Diagnostic performance of deep learning on 12-leads electrocardiography for recurrence after pulmonary vein isolation in patients with persistent atrial fibrillation
Abstract Background Several patients with persistent atrial fibrillation (per-AF) suffer from recurrence after pulmonary vein isolation (PVI). Various methods to predict the recurrence were tried, but deep learning on 12-leads electrocardiography (ECG) after PVI was not studied. Purpose To elucidate diagnostic performance of deep learning on 12-leads ECG after PVI in patients with per-AF Methods We enrolled consecutive 109 patients with per-AF who underwent PVI (68.8±10.0 years, 83 males) excluding failure cases. We defined recurrence in 3–12 months after PVI. From the ECG just after PVI, five beats of each lead were sampled separately. Deep learning (convolutional neural network on bitmap ECG image) was performed by transfer learning of Inception-Resnet-V2 model. Gradient weighted class activation color mapping (GradCam) was performed to detect convolutional importance in the lead. Results Thirty-six patients showed recurrence in the period. Lead II (accuracy 0.701), aVR (0.690) were the top 2 leads of prediction, which showed larger accuracy than statistical accuracies of Non PV foci = SVC (accuracy = 0.541) and left atrial diameter >50mm (0.596). In lead II, GradCam spotlighted strong convolution of latter half of P wave in recurrent case, and former half of P wave and T wave in no-recurrent case. Conclusions Deep learning on ECG was a powerful tool to predict recurrence of per-AF after PVI. FUNDunding Acknowledgement Type of funding sources: None. Results of deep learning Results of GradCam