Patient-specific hemodynamic simulation for left ventricular assist device via non-invasive monitoring based on lumped parameter model and hierarchical neural network

Introduction: Left ventricular assist devices (LVADs) are widely used in advanced heart failure, but require accurate hemodynamic assessment for optimal management. Current invasive methods such as right-heart catheterisation (RHC) are limited in routine use, highlighting the need for non-invasive alternatives. Methods: A non-invasive framework combining a lumped parameter model (LPM) with a hierarchical neural network (CLPM-Net) was developed to estimate patient-specific hemodynamic parameters from echocardiography and blood pressure. Model identifiability analysis was performed to select key parameters. The model was trained on synthetic data and validated with clinical cases. Results: The proposed method achieved accurate parameter estimation with errors below 10% (RMSE). Simulated hemodynamic indicators showed strong agreement with ground truth (nMED < 1%). Clinical validation demonstrated close consistency with invasive measurements. Discussion: This framework enables non-invasive, patient-specific hemodynamic assessment for LVAD management. It shows potential as an alternative to invasive monitoring, though further large-scale clinical validation is required.