Utilization of Tree-Based Machine Learning Models for Predicting Low Birth Weight Cases

Abstract Background: The monitoring of medical teams is crucial for assessing maternal and fetal health and early identification of complications such as Low Birth Weight (LBW), affecting over 20 million cases globally annually. This issue is associated with various risks, negatively impacting the development and quality of life of newborns. In Brazil, data from the Information System on Live Births (SINASC) indicate that 8% to 9% of newborns are diagnosed with LBW annually, emphasizing the need for comprehensive health policies for pregnant women. This study proposes the application of machine learning models to predict pregnant women at risk of having LBW newborns, aiming to contribute to prevention and adequate care. Methods: The methodology employed five models in four different scenarios, considering variations in attribute selection and experimenting with the removal or retention of duplicate data after preprocessing. Results: The results indicate that the models can identify newborns with Normal Birth Weight (NBW) but have limitations in detecting LBW cases, facing challenges in generalization, especially when tested with more recent data. The analysis of attribute importance highlights sociodemographic characteristics and gestational history as influential in the models’ learning. Conclusions: The study provides a comprehensive view of the capabilities and limitations of machine learning models in predicting LBW, with Adaboost and Gradient Boosting standing out as more robust in dynamic data scenarios.