Abstract / Summary
Hypertensive disorders of pregnancy arise from early placental and systemic vascular dysfunction, yet current first-trimester screening remains imprecise, biologically opaque and logistically challenging. Here we present Visionary AI, an interpretable platform that integrates retinal imaging with graph-based vascular modeling to predict hypertensive disorders before clinical onset. Visionary AI converts retinal images into topological and geometric representations of microvascular structure, enabling risk prediction from biologically grounded vascular features rather than generic image embeddings or clinical variables. In a prospective development cohort of 1,267 pregnancies, Visionary AI achieved strong performance for preeclampsia across population-wide control settings (area under the curve (AUC) = 0.91, average precision (AP) = 0.81). A stability-optimized model transferred without retraining to an independent external validation cohort, achieving AUC = 0.81 and AP = 0.68 and outperforming clinical and deep learning benchmarks. Recurrent features involving retinal vascular topology, geometry, network complexity and hierarchical organization position the maternal retina as a minimally invasive biosensor of systemic vascular health and Visionary AI as a biologically interpretable framework for pregnancy risk stratification. An interpretable artificial intelligence framework predicts hypertensive disorders of pregnancy using retinal images and vascular modeling.