Machine learning for predicting preterm birth: A cross-database analysis on Pernambuco, Brazil cases

Preterm birth remains a major public health challenge globally and is the leading cause of death in children under five. This study investigates the predictive capacity of machine learning models for preterm birth using two health datasets from Pernambuco, Brazil: the national-level SINASC and the state-level SIS-MC. The research evaluates the models’ performance under different data balancing techniques—random undersampling and hybrid sampling—and examines the influence of threshold adjustment to optimize clinical decision-making metrics. A cross-database approach was employed to assess the performance of the models across distinct data collection frameworks, considering variations in attribute availability, data completeness, and contextual relevance. Five tree-based classifiers (Decision Tree, Random Forest, XGBoost, CatBoost, and LightGBM) were trained and evaluated using a standardized preprocessing pipeline and hyperparameter optimization. Findings indicate that although models generally achieved higher accuracy for term births, sensitivity for preterm cases remained limited, especially when trained with imbalanced data. The undersampling strategy, when coupled with threshold tuning based on ROC curve analysis, resulted in the most favorable trade-off between sensitivity and positive predictive value (PPV). Cross-database evaluation revealed performance degradation when models trained on one dataset were applied to another, highlighting the influence of data heterogeneity and the importance of local context in model development. This study underscores the necessity for attribute harmonization and transfer learning strategies to improve model adaptability across diverse healthcare settings.