Machine Learning Approaches for Computational Modeling of Bioprosthetic Heart Valves

Heart valve disease (HVD), a significant complication of cardiovascular disease, is one of the leading global causes of death. HVD can be treated using medical devices such as artificial valves, stents, and patches. However, designing and optimizing these devices requires a comprehensive understanding of the interactions of the device with the cardiovascular system. Experimentally studying these interactions is challenging, if not intractable. Computer simulation frameworks for fluid-structure interaction (FSI) have emerged as a valuable tool for studying the biomechanics and hemodynamics of the heart and its valves, complementing experiments. However, a significant research gap exists between methods developers and framework users due to expensive, proprietary, and inflexible existing software frameworks for FSI simulations. In this talk, I will discuss a new open-source, user-friendly, and expandable framework for heart valve FSI. Built on the FEniCS open-source computing platform, our framework can automatically convert scientific models into efficient finite element code. Additionally, it integrates state-of-the-art open-source libraries to extend FEniCS functionality to perform isogeometric analysis (IGA) and immersogeometric (IMGA) FSI simulations. The second challenge is the tedious nature of generating analysis-suitable bioprosthetic heart valve (BHV) geometries for simulations. Recent geometry-aware machine learning approaches can be adapted to create patient-specific valve geometry; still, they are limited to voxelized or triangular mesh geometry, which are not directly analysis-suitable. IGA uses Non-Uniform Rational B-splines (NURBS) for geometry and analysis. We have developed a differentiable NURBS module to integrate NURBS geometry representation with deep learning methods. In the second part of this talk, I will demonstrate the efficacy of our NURBS-Diff module in biomedical applications, such as fitting a NURBS surface to a point cloud data of an Aortic valve. The final challenge is the computational cost of performing several simulations to identify an optimal BHV design. To address this, we have developed a deep-learning framework to predict valve deformations, which provides the same accuracy as existing numerical approaches but is significantly faster. In addition, using our NURBS-Diff module, we can enforce different boundary constraints accurately to predict the deformed shape of valves. In the final part of my talk, I will demonstrate how this framework can be used to develop an optimal BHV design for a given patient. We envision these tools will address a critical gap in cardiovascular research by providing accessible, powerful tools for studying and treating HVD and have the potential to transform patient care through personalized, optimized heart valve treatments.