Abstract 11178: Non-Invasive Identification of Fibrotic Atrial Cardiomyopathy with Neural Networks Based on 642,400 12-Lead Ecgs from an Extensive Cohort Simulation Study

Abstract

The clinical picture of fibrotic atrial cardiomyopathy is caused by a partial replacement of atrial myocardial tissue with fibrotic substrate. Being subject to electrical and structural remodeling processes, the substrate provides the necessary properties for the maintenance of atrial fibrillation and could thus be a good risk marker. We hypothesize that the P waves in the 12-lead ECG present a non-invasive and cost-effective tool to quantify the extent of atrial fibrosis as an alternative to state-of-the-art approaches. Our virtual patient cohort comprised a combination of 80 atrial, 25 thoracic geometries and 27 rotation angle variations of the atria. For each atrial model, 10 stages of fibrosis were defined by replacing different fractions of the myocardial volume with fibrotic tissue patches exhibiting remodeled ionic and conduction velocity properties. Conducting electrophysiological simulations in sinus rhythm on each model combination with one to three different healthy baseline conduction velocities yielded 642,400 synthetic ECGs. P wave duration, dispersion, peak-to-peak amplitudes and terminal force in V1 were calculated and served, together with atrial volumes and torso sizes, as an input for a regression neural network to estimate the volume fraction of atrial fibrosis. The root mean square error between the ground truth and the estimated extent of fibrosis was 8.74% left atrial fibrotic volume. Based on the regression results, a distinction between the healthy and the diseased cohorts was possible with an accuracy of 99.01%, a sensitivity of 99.57% and a specificity of 62.38%. In conclusion, we have gauged the potential of the 12-lead ECG for estimating fibrotic atrial volume fractions in an extensive cohort simulation study comprising both, functional and anatomical inter-subject variability. As a next step, the methods need to be validated with clinical data for which robust feature extraction algorithms are required.