Estimating Respiratory Modulation in Atrial Fibrillation Using a Convolutional Neural Network

Abstract:
The quantification of autonomic nervous system (ANS) activity from ECG data may provide useful information for personalizing atrial fibrillation (AF) treatment, but is currently not possible. Since respiration is known to elicit an ANS response, the purpose of this study was to assess whether the respiratory modulation in AV nodal refractory period and conduction delay during AF can be estimated from ECG data. We trained a 1-dimensional convolutional neural network (1D-CNN) on synthetic data generated using a network model of the AV node to estimate the respiratory modulation in AV nodal conduction. The synthetic data replicates clinical ECG-derived data and contained 1-minute segments of RR series, respiration signals and atrial fibrillatory rates (AFR). Further, the synthetic data was generated using a total of 4 million unique parameter sets. We showed using synthetic data that the 1D-CNN can estimate the respiratory modulation from an RR series, respiration signal and AFR with a Pearson sample correlation of ρ = 0.855. The results of the present study suggest that the proposed method can be used to quantify respiratory-induced variations in ANS activity from ECG data. Further studies are needed to verify the estimates and to investigate the clinical value of the estimates.