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

Area of Science:

• Computational mechanics
• Biomedical engineering
• Artificial intelligence in medicine

Background:

• Finite element analysis (FEA) is crucial for cardiac biomechanics but computationally intensive for creating patient-specific digital twins.
• Current FEA emulators struggle with diverse geometries, material models, and boundary conditions, limiting clinical applications.

Purpose of the Study:

• To develop a rapid and accurate FEA emulator for predicting passive biventricular myocardial displacements.
• To overcome the limitations of existing emulators in handling complex cardiac geometries and material properties.

Main Methods:

• Developed an attention-enhanced graph neural network (GNN) named HeartSimSage, inspired by Graph Sample and Aggregate (GraphSAGE).
• Integrated Laplace-Dirichlet solutions for spatial encoding and subset-based training for efficiency.
• Employed an attention mechanism to adaptively weigh neighbor contributions and filter information flow.

Main Results:

• HeartSimSage accurately predicted biventricular myocardial displacements with a median error of 0.280 mm (IQR [0.167, 0.484] mm) compared to traditional FEA.
• Achieved significant computational speedups: ~13000X on GPU and ~190X on CPU.
• Demonstrated robustness across diverse 3D biventricular geometries, mesh types, and material models.

Conclusions:

• HeartSimSage offers a computationally efficient and accurate alternative to traditional FEA for cardiac biomechanics.
• The model's ability to handle complex inputs and its speedup potential pave the way for clinical applications of cardiac digital twins.