Learning reduced-order models for cardiovascular simulations with graph neural networks.

Area of Science:

• Cardiovascular physiology and computational fluid dynamics.
• Application of artificial intelligence in medical modeling.
• Hemodynamics and biomechanical engineering.

Background:

• Physics-based reduced-order models are efficient for cardiovascular modeling but struggle with complex anatomies.
• Existing models may lack accuracy in cases with numerous vessel junctions or pathological conditions.
• Need for accurate and efficient simulation of blood flow dynamics in diverse physiological scenarios.

Purpose of the Study:

• To develop a novel one-dimensional reduced-order model for blood flow simulation using graph neural networks.
• To enhance the accuracy and generalizability of cardiovascular models, particularly for complex geometries.
• To achieve high efficiency during inference time for hemodynamic simulations.

Main Methods:

• Development of a graph neural network (GNN) trained on three-dimensional (3D) hemodynamic simulation data.
• The GNN iteratively predicts pressure and flow rate at vessel centerline nodes based on initial conditions.
• Validation of the GNN model on diverse physiological geometries and boundary conditions.

Main Results:

• The GNN-based model demonstrates high accuracy and generalizability across various anatomies.
• Achieved prediction errors below 3% for both pressure and flow rate with sufficient training data.
• Outperformed traditional physics-based one-dimensional models in accuracy while maintaining computational efficiency.

Conclusions:

• The proposed graph neural network approach offers a superior alternative to conventional physics-based models for cardiovascular blood flow simulation.
• This method effectively handles complex anatomies and pathological conditions with high accuracy and efficiency.
• The findings highlight the potential of AI in advancing hemodynamic modeling and cardiovascular research.