Rapid prediction of prosthetic heart valve haemodynamic performance using physics-informed machine learning

点击翻译按钮获取中文摘要

The project aims to create rapid and scalable deep learning-based simulation techniques for predicting the haemodynamic performance of transcatheter aortic valve (TAV) implants. TAVs are the de facto treatment for severe aortic valve stenosis in patients with med-/high-surgical risk. Yet, various clinical and technological challenges prevent uptake of the procedure in a wider spectrum of patients. Computational modelling tools can allow performance of TAVs in so-far untested scenarios (low-risk groups, bicuspid anatomy, etc.) to be assessed non-invasively and safely. Such capabilities underpin the emerging concept of in-silico trials (ISTs). Complex interactions of blood flow and valves necessitate use of fluid-structure interaction models for haemodynamic assessment. FSI models are expensive, and, like all computational techniques, can encounter convergence problems especially in complex scenarios. This is problematic for ISTs, where large simulation cohorts are required to run quickly and automatically. Recent research demonstrates the feasibility of accelerating computational multiphysics via learning-based approaches. So-called physics-informed neural networks (PINNs) are particularly attractive as they guarantee predictions conform to physical laws, rather than proceeding from data observations only. They may also alleviate training data requirements by imposing strong regularisation. The project will build on our existing efforts to develop distinct structural- and flow-based PINN models. The aim is to produce efficient and automated FSI simulation tools that will enable TAV haemodynamic performance to be predicted in simulation studies involving potentially thousands of virtual patients.