Phenotyping Preeclampsia Using Unsupervised Machine Learning: A Prospective Cohort Study

OBJECTIVE: To explore clinically meaningful phenotypes of preeclampsia using unsupervised machine learning. DESIGN: Prospective cohort study. SETTING: BCNatal, a tertiary maternal-foetal medicine centre (Barcelona, Spain). POPULATION: A total of 482 pregnant women diagnosed with preeclampsia between August 2013 and April 2024. METHODS: Maternal demographic, clinical, ultrasound and laboratory data were prospectively collected, together with delivery details and maternal and neonatal complications. We generated a patient representation using maternal age, height, weight, body-mass index, blood pressure, angiogenic factors, urinary albumin/creatinine ratio, gestational age at birth and birthweight centile. Dimensionality reduction was performed using Uniform Manifold Approximation and Projection, followed by k-means clustering to identify phenotypes. MAIN OUTCOME MEASURES: Maternal and neonatal characteristics and complication rates were compared across clusters to evaluate the clinical significance of the data-driven phenotypes. RESULTS: Three phenotypes were identified. Cluster A (n = 223; 46.2%) showed earlier delivery (mean 33.1 ± SD 3.3 weeks), marked angiogenic imbalance, prevalent foetal growth restriction (64%) and the highest rates of maternal (23%) and neonatal (41%) complications. Cluster B (n = 147; 30.5%) had later delivery (37.0 ± 2.1 weeks), moderate angiogenic imbalance and intermediate birthweight centiles (15.5 ± 9.7). Cluster C (n = 112; 23.2%) comprised mostly term cases (38.2 ± 1.5 weeks) with the lowest angiogenic imbalance and the highest birthweight centile (67.9 ± 27.3); obesity (31%) and diabetes (15%) were most prevalent and maternal (4%) and neonatal (9%) complications were least frequent. CONCLUSIONS: Unsupervised learning delineated three preeclampsia phenotypes. These phenotypes could support the need for future risk stratification and more personalised management; prospective external validation is warranted.