Rapid prediction of pulsatile hemodynamics in the aortic arch based on physics-informed neural networks

Background Computational Fluid Dynamics (CFD) has demonstrated significant value in the clinical assessment of cardiovascular diseases[1]. In contrast to CFD, which requires substantial computational time, Physics-Informed Neural Networks (PINNs) can rapidly deliver the predicted results. PINNs provide a novel approach for the rapid hemodynamic assessment of acute conditions such as ascending aortic dilation and aortic dissection[2]. Purpose This study aims to assess the precision and computational efficiency of PINNs for hemodynamic pressure reconstruction in patient-specific anatomical models. Methods In this study, a personalized three-dimensional anatomical model of the aortic arch was reconstructed based on clinical CT imaging data. Subsequently, a periodic pulsatile inlet condition[3] (period T=1s, Figure 1a) was applied, and the impedance characteristics of peripheral vessels were defined using a three-element Windkessel model. High-fidelity CFD simulations were performed to obtain benchmark hemodynamic data over the cardiac cycle, revealing flow variations in the aortic branches (Figure 1b). On this basis, a PINN architecture (Figure 1c) incorporating three-dimensional Navier-Stokes equations was constructed, and the network was trained using numerical simulation data to achieve rapid prediction of pressure field distribution. Results Figure 2 presents the pressure contours from both numerical simulations and PINNs predictions. With a baseline pressure of 110 mmHg, the PINNs predictions showed excellent agreement with CFD results in the ascending aorta (mean relative error of approximately 1%) and descending aorta (mean relative error of approximately 0.5%). However, localized deviations were observed in the Truncus, left common carotid artery (LCA), and left subclavian artery (LSA) during peak systole (PS) and mid-deceleration (MD), with a maximum relative error of 8.5%. The differences are primarily attributed to the simplified treatment of outlet boundary conditions and the complex dynamics of blood flow. Conclusion This study demonstrates the feasibility of using PINNs for rapid hemodynamic pressure reconstruction in the aortic arch. In the future, by integrating multimodal clinical data, PINNs are expected to evolve into hemodynamic assessment tools for diagnosing and planning surgeries for aortic diseases.