AI in Hypertensive Disorders of Pregnancy: Review

Abstract Background Hypertensive disorders of pregnancy (HDP) are a leading cause of maternal and fetal mortality worldwide. Early detection and risk stratification are critical for timely intervention to prevent severe complications such as eclampsia, stroke, and preterm delivery. However, traditional clinical methods often lack the precision needed to identify high-risk individuals effectively. Machine learning (ML) has emerged as a powerful tool, leveraging complex data to enhance prediction, diagnosis, and clinical decision-making in HDP. This review aims to systematically evaluate ML applications in HDP, highlighting trends, methodologies, and gaps to guide future research and improve maternal and fetal outcomes. Methods This study adheres to the PRISMA-ScR guidelines for scoping reviews, focusing on full-text, English-language publications that apply ML models to HDP. A comprehensive search across three databases captured studies involving at-risk patient populations. Data extraction followed the CHARMS checklist, summarizing study characteristics, outcomes, and ML methodologies, while also identifying gaps and opportunities for further research. Results Most studies targeted preeclampsia (n=70, 75.27%), with limited focus on other HDP phenotypes such as gestational hypertension (n=4, 4.3%) and postpartum hypertension (n=1, 1.07%). Sample sizes ranged from 20 to over 700,000 participants. Studies have been increasing since 2014 emphasizing diagnosis/onset detection (n=58, 62.37%) and risk prediction (n=26, 27.95%). Random Forest, Logistic Regression, Decision Trees, and SVM were the most common ML methods. Geographic analysis revealed concentration in China (n=29, 31.18%) and North America (n=18, 19.35%), with underrepresentation in other regions. Input data predominantly comprised demographics (n=50, 53.76%), patient/family history (n=43, 46.24%), and functional tests (n=43, 46.24%), whereas omics (n=29, 31.18%) and imaging data (n=2, 2.15%) were infrequently used. Outcomes related to time-to-intervenes and readmission were each reported once. Conclusions Machine learning is increasingly applied to HDP, with significant growth in diagnostic and risk prediction models. However, geographic disparities, limited phenotype representation, and models to help intervene at critical time points throughout the perinatal lifecycle remain barriers. Notably, models addressing time-to-intervene predictions and hospital readmissions are underrepresented, highlighting critical gaps in the current literature. Addressing these limitations—by developing models to help improve the timing of medical interventions, higher risk profiling, and diverse datasets—can advance ML's role in improving maternal and fetal outcomes and reducing mortality globally. Future research should focus on refining ML models to support clinicians and advance care for patients with HDP.