Scanned ECG-Based Deep Learning for Localization of Premature Ventricular Complexes Origins

Background: Premature ventricular complexes (PVCs) frequently originate from the right or left ventricular outflow tract (RVOT or LVOT). Accurate localization is essential for guiding catheter ablation but remains challenging due to morphological overlap on surface Electrocardiogram(ECG). Objective: To develop and evaluate an image-based deep learning framework for automated localization of PVC origins using scanned 12-lead ECGs. Methods: Scanned ECGs from 59 patients with confirmed LVOT or RVOT PVCs were preprocessed and augmented to improve quality and generalizability. Two models, a custom convolutional neural network (CNN) and a transfer learning-based MobileNetV2, were trained to classify PVC origin. Performance was assessed using fivefold cross-validation and metrics including accuracy, sensitivity, specificity, and area under the ROC curve (AUC). Grad-CAM was applied for interpretability. Results: MobileNetV2 outperformed CNN across all metrics, achieving an accuracy of 0.74, sensitivity of 0.78, specificity of 0.70, and AUC of 0.84. Grad-CAM visualizations confirmed the model's attention to clinically relevant features of QRS morphology, particularly precordial leads. Conclusion: This study demonstrates the feasibility of image-based deep learning for PVC localization using paper-based ECGs. MobileNetV2 offers a promising, interpretable solution for real-world deployment, especially in resource-limited settings where digital ECG signals are unavailable.