Physics-Informed Emulation of Systemic Circulation for Fast Parameter Estimation and Uncertainty Quantification.

Area of Science:

• Computational fluid dynamics
• Biomedical engineering
• Machine learning in healthcare

Background:

• Computational models for blood flow are essential but computationally expensive for clinical applications requiring repeated simulations.
• Patient-specific parameter estimation and model calibration necessitate efficient simulation methods.

Purpose of the Study:

• To develop a physics-informed neural network framework for rapid, patient-specific blood flow and pressure prediction in vascular networks.
• To enable efficient parameter inference and inverse uncertainty quantification for clinical applications.

Main Methods:

• Utilized physics-informed neural networks (PINNs) as a surrogate modeling approach.
• Focused on patient-specific model calibration for vascular networks.
• Applied the framework to clinical data from patients with Double Outlet Right Ventricle (DORV).

Main Results:

• The trained machine learning model significantly reduces computation time compared to traditional numerical solvers.
• Achieved accurate predictions of flow and pressure waveforms.
• Demonstrated the framework's effectiveness in a comparative study against alternative machine learning methods.

Conclusions:

• Physics-informed neural networks offer a computationally efficient and accurate solution for patient-specific vascular modeling.
• The developed framework facilitates faster parameter inference and inverse uncertainty quantification in clinical settings.
• This approach holds promise for improved monitoring and management of congenital heart defects like DORV.