ECGI Periodicity Unraveled: A Deep Learning Approach for the Visualization of Periodic Spatiotemporal Patterns in Atrial Fibrillation Patients

This work proposes a novel deep learning based method for the identification of periodic patterns in Electrocardiographic Imaging (ECGI) signals and demonstrates its ability to identify, quantify, and visualize recurring patterns.
ECGIs from AF patients obtained prior to pulmonary vein isolation (PVI) are encoded to a lower-dimensional feature space using a 3D-CNN autoencoder, and further processed with principal component analysis to aggregate recurring patterns and quantify their contribution to the overall spatiotemporal propagation pattern.
Several markers are evaluated as potential predictors of AF recurrence. The variance captured by the first 3 principal components (PCs) varied from 19.8% to 59.2% (32.2±9.35) in different patients showed an inter-segment correlation exceeding 64%. Similarly, the number of PCs necessary to explain 90% of variance in ECGI recordings varied from 20 to 90 (56.2±20.1) demonstrating a varying number of propagation patterns across patients, which was reproducible intra-patient with an inter-segment correlation higher than 69%.
In addition, backpropagation-based saliency maps are computed to identify in which atrial regions the captured patterns occur. Saliency maps are visualized on 3D atrial models to aid in the interpretation and anatomical contextualization