A machine learning method for real-time numerical simulations of cardiac electromechanics

We propose a machine learning-based method to build system of differential equations that approximates the dynamics 3D electromechanical models for human heart, accounting dependence on set parameters. Specifically, our permits create reduced-order model (ROM), written as Ordinary Differential Equations (ODEs) wherein forcing term, given by right-hand side, consists an Artificial Neural Network (ANN), possibly depends parameters associated with be surrogated. This is non-intrusive, it only requires collection pressure and volume transients obtained from full-order (FOM) cardiac electromechanics. Once trained, ANN-based ROM can coupled hemodynamic blood circulation external in same manner original model, but at dramatically lower computational cost. Indeed, allows real-time numerical simulations function. demonstrate effectiveness proposed two relevant contexts modeling. First, we employ perform global sensitivity analysis both models. Second, Bayesian estimation starting noisy measurements scalar outputs. In these cases, replacing FOM electromechanics makes possible few hours time all would otherwise unaffordable, because their overwhelming cost, if carried out FOM. As matter fact, able speedup more than three orders magnitude.