Comparação de métodos de aprendizado de máquina na predição de acidemia fetal

Abstract

Background: Fetal acidemia, characterized by an umbilical artery pH lower than 7.1, is the main cause of intrapartum asphyxia. The umbilical artery pH indicates comorbidities during the delivery as well as in the child's development. There is a need for non-invasive methods to access this parameter to help with decision-making. Objective: To compare machine learning models to predict fetal acidemia in a university hospital in southern Brazil. Study design: The study reviewed records from 567 patients with single pregnancies who gave birth in the Hospital Geral de Caxias do Sul between 2011 and 2016. Statistical analysis was performed, and the following machine learning algorithms were developed using Python: Extra Trees Classifier, Random Forest Classifier, Support Vector Machine, Nonlinear Support Vector Machine with RBF Kernel, Artificial Neural Networks, Gradient Boosting Machine, and Logistic Regression. We also used the GridSearchCV function to find the best parameters of each model for optimization. Results: Group I presents 397 newborns with an umbilical artery pH higher than 7.1 while Group II contains 170 newborns with an umbilical artery pH lower than 7.1. The statistical analysis indicated a significant difference between Group I and Group II in the following variables: parity, previous stillbirth, gestational age, diabetes, fetal presentation, type of delivery, and Apgar in the first and fifth minute. The model with the highest AUROC after optimization was Artificial Neural Networks (0.82), followed by the Logistic Regression (0.81), Support Vector Machine (0.80), Gradient Boosting Machine (0.78), Random Forest Classifier (0.78), Extra Trees Classifier (0.76), and Nonlinear Support Vector Machine with RBF Kernel (0.71). The Artificial Neural Networks and Logistic Regression models had outstanding precision (0.78 vs 0.79), recall (0.95 vs 0.95), F1-score (0.86 vs 0.87), and accuracy (0.79 vs 0.80). Conclusion: Machine learning models can be used to assist healthcare professionals in decision-making. Among the algorithms proposed, the Artificial Neural Networks and Logistic Regression presented the best performance in identifying both groups to predict fetal acidemia; Even though those models have the best accuracy, we recommend using Logistic Regression because it requires less computational capability. This study represents a non-invasive method to identify the umbilical artery pH. [resumo fornecido pelo autor]