HeartSimSage: Attention-Enhanced Graph Neural Networks for Accelerating Cardiac Mechanics Modeling

Finite element analysis (FEA) is a powerful tool that forms the cornerstone of modeling cardiac biomechanics. However, FEA is computationally expensive for creating digital twins, which typically involves performing tens or hundreds of FEA simulations to estimate tissue parameters, limiting its clinical application. We have developed an attention-enhanced graph neural network (GNN)-based FEA emulator, HeartSimSage, to rapidly predict passive biventricular myocardial displacements from patient-specific geometries, chamber pressures, and material properties. Designed to overcome the limitations of current emulators, HeartSimSage can effectively handle diverse three-dimensional (3D) biventricular geometries, mesh topologies, fiber directions, structurally based constitutive models, and physiological boundary conditions. It also supports flexible mesh structures, allowing variable node count, ordering, and element connectivity. To optimize information propagation, we developed a neighboring connection strategy inspired by Graph Sample and Aggregate (GraphSAGE) that prioritizes local node interactions while maintaining mid-to-long-range dependencies. Additionally, we integrated Laplace-Dirichlet solutions to enhance spatial encoding and employed subset-based training to improve computational efficiency. Incorporating the attention mechanism allows HeartSimSage to adaptively weigh neighbor contributions and filter out irrelevant information flow, enhancing prediction accuracy. As a result, errors in the predicted biventricular myocardial displacements by HeartSimSage were limited to a median of 0.280 mm with an interquartile range of [0.167, 0.484] mm compared to traditional FEA, while achieving a computational speedup of approximately 13000X on a GPU and 190X on a CPU. We validated our model on a published left ventricle dataset and analyzed the model's sensitivity to hyperparameters, neighboring connection strategies, and the attention mechanism.