Machine learning identification of factors associated with exclusive breastfeeding in Brazilian infants: Cross-sectional analysis of the ENANI-2019 survey.

BACKGROUND: Exclusive breastfeeding (EBF) for the first six months is recommended by the World Health Organization, yet global rates remain suboptimal. In Brazil, EBF rates show significant regional disparities and socioeconomic gradients. This study aimed to identify factors associated with EBF status among Brazilian infants under six months using machine learning techniques. METHODS: This cross-sectional study utilized data from the National Study of Child Nutrition (ENANI-2019), a nationwide Brazilian household survey. Among 1960 infants under six months, EBF was defined according to WHO criteria as receiving only breast milk in the 24 hours preceding interview. Following systematic feature reduction and Elastic Net regularization, 49 features were selected from 118 available variables. Nine machine learning algorithms were compared using nested cross-validation, with Logistic Regression selected as the optimal approach due to clinical interpretability. RESULTS: The final model achieved AUC of 0.865 (95 % CI: 0.829-0.898) on independent validation data. Among 49 selected features, 19 showed statistically significant associations with EBF. Previous bottle exposure demonstrated the strongest negative association (coefficient = -2.355, 95 % CI [-2.586, -2.173]), followed by infant age in months and previous pacifier exposure. Protective factors included seeking breastfeeding information, breast milk donation, and higher birth weight. SHAP analysis validated feature importance rankings (correlation r = 0.968). Model performance varied across demographic subgroups (AUC range: 0.820-0.947). CONCLUSIONS: Previous bottle exposure emerged as the strongest factor associated with EBF discontinuation among Brazilian infants. The systematic comparison of machine learning algorithms and feature importance validation provides methodological insights for identifying factors associated with infant feeding practices. These findings may inform targeted interventions focusing on counseling regarding artificial nipple introduction and extending support during vulnerable periods, though prospective validation is needed before clinical application.