Methodological Innovations in Predictive Modeling of Birth Outcomes Using Brazilian Vital Statistics Data (2012–2024)

This study presents methodological innovations in applying machine learning to Brazilian vital statistics for predicting adverse neonatal outcomes using data from the Sistema de Informações sobre Nascidos Vivos (SINASC, 2012–2024). The dataset, comprising about 67.8 million live births, enables large-scale population analysis through harmonized and reproducible workflows. Two supervised learning algorithms — Logistic Regression and XGBoost — were implemented to predict three outcomes: low birth weight, prematurity, and their combined occurrence. The modeling design strictly prevents target leakage by using gestational age and birth weight to define dependent variables. The feature matrix includes 22 predictors across six conceptual domains covering maternal sociodemographic characteristics, reproductive history, and prenatal care. Data completeness analyses confirmed high reliability, with most variables presenting less than 5% missingness. Both models were evaluated using stratified 5-fold cross-validation and ROC-AUC and F1 metrics. Results indicate comparable predictive performance, with Logistic Regression achieving slightly higher weighted F1 scores and XGBoost showing marginally better macro-level balance. These findings highlight the complementarity between interpretable and ensemble methods in addressing class imbalance in population-based prediction. The reproducible pipeline, fully implemented in Python and DuckDB, demonstrates a scalable and transparent framework for demographic and epidemiological modeling with administrative microdata. Overall, the research contributes to methodological advances for data-driven monitoring of maternal and child health and provides transferable tools for large-scale public health surveillance.