Accelerated Patient-Specific Hemodynamic Simulations with Hybrid Physics-Based Neural Surrogates

1. Problem Statement

Cardiovascular hemodynamic simulations are critical for clinical decision-making, surgical planning, and medical device design. However, there is a trade-off between accuracy and computational cost:

• 3D Finite Element Models (FEM): Highly accurate, resolving detailed velocity and pressure fields, but computationally expensive (requiring hours to days on high-performance clusters). They are unsuitable for real-time applications or many-query tasks like optimization and uncertainty quantification.
• 0D Reduced-Order Models (ROMs): Computationally lightweight (solving in fractions of a second on personal laptops) by treating the vasculature as a network of lumped parameters (resistors, inductors). However, standard 0D models rely on simplified Poiseuille flow assumptions and idealized junction dynamics. This leads to significant accuracy errors, particularly in complex anatomies with extensive branching (e.g., pulmonary arteries) or geometric irregularities (e.g., stenosis), where flow deviates from idealized pipe flow.

The Goal: Develop a hybrid framework that retains the low computational cost and interpretability of 0D models while achieving the accuracy of high-fidelity 3D simulations by learning optimal parameters from 3D data.

2. Methodology

The authors propose a Hybrid Physics-Based Neural Surrogate framework. Instead of deriving 0D parameters (resistance $R$, inductance $L$) from analytical Poiseuille formulas, they use Neural Networks (NNs) to predict these parameters directly from vessel and junction geometry.

A. The Hybrid Model Formulation

The framework utilizes a Resistor-Resistor-Inductor (RRI) model (or a simplified Resistor-Inductor RI model) for both vessels and junctions:

• Standard 0D: Uses fixed formulas for $R_{lin}$, $R_{quad}$, and $L$ based on vessel radius, length, and viscosity.
• Hybrid 0D: Replaces these fixed formulas with outputs from neural networks ($R_{lin}, R_{quad}, L = f(\text{Geometry})$).
• Training Target: The "ground truth" for training is not the 3D simulation data directly, but the optimal 0D parameters ($\Theta_{opt}$) that, when used in a forward 0D simulation, minimize the error against the high-fidelity 3D solution. These optimal parameters are found using a calibration tool (svZeroDCalibrator).

B. Geometric Preprocessing

To make the 0D discretization compatible with the hybrid approach, two critical preprocessing steps were introduced:

1. Bifurcation Splitting: Standard 0D discretization often creates junctions with many outlets (e.g., 10+ in pulmonary arteries). The NN requires a fixed-size input vector. The authors developed an algorithm to split $N$-outlet junctions into a series of binary bifurcations connected by artificial "connector vessels" of zero length.
2. Entrance Length Adjustment: Standard discretization places junction boundaries too close to the split site, where 3D effects (flow separation, recirculation) create "noise" in pressure drop measurements. The authors extend the junction boundaries into the outlet vessels by a pseudo entrance length ($L_e = 10 \times r_{outlet}$). This ensures the flow is fully developed at the boundary, providing cleaner data for the NN to learn bulk hemodynamic behavior.

C. Neural Network Architecture & Training

• Inputs: Geometric features extracted from centerlines (e.g., inlet/outlet radii, tortuosity, length ratios, flow split ratios).
• Outputs: $R_{lin}$, $R_{quad}$, and $L$.
• Loss Function: A standard Mean Squared Error (MSE) was compared against a Proximity-Weighted Loss. The latter weights elements closer to the inlet (proximal) more heavily, correcting for the over-representation of distal, less influential elements in highly branched networks (like pulmonary trees).
• Architecture: Multi-layer perceptrons (MLPs) with 2–4 hidden layers.

3. Key Contributions

1. Hybrid Framework: Successfully integrated data-driven parameter prediction into physics-based 0D models, bridging the gap between speed and accuracy.
2. Novel Preprocessing: Introduced bifurcation splitting and entrance length adjustment to handle complex, patient-specific geometries and reduce noise in training data.
3. Loss Function Innovation: Demonstrated that proximity-weighted loss significantly improves performance in extensively branched anatomies by prioritizing hemodynamically critical proximal regions.
4. Model Simplification Discovery: Found that excluding the quadratic resistor term (using the RI model instead of RRI) actually improved convergence and accuracy in this specific context, likely because the NN struggled to disentangle linear and quadratic effects with limited data.

4. Results

The framework was validated on three patient-specific cohorts: Aortic, Aortofemoral, and Pulmonary (totaling 32 geometries). Performance was measured using Maximum Percent Error (MPE) of inlet pressure compared to 3D ground truth.

• Accuracy Improvement:
  • Aortic: Error reduced from 5.5% (Baseline) to 2.3%.
  • Aortofemoral: Error reduced from 8.1% to 3.5%.
  • Pulmonary: Error reduced dramatically from 30.4% to 7.2%.
  • Overall: Across all cohorts, learned parameters reduced error by at least 50%.
• Design Choice Impact:
  • RI vs. RRI: The simpler RI model (excluding $R_{quad}$) outperformed the full RRI model, which suffered from convergence issues in complex pulmonary cases.
  • Preprocessing: Entrance length adjustment and proximity-weighted loss were crucial for the pulmonary cohort but had minimal effect on simpler aortic geometries.
• Computational Efficiency: The hybrid model runs in <2 seconds on a personal laptop, enabling real-time simulation.

5. Significance

This work presents a viable path toward real-time, interpretable digital twins for cardiovascular medicine.

• Clinical Utility: By achieving 3D-level accuracy with 0D speed, clinicians can perform rapid "what-if" scenarios for surgical planning or device sizing without waiting days for 3D simulations.
• Scientific Insight: It demonstrates that machine learning can effectively "inject" high-fidelity physics insights into simplified models without sacrificing the interpretability required for clinical trust.
• Scalability: The framework is generalizable to various vascular anatomies and can be integrated into existing open-source tools (SimVascular), facilitating broader adoption in cardiovascular research and engineering.

In summary, the paper proves that learning optimal parameters from 3D data is a superior strategy to relying on analytical assumptions for 0D modeling, particularly for complex vascular networks, effectively solving the accuracy-speed trade-off in hemodynamic simulation.