MultiQRSNet: a deep learning approach for high-accuracy localization of outflow tract ventricular premature contractions using 12-lead ecg

Background Idiopathic outflow tract Ventricular Premature Contractions (VPCs) are generally considered benign, most commonly originating from the Right Ventricular Outflow Tract (RVOT). While the success rate of catheter ablation is largely dependent on the VPC's site of origin, accurate pre-procedural localization remains a clinical challenge. Purpose This study aimed to develop a high-accuracy Deep Learning (DL) model using 12-lead surface ECGs to precisely determine the localization of outflow tract VPCs. Methods A total of 129 patients who underwent outflow tract VPC ablation between July 2020 and July 2025 were retrospectively included in the primary cohort. An External Validation Cohort of 50 patients from a different hospital was used to assess model generalizability. We developed a 1D Convolutional Neural Network (1D CNN) named MultiQRSNet, utilizing a 13-channel input (standard 12-lead ECG plus the normalized Maximum Deflection Index, MDI). The network architecture comprised five convolutional blocks and a three-layer fully connected block, culminating in a softmax layer for binary prediction of LVOT vs. RVOT(Fig 1a). To investigate the contribution of each lead and time segment to the predictive power, a Grad-CAM analysis was performed for each patient(Fig 2). Model interpretability was enhanced using SHapley Additive exPlanations (SHAP), and misclassification analysis was performed using UMAP-based dissimilarity. Traditional XGBoost was also evaluated for comparison. Results Baseline characteristics revealed that RVOT patients were younger and leaner, and had lower rates of hypertension and diabetes mellitus, compared to the LVOT and Intramural groups (Table 1). SHAP analysis indicated that V6 and V3 were most influential for LVOT predictions, while V2 and V4 dominated for RVOT predictions, consistent with anatomical localization(Fig 3). MultiQRSNet achieved a test accuracy of 98.22% in separating RVOT and LVOT labeled test sets, significantly outperforming the traditional XGBoost model (accuracy 89.64%). In the external validation cohort, the model demonstrated excellent generalizability, achieving an accuracy of 96.99% (p < 0.01) for the RVOT-LVOT distinction. The more challenging RVOT-LVOT-Intramural distinction achieved 60.73% accuracy (p < 0.01)(Table 2). UMAP projection confirmed the model's effectiveness by revealing two well-defined clusters corresponding to the LVOT and RVOT samples (Fig 1b). Conclusion(s) The developed MultiQRSNet DL model successfully identified the localization of outflow tract VPCs and intramural-origin VPCs with high accuracy. The model demonstrated strong performance not only in the primary cohort but also upon external validation, proving its generalizability across different hospitals and devices. This DL approach offers a powerful, non-invasive tool for pre-procedural planning.Table 1,2Figure 1,2,3