Study of PINN sensor layout in evaluating WSS with application to patient-specific carotid flow

Atherosclerosis, a chronic inflammatory disease characterized by lipid-rich plaque buildup within arterial walls, is a major contributor to stroke and heart attack. It is strongly influenced by blood flow characteristics, such as wall shear stress (WSS). Traditional methods like computational fluid dynamics (CFD) and magnetic resonance imaging (MRI) struggle to accurately capture near-wall flow due to the complex and uncertain nature of blood dynamics. This study investigates whether sensor density and location affect the accuracy of Physics-Informed Neural Networks (PINNs) in modeling blood flow in carotid artery bifurcations. We hypothesized that: (a) increasing the number of sensors would improve model accuracy due to increased quantities of probed data, and (b) sensors placed closer to the artery wall would more accurately predict WSS, better capturing the flow gradient within the boundary layer. Using idealized bifurcation models, we tested various sensor configurations. Higher sensor densities improved accuracy to a certain extent, with no significantly positive impact beyond N_sample = 100, and sensors placed at 10% of the radius from the wall produced balanced predictions for both the near-wall flow and the general flow patterns. To test real-world applicability, we applied the optimized PINN model to a patient-specific carotid geometry and achieved an average velocity error of only 0.0037 m/s. These results successfully demonstrate how sensor configuration significantly affects PINN accuracy and highlight PINNs’ potential as an efficient tool for personalized diagnosis of cardiovascular disease.