Geometry-Aware Abdominal Aortic Aneurysm Digital Twin for Patient-Specific Wall Stress Mapping

Accurate estimation of wall stress distributions in abdominal aortic aneurysms (AAAs) is critical for improving rupture risk prediction beyond the traditional maximum diameter criterion. While high-fidelity finite element analysis (FEA) provides precise patient-specific wall stress estimates, its computational cost limits its use in real-time clinical decision making. We introduce a graph-based deep learning framework that leverages three types of graph neural networks (GNNs), namely a Gated Graph Convolutional Network (GGCN), an Equivariant Graph Neural Network (EGNN), and a Graph Transformer (GT) to rapidly predict FEA-derived wall stress distributions directly from AAA outer wall surface meshes. Computed tomography angiography (CTA) images from 202 AAA patients treated at three clinical centers, were segmented using an in-house U-Net pipeline and refined manually to generate high quality AAA volume and surface meshes. From these, several node-specific geometric and biomechanical features were recorded, including wall thickness (WallTHK), intraluminal thrombus thickness (ILTTHK), wall strength $\left({\sigma }_{\text{ult}}\right)$ , the principal minor and major curvatures ( $k_1$ and $k_2$ ) and their derived curvatures (Gaussian $K$ and mean $M$ ), distance to the lumen centerline ( $r_L$ ), normalized local diameter (NORD), and the ILTTHK-to-WallTHK ratio ( ${\rho }^{\text{ILT}}$ ). Local information was incorporated via neighborhood averaged features from the six nearest nodes. All GNNs achieved high fidelity in reproducing patient-specific FEA wall stress fields, with the GT providing the best overall agreement, while the GGCN and EGNN also delivered strong performances. The best performing GT captured spatial stress patterns, preserved biomechanical trends, and enabled near real-time inference. It reproduced FEA-derived wall stress with high accuracy (node-specific $R^2>0.97$ , graph-level $R^2>0.98$ ) while reducing per-case inference time from hours for a complete FEA simulation to a few seconds on a single GPU. With further refinement and larger cohorts, this graph-based framework may strengthen rupture risk assessment and facilitate routine, patient-specific biomechanical analyses in clinical workflows.