Inferring left atrial appendage (LAA) hemodynamics from 4D CT contrast dynamics by physics informed neural networks (PINNs)

Atrial fibrillation (AFib) is a common arrhythmia with a lifetime incidence of ~ 1 in 3 people that is associated with a 5X increase in ischemic stroke risk. During AFib, the atrial walls move weakly and irregularly causing blood stasis inside the LAA. Current clinical risk scores for stroke in AFib patients are based on demographic factors and have moderate accuracy. Here we present a computational tool for patient-specific clotting risk assessment based on 4D CT acquisitions of LAA contrast dynamics. We test PINNs with different underlying models (e.g., continuity only vs. Navier-Stokes and continuity), boundary conditions (e.g., temporally periodic flow vs. non-periodic flow), and input data with varying spatio-temporal resolutions. Our ground truth comprises CFD simulations in idealized, fixed-wall geometries as well as patient-specific, moving-wall left atrial meshes. We find that PINNs can accurately infer LAA hemodynamics using each patient's contrast agent concentration fields from CFD as training data with only a continuity underlying model, as long as temporal periodicity is imposed. Finally, we show proof of concept of the application of ROMs to infer LAA residence time using 4D CT data acquired in the clinical setting.