Intracranial aneurysms (IAs) are life-threatening vascular pathologies requiring accurate risk assessment to guide clinical interventions. Hemodynamic biomarkers, such as wall shear stress and oscillatory shear index, have shown potential in predicting rupture risk. However, their use in clinical routine remains challenging due to the computationally intensive nature of traditional computational fluid dynamics (CFD) methods required for their estimation. In this work, we propose a novel physics-constrained graph neural network (GNN) framework trained on high-fidelity CFD data for predicting full 3D, time-resolved hemodynamic fields throughout the cardiac cycle. To improve accuracy in capturing complex spatio-temporal dynamics, the framework incorporates node features enhancement into the graph representation, enabling precise modeling of intricate hemodynamic variations. The proposed model achieves near-real-time predictions, maintaining high accuracy while drastically reducing computational costs. Additionally, we introduce a dataset of 105 patient-specific aneurysm geometries with corresponding CFD simulations, designed to benchmark Machine Learning (ML) models in aneurysm hemodynamics. This work represents the first application of GNNs to time-dependent 3D flow field predictions in aneurysms, offering a transformative step toward real-time, AI-driven clinical decision support for rupture risk stratification and treatment planning.
publication under review