Abstract
BACKGROUND: Low birth weight (LBW) is a critical factor linked to neonatal morbidity and mortality. Early prediction is essential for timely interventions. This study aimed to develop and evaluate predictive models for LBW using machine learning algorithms, including Random Forest, XGBoost, Catboost, and LightGBM. METHODS: We analyzed data from 1,579 pregnant women enrolled in the Araraquara Cohort, a population-based longitudinal study. Predictor variables included maternal sociodemographic, clinical, and behavioral factors. Four ML algorithms Random Forest, XGBoost, CatBoost, and LightGBM, were trained using an 80/20 train-test split and 10-fold cross-validation. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied. Model performance was assessed using metrics such as area under the receiver operating characteristic curve (AUROC), F1-score, and precision-recall. Variable importance was evaluated using Shapley values. RESULTS: XGBoost demonstrated the best performance, achieving an AUROC of 0.94, followed by CatBoost (0.94), Random Forest (0.94), and LightGBM (0.94). Maternal gestational age was the most influential predictor, followed by marital status and prenatal care frequency. Behavioral factors, such as physical activity, also contributed to LBW risk. Shapley analysis provided interpretable insights into variable contributions, supporting the clinical applicability of the models. CONCLUSION: Machine learning, combined with SMOTE, proved to be an effective approach for predicting LBW. XGBoost stood out as the most accurate model, but Catboost and Random Forest also provided solid results. These models can be applied to identify high-risk pregnancies, improving perinatal outcomes through early interventions.