Aplicação da aprendizagem supervisionada na determinação do factores associados a perdas gestacionais

Abstract

Pregnancy loss occurs when a pregnancy, for various reasons, does not result in the birth of a live baby, representing an experience of significant physical and emotional impact on the woman. This study aimed to identify the factors associated with pregnancy loss in Mozambique through the application of supervised learning techniques, namely logistic regression and Árvore de decisãos. A total of 2,668 observations from the 2023 Demographic and Health Survey (DHS) were analy- zed. Eleven variables were subjected to descriptive analysis, multicollinearity diagnosis, missing data treatment, response variable balancing using the ROSE (Random Over-Sampling Examples) method, and predictive model construction. Advanced maternal age (35–49 years) emerged as the most significant risk factor for pregnancy loss, with odds ratios ranging from 11.9 to 14.4. The province of residence also showed a strong association: women living in Maputo City had a 10.5 times higher probability of pregnancy loss compared to those living in Niassa Province. Age at first sexual intercourse also proved influential: women who initiated sexual activity later had a lower risk of pregnancy loss (OR = 0.87). Regarding model performance, the Árvore de decisão outperformed logistic regression in discriminative ability, with an F1-score of 0.504 versus 0.465 and an accuracy of 69.8% versus 64.7%. Logistic regression showed higher sensitivity (81.7%) compared to the tree (74.5%), while the tree obtained better accuracy (38.1% versus 32.5%). Both models showed similar balanced accuracy: 74.6% for the tree and 71.6% for logistic regression. These results suggest a slight advantage of the tree in identifying patterns and interactions between relevant variables