Graph Neural Networks to Predict Vascular Junction Pressure Losses for Reduced-Order Cardiovascular Modeling

  • Rubio, Natalia (Stanford University)
  • Pegolotti, Luca (Stanford University)
  • Pfaller, Martin (Stanford University)
  • Pham, Jonathan (Stanford University)
  • Darve, Eric (Stanford University)
  • Marsden, Alison (Stanford University)

Please login to view abstract download link

Reduced-order models (ROMs) are often used to approximate bulk quantities in patient-specific cardiovascular hemodynamic models due to their efficiency. Compared to expensive 3D computational fluid dynamics (CFD) simulations, they reduce computation by solving a smaller system of equations that governs a simpler but analogous physical system. In particular, the 0D ROM describes a vasculature as an electric circuit in which flow is represented by current and pressure is represented by voltage. These models can be run in seconds on inexpensive hardware (e.g., laptops). 0D models are widely used in practice but suffer from the inability to accurately characterize fluid dynamics in junctions. This work addresses this issue by using past CFD simulations to train a machine learning model to predict pressure drops over vascular junctions. We chose to develop a graph neural network (GNN) model to synthesize geometric and flow parameters for any number of inlets and outlets. During training, the GNN parameters are tuned to minimize the mean squared error with respect to the true pressure drop defined by the CFD simulations. We show that our model achieves lower errors than state-of-the-art options. Ultimately, we will incorporate this data-driven junction model in existing 0D frameworks to increase their accuracy. The training data for this work centers around the Vascular Model Repository, a publicly available collection of resolved hemodynamic 3D CFD simulations of diverse vasculatures. The complexity of this dataset poses challenges, particularly with respect to the generalization properties of our algorithm. These challenges are commonly encountered when applying machine learning techniques to real-world problems in which data are scarce and expensive to generate. This work proposes ways to overcome these issues by augmenting the dataset using synthetic data.