Defining refractoriness in single atrial beats using autoencoder neural networks. (19th May 2022)

• Record Type: Journal Article
• Authors: Ruiperez-Campillo, S; Deb, B; Feng, R; Ganesan, P; Clopton, P; Rogers, AJ; Narayan, SM
• Is Part Of: Europace. Volume 24:Supplement 1(2022)
• Publication Date: 2022-05-19
• Subjects: Arrhythmia -- Treatment -- Periodicals; Cardiac pacing -- Periodicals; Catheter ablation -- Periodicals; Heart -- Physiology -- Periodicals; Electrophysiology -- Periodicals; 617.4120645
• DOI: 10.1093/europace/euac053.614
• Languages: English
• ISSNs: 1099-5129

Abstract

Abstract: Funding Acknowledgements: Type of funding sources: Public grant(s) – National budget only. Main funding source(s): NIH Background: Mapping atrial fibrillation (AF) is complicated by signals which may be local or far-field, but which cannot currently be separated. This could be clarified by a knowledge of atrial refractory periods, yet these are difficult to define from monophasic action potentials (MAP) in patients. We hypothesized that transfer learning using an autoencoder neural network (ANN), first trained with less-noisy ventricular signals, can be applied to de-noise and classify atrial MAPs. Methods: We first developed an ANN to encode MAPs in 5706 ventricular MAPs from N=42 patients (age 65±13y) during pacing (fig1. A-B). This created a latent feature space. We now tuned the ANN to classify atrial MAPs in a different cohort of patients with AF. We used a statistical loss function based on mathematical optimization to evaluate the accuracy of final representations of the MAP and classify the different signals. Results: The autoencoder ANN reconstructed ventricular MAPs with an average similarity of 0.85 (range 0-1) (an example is shown in fig 1.B). We tested on 3000 atrial MAPs in AF patients (N=21; 67±5y, 13 women). Atrial MAPs were accurately represented (fig 1.E-F) with similarity indices that were higher than those obtained by a panel of 3 experts. Fig. 1 shows the reconstruction of different signals: ventricular MAP (fig 1.A-B), ventricular MAP with pacing artifact (fig. 1.C-D), atrial MAP (transfer learning is assumed in here; fig 1.E-F) and noise or signals with morphologies of no interest (fig 1.G-H). Fig. 2 shows the classification of signals according to the similarity metric that allows distinguishing among the different types of signals without manual annotation (p<0.05 between groups). Conclusion: Atrial refractory periods can be defined in single beats in AF patients using autoencoder neural networks and transfer learning. This approach can separate atrial beats from far-field ventricular beats and other sources of noise. Future work can study if this approach can be used to improve AF mapping or define novel physiological phenotypes.