Reproductive
Identification of preeclampsia subtypes via machine learning and associations with risk of perinatal and postpartum outcomes Hongyi Chen* Hongyi Chen Amanda L Ngo Lauren D Liao Yeyi Zhu
There has been increasing recognition of pre-eclampsia (PE) heterogeneity. We aimed to identify PE subtypes using routinely available clinical data via Uniform Manifold Approximation and Projection and Density-Based Spatial Clustering of Applications with Noise. We assessed associations of identified PE subtypes with risk of perinatal and postpartum outcomes via modified Poisson regression and Cox regression, adjusting for maternal age, race/ethnicity, parity, and pre-pregnancy body mass index. In a cohort study of 14,132 racially and ethnically diverse individuals with PE in Northern California in 2011-2021, we split the sample into one training (80% of 2011-2020) and two validation sets (1: 20% of 2011-2020; 2: 2021). In the training set, 4 clusters (C) were identified: C1 (68.3%), C2 (20.4%), C3 (8.7%), and C4 (2.6%). C1 (0.3%) and C2 (0.6%) had nearly no gestational diabetes mellitus (GDM), whereas all individuals in C3 and C4 had GDM. Compared to no gestational hypertension (GH) in C1 and C3, all individuals in C2 and C4 had GH. C1 (92.2%) and C3 (93.4%) were more likely to have late onset PE (≥ 34 gestational weeks), compared to C2 (88.3%) and C4 (89.9%). C3 and C4 were more likely to have pre-existing conditions (prediabetes, polycystic ovary syndrome, dyslipidemia and obesity), and to be older and multiparous, compared to C1 and C2. Compared to C1 as the mildest and most common subtype, C2 had higher risk of small for gestational age (RR 1.14; 95% CI: 1.04-1.25), C3 had higher risk of cesarean delivery (1.08; 1.00-1.16) and large for gestational age (1.49; 1.30-1.71), and C2, C3, and C4 had 1.23/1.11/1.36-fold higher risk of preterm birth, 1.14/1.11/1.23-fold higher risk of neonatal intensive care unit admission, and 1.60/1.26/1.62-fold higher risk of postpartum hypertension (all P <0.05). Findings were similar in the two validation sets. Subtyping PE may inform personalized risk assessment and management to improve associated maternal and child outcomes.