A Deep Learning Approach for Patient-Specific Aortic Wall Mechanical Stress Prediction in Transcatheter Aortic Valve Implantation Procedure Using Graph Neural Networks

Cardiovascular disease remains the leading cause of noncommunicable mortality globally, with aortic valve (AV) disorders, especially aortic stenosis (AS), representing a substantial burden. Transcatheter aortic valve implantation (TAVI) is an effective treatment for AS, where selecting the correct valve size is crucial and preimplantation mechanical assessment is desirable to predict possible complications that may occur post-TAVI. Finite-element analysis (FEA) is a powerful technique for predicting wall stresses, such as contact pressure and von Mises stresses. Predicting post-TAVI complications such as conduction abnormalities and paravalvular leakage is challenging. FEA is useful for that purpose, but it is technically demanding and time-consuming. This study introduces a deep learning (DL) approach to efficiently predict patient-specific aortic wall stresses for the TAVI procedure. We collaborated with Hamad Heart Hospital’s TAVI program, analyzing computed tomography (CT) data and implanted valve information from 55 patients. We conducted FEA for virtual implantation simulations to derive contact pressures and von Mises stresses. Using a graph neural network model that features SageConv and SplineConv layers within a Geometric PointNet architecture, we trained on data that included the actual implanted valve size and simulations with marginally smaller and larger sizes on the same patients to enhance training variability. Our model achieved a normalized mean squared error of 0.0020 for contact pressure and 0.0057 for von Mises stress predictions. Integrating the pre-TAVI aortic geometry with FEA results significantly improved prediction accuracy for contact pressure, though alignment of von Mises stress predictions was occasionally inconsistent.