Identifying high-risk pregnancies in rural areas with machine-manifold learning

Abstract

Maternal mortality remains unacceptably high, particularly in developing countries, where 99% of cases occur, most of which are preventable. In Guatemala, disparities in maternal healthcare quality between urban and rural areas exacerbate this issue. The Healthy Pregnancy project, initiated by the EHAS Foundation, sought to address this gap by equipping rural nurses with prenatal care kits, enabling them to perform screenings comparable to those in urban settings. This study presents a retrospective secondary analysis of data collected during the project (2014–2016), encompassing 10,108 cases and 108 features. Using Data Science and machine learning (ML) methodologies, we compared Support Vector Machines (SVM) and Random Forest (RF) algorithms to classify pregnancies requiring referral to higher-level care, where the referral decision was defined based on the retrospective assessment of obstetric specialists using clinical, laboratory, and ultrasound data. Their performance was evaluated against the benchmark sensitivity (0.62) and specificity (0.98) of trained nurses. SHAP (Shapley Additive exPlanations) values were employed to interpret model predictions and identify the most critical features influencing classification. Furthermore, manifold learning techniques, specifically Uniform Manifold Approximation and Projection (UMAP), were utilized to uncover latent structures within the data, offering additional interpretability via SHAP analysis. Our results show that both models (RF and SVM) achieve sensitivity and specificity values comparable to those obtained by trained nurses when techniques such as SMOTE (Synthetic Minority Oversampling Technique) and cost-sensitive learning are applied to address the imbalanced dataset. UMAP and SHAP analyses revealed the most globally relevant features. These findings demonstrate the potential of ML-driven approaches to support clinical decision-making in resource-limited settings, enhancing the detection of high-risk pregnancies, reducing training demands, and facilitating the monitoring of prenatal checkups in such contexts.