Combining Biophysical Modeling and Machine Learning to Predict Location of Atrial Ectopic Triggers

Eduardo Godoy1, Miguel Lozano1, Ignacio Garcia-Fernandez1, Rafael Sebastian2
1Universitat de Valencia, 2CoMMLab, University of Valencia


Abstract

The localization of ectopic triggers in the atria requires that an electrophysiologist maps the electrical activity in the endocardium to infer the area where tissue should be ablated to eliminate an arrhythmia. When triggers are located in non-standard regions, or only generate a few ectopic activations, it can be complex and time consuming its localization. This study aims to show a methodology to estimate the location of ectopic foci by means of a machine learning system trained with body surface potential maps (BSPMs) generated by means of a detailed biophysical model of the atria and torso run for hundreds of scenarios. A total of 500 simulations in 26 atria configurations (i.e., different distributions of fibrosis) were carried out to obtain the BSPMs. Courtemanche model was used for the healthy tissue and MacCannell for fibrotic tissue. BSPMs were clustered and labelled using a hierarchical clustering method, and a support vector machine was trained with BSPMs and labels. Finally, a cross-validation was performed to analyse the accuracy of the method to predict the label of a BSPM, which is related to the location of the corresponding ectopic focus on the atria. Focal activity was triggered from 20 different atrial locations, and BSPMs were collected using a number of electrodes that ranged from 2 to 256. The ectopic foci were grouped in the atria into regions (between K=2 and K=10). The results from the cross-validation showed that 32 electrodes were enough to predict the region of an ectopic focus (90% accuracy) in patients with little amount of fibrosis (< 5%). As fibrosis increased, the system required at least 64 electrodes to reach the same accuracy. However, in cases with fibrosis over 15% the labels assigned to the atrial regions overlapped, which blurred the association between BSPM and specific atrial regions, hampering its utility.