PREDIÇÃO DE NASCIMENTO PRÉ-TERMO EXTREMO EM DUAS CIDADES BRASILEIRAS

Abstract

Introduction: Extreme preterm birth (EPTB) is a major public health problem associated with high neonatal and infant morbidity and mortality. Its occurrence is influenced by multiple clinical, sociodemographic, and behavioral factors, making prevention particularly challenging. Although relatively uncommon, EPTB has substantial consequences for maternal and child health. In this context, machine learning (ML) techniques offer a promising approach for identifying pregnant women at increased risk of EPTB. Objective: To develop and evaluate machine learning-based predictive models to identify predictors of extreme preterm birth among live births from two Brazilian cities. Methods: This analytical cross-sectional study was nested within the 2010 BRISA prenatal and birth cohorts conducted in São Luís, Maranhão, and Ribeirão Preto, São Paulo, Brazil. The inclusion of these cohorts allowed the investigation of maternal and child health in distinct socioeconomic and demographic settings. The study included singleton pregnancies receiving care in public and private healthcare services, comprising 6,414 mother–child pairs from São Luís and 8,273 from Ribeirão Preto, totaling 14,687 observations. A sensitivity analysis was performed using the variable threatened preterm labor during the current pregnancy, which represents ongoing pathophysiological changes potentially related to the onset of the outcome and is considered an important predictor. Six analytical scenarios were developed, corresponding to the combined cohort and the São Luís and Ribeirão Preto cohorts, each analyzed with and without this variable. Based on expert knowledge and previous literature, 119 variables common to both cohorts, 121 variables in São Luís, and 119 variables in Ribeirão Preto were selected as potential predictors, including socioeconomic, demographic, behavioral, reproductive, and maternal healthcare characteristics. The BORUTA feature selection algorithm identified 50 relevant variables, from which the eight most important predictors were retained. Fourteen machine learning algorithms were evaluated in each scenario: penalized logistic regression (PLR), support vector machines (svmLinear and svmRadial), random forest (rf, ranger, and Rborist), boosting methods (glmboost and AdaBoost.M1), nearest neighbor methods (Knn e KKnn), artificial neural networks (nnet), tree-based partitioning models (cforest and ctree2) and generalized additive models (GAM). Model performance was assessed using accuracy, sensitivity, specificity, precision, F1-score, and the area under the receiver operating characteristic curve (AUC-ROC). Results: All analytical scenarios showed class imbalance between EPTB and non-EPTB births. Model performance varied across the six scenarios. When threatened preterm labor was included, kNN showed the best overall performance in the combined cohort (sensitivity = 0.774, AUC-ROC = 0.806, F1-score = 0.049), svmRadial performed best in São Luís (sensitivity = 0.653, AUC-ROC = 0.732, F1-score = 0.063), and ranger achieved the highest performance in Ribeirão Preto (AUC-ROC = 0.860, sensitivity = 0.771, F1-score = 0.053). When this variable was excluded, nnet performed best in the combined cohort (sensitivity = 0.561, AUC-ROC = 0.648, F1-score = 0.019), GAM in São Luís (sensitivity = 0.623, AUCROC = 0.587, F1-score = 0.024), and Rborist in Ribeirão Preto (AUC-ROC = 0.629, sensitivity = 0.573, F1-score = 0.016). Overall, kNN and nnet demonstrated the best predictive performance in scenarios with and without the threatened preterm labor variable, respectively. The eight most important predictors were threatened preterm labor during the current pregnancy, threatened miscarriage during the current pregnancy, maternal occupation, per capita income, alcohol consumption during pregnancy, self-reported maternal race/skin color, planned pregnancy, and hypertension diagnosed before or during pregnancy. Conclusion: Machine learning models can predict extreme preterm birth with good predictive performance using maternal health, fetal, socioeconomic, behavioral, and prenatal care variables. Early identification of pregnancies at increased risk may support timely interventions and improve maternal and neonatal outcomes. These findings provide evidence for the potential application of machine learning in prenatal risk stratification and encourage further validation in different populations and clinical settings.