Predicting Low Birth Weight Using Machine Learning Models

The benefits of prenatal care are well-established, not least the relationship between prenatal care and better birth outcomes. Unsurprisingly, prenatal care is a cornerstone of modern public health policy. Prenatal care is associated with both reduced mortality and reduced morbidity risks. In particular, prenatal care can identify at-risk mothers and support interventions to reduce the incidence of low birth weights and associated adverse pregnancy outcomes. The objective of this work is to evaluate the performance of selected machine learning models in predicting whether a pregnancy is at risk of a low birth weight pregnancy outcome. A data set from the Brazilian Live Births Information System (SINASC) was used comprising data on pregnant women, prenatal care, and newborns. Three tree-based machine learning models were selected for evaluation using the main attributes found in the current literature. The Adaboost model presented the best metrics in the test dataset with an f1-score of 60.65\% and a sensitivity of 51.34\%; the attributes with the greatest impact on the prediction process were age, education, maternal occupation, and multiple gestations. Despite the models achieving relatively low metrics, this paper contributes to an important public health topic, namely the prediction of low birth weights, and the role that machine learning can play in improving support to obstetricians and assistance to pregnant women.