Predição de risco de mortalidade infantil em neonatos prematuros baseado em aprendizagem de máquina

Abstract

This study investigates the application of Machine Learning (ML) models to predict mortality risk in preterm neonates, one of the most significant challenges in contemporary neonatal medicine. The research explores how advanced ML techniques, such as Gradient Boosting, Neural Networks, Logistic Regression, and Naive Bayes, can develop accurate and reliable predictive systems to identify neonates at higher risk, aiming to improve interventions and reduce infant mortality. The central hypothesis suggests that each model has distinct potential in predicting these risks, evaluated based on performance, clinical integration, and contributions to public health policies. The methodology included a systematic literature review and the development of a custom dataset built from public data sources, SIM and SINASC. These data were refined and adjusted to eliminate inconsistencies and ensure their applicability in ML techniques, resulting in a targeted and consistent database. The analysis of the models indicated that, while all showed applicability, aspects such as data bias, interpretability, and clinical integration varied among them. It was concluded that the choice of the ideal model must balance accuracy, transparency, and clinical applicability, while also carefully considering ethical issues related to data privacy and informed consent.