Identifying Preeclampsia Subtypes Using Machine Learning for Cardiovascular Health Insights

Background: Preeclampsia is a leading cause of maternal-fetal morbidity and strongly linked to future cardiovascular disease. Prediction and prevention remain limited, in part due to preeclampsia heterogeneity and poorly defined biological subtypes. Better delineation of preeclampsia subtypes could inform strategies to reduce preeclampsia and its short- and long-term morbidity. We aimed to identify novel clinical and biomarker defined subtypes of preeclampsia using unsupervised machine learning and evaluate their associations with postpartum cardiovascular health. Methods: Participants with preeclampsia, defined per ACOG criteria, were identified from the multicenter, prospective Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be (NuMoM2b)-Heart Health Study. Participants with pre-pregnancy hypertension were excluded. Clinical and demographic data, placental and cardiovascular biomarkers, and proteomic profiles were analyzed. Two-stage feature selection (Boruta SHAP and volcano filtering) identified informative proteomic markers, integrated with clinical variables. UMAP dimensionality reduction and Gaussian Mixture Model clustering (optimized by silhouette scores) identified subtypes, followed by refinement to maximize between-cluster variance. The resulting subgroups represented phenotypically and biologically distinct preeclampsia endotypes. Postpartum incident hypertension (>=130/80 mmHg or medication use at 2-7 years) was compared across subtypes using modified Poisson regression. Results: Among 298 women with preeclampsia, three subtypes emerged (Figure 1): (1) cardiometabolic – higher body mass index, blood pressure, and maternal inflammatory markers; (2) placentally-mediated – elevated placental analytes and normal maternal cardiometabolic features; and (3) mixed – intermediate cardiometabolic and placental analyte profiles. All subtypes showed high risk of incident hypertension versus women without adverse pregnancy outcomes (n=755). Risk was highest in the cardiometabolic subtype (RR 3.09 [95%CI 2.39-3.97]), followed by placentally mediated (RR 1.97 [1.43-2.73]) and mixed (RR 1.62 [0.98-2.67]) (global Wald test p<0.001; Figure 2). Conclusion: Preeclampsia comprises biologically distinct subtypes with heightened but differential postpartum cardiovascular risk. Recognition of this heterogeneity provides new insights into pathophysiology and may advance precision prevention of preeclampsia and its long-term cardiovascular sequelae.