Interpretable machine learning models for hypertensive disorders of pregnancy prediction

Abstract

Hypertensive disorders of pregnancy (HDP), including chronic hypertension, gestational hypertension, preeclampsia, and eclampsia, remain a major cause of maternal and neonatal morbidity and mortality. Although machine learning (ML) models show promise for HDP prediction, many high-performing models lack interpretability, limiting clinical trust and adoption. This thesis develops and evaluates interpretable ML models for HDP prediction using the 2022 National Vital Statistics System (NVSS) natality dataset. After preprocessing, the final dataset consisted of approximately 2.3 million births with 38 predictors including demographics, prenatal care, medical history, and socioeconomic factors. HDP was modeled as a highly imbalanced binary outcome (~ 9.8% positive). Falling Rule Lists (FRL) and Softly Falling Rule Lists (SoftFRL) were evaluated and compared with logistic regression, decision trees, random forests, support vector machines, and XGBoost using stratified 10-fold cross-validation with cost-sensitive learning. Due to the computational complexity of FRL on the full dataset, a graph coarsening strategy based on class-specific k-nearest neighbor graphs and algebraic multigrid was applied, reducing the training set to 9,205 representative observations. On the coarsened dataset, FRL achieved the highest sensitivity (0.762 ± 0.040) with competitive discrimination (AUC 0.671 ± 0.023) and balanced performance (G-Mean 0.638 ± 0.013), while SoftFRL showed comparable results. Statistical testing indicated no significant performance differences between FRL-based models and tuned baselines. The resulting rule lists revealed clinically meaningful risk patterns, supporting transparent and interpretable HDP risk stratification.