to its implications for morbidity and mortality. It is one of the main risk factors for neonatal and infant mortality. Preterm delivery is defined as one whose pregnancy ends between the 20th and 37th weeks or between 140 and 257 days after the first day of the last menstrual period. For this study, data from the Information System on Live Births (SINASC) from the capitals of the Northeast region of Brazil, between 2018 and 2021, were used. of the performance metrics, compared to what was used for training and validation of the models. Six machine learning algorithms (Logistic Regression, Linear Discriminant Analysis, Multilayer Perceptron, AdaBoost, Decision Tree and Random Forest) were applied to predict prematurity. The models showed a drop in the Area Under the Roc Curve (AUC) metric in the years 2020 and 2021 compared to 2018 and 2019, with emphasis on the Adaboost, Random Forest and Decision Tree models, with drops greater than 10% attested by the statistical tests of Kruskal-Wallis and Nemenyi. As causes of the drop in performance of the models, it was identified that the variables month of beginning of prenatal care and age lost adherence in relation to the training base. The models showed good predictive performance, however, the use of tree-based models should be done with caution, since they are more unstable and that covid-19 had an impact on the distribution of the variables age and month of beginning of prenatal care. For training new models, pay attention to the input variables and the period used for training. For already established solutions, consider your retraining.

KEYWORDS: Depersonalization. Emotional Exhaustion. Professional Achievement. Predictors.