Deep learning formulation of ECGI evaluated on clinical data

Abstract

Background: Electrocardiographic Imaging (ECGI) is an exceptional resource in cardiology practice and research, allowing for non-invasive assessment of local cardiac electrical activities, through the acquisition of ECGs signals acquired with multi-electrodes vests. This approach is largely based on solving an ill-posed inverse problem. However, to date, there is no method sufficiently convincing to solve the inverse problem, to establish ECGI as the clinical modality of choice. Previously, we proposed a deep learning (DL) based method for ECGI reconstruction by exploiting multimodal information and prior knowledge from previous cases. Tested with synthetic data, the method proved to be effective and convincing, but clinical validation is lacking. Purpose: To propose a model allowing for fast generation of activation maps from combined Body Surface Potentials (BSP) and imaging data acquired in patients. Methods: The entire ECGI problem was reformulated with a probabilistic generative model, the Conditional Variational AutoEncoder (CVAE). During training, the model learned the conditional probability of the activation map based on conductivity maps and BSP of synthetic data. The model was constructed with convolutional layers, which allowed to extract the spatio-temporal correlations from the multimodal conditional data. To construct training and evaluation datasets, we proposed the following data preparation pipeline. First, the clinical data from 7 patients was processed, then the geometrical modelling of the torso and the heart was done. Fast personalized cardiac simulation using an eikonal model was applied to build the training dataset. Then the ground truth (GT) data was transferred to create the evaluation dataset. In addition, BSPs were processed to fit the simulated data onto the GT. The model was trained on the simulated data. To evaluate our method, we compared the GT activation map, captured by the CARTO system, with the reconstructed output of the trained model from the registered BSPs. Results: Our model was trained on 5 different cardiac anatomies, with 4500 generated activation maps and BSPs. It was evaluated on two additional patients. The method allowed for the generation of activation maps from BSP signals in less than 1s. Comparison of our generated activation maps with the GT showed a good identification of the cardiac activation pattern (mean distance between generated and GT 32+/-8.3 ms). Validation from 7 activation onsets indicated a quite accurate onset localization (11.5+/- 8.6 mm). Conclusion: Our DL ECGI solver allows for fast generation of activation maps from clinically-acquired BSPs. By probabilistically learning the spatio-temporal correlations and integrating anatomical information, the method goes beyond the current state-of-the-art, with comparison to invasive contact mapping data demonstrating accurate identification of activation patterns and a precise localization of the onset of focal activations. (Figure Presented).