BACKGROUND: Low birth weight (LBW) is a health condition that affects over 20 million gestational outcomes worldwide. The current literature indicates that machine learning models have the potential to assist healthcare professionals in predicting LBW and giving them the opportunity to intervene earlier in the pregnancy, which might include adjusting medical treatments or suggesting changes in diet. PURPOSE: This study proposes the evaluation of machine learning models to predict which pregnant women are at risk of neonatal outcomes with LBW. METHODS: The methodology involves six phases, including data analysis and attribute selection through different techniques, which generated four distinct scenarios. We used five machine learning models and validated them through cross-validation and hyper-parameter optimization and evaluated their performance considering seven distinct metrics and statistical analysis, focusing on the effectiveness of the models in predicting LBW. RESULTS: The results revealed that the models achieved varying levels of performance across the scenarios, with the removal of duplicate data resulting in improvements in recall (0.83) and f1-score (0.64). Statistical analysis confirmed significant differences (p <
0.05) among most models. CONCLUSIONS: The conclusions of this study indicate that the removal of duplicate data and careful attribute selection positively influenced the performance of the machine learning models in predicting low birth weight. Additionally, the analysis of attribute importance highlighted socio-demographic characteristics and gestational history as the most influential in the training of the models.