Deep Learning Model for High-Accuracy Classification of Premature Ventricular Contractions With Precordial Transition Zones in Leads V3 or V4

Abstract

Background: Predicting the origin of premature ventricular contractions (PVCs) is challenging when a transition zone (TZ) appears in leads V3 and V4. The aim of this study was to develop a deep-learning model to predict PVC origins and identify electrocardiographic (ECG) features that contribute to the model’s decisions.

Methods and Results: ECG data from 314 patients with PVCs showing an inferior axis and TZ in leads V3 or V4 who underwent catheter ablation were analyzed. A convolutional neural network (CNN) was trained to predict an origin in the right or left ventricular outflow tract. Patients were divided into 3 cohorts for training, validation, and holdout (3 : 1 : 1 ratio). The CNN model was trained using paired data consisting of PVC and intrinsic QRS (iQRS). Five datasets per patient were used for training and validation; performance was evaluated using a single holdout dataset per patient. The CNN model achieved 92.1% accuracy, an F1 score of 0.91, and an area under the receiver operating characteristic curve of 0.96 on the holdout. Our model demonstrated superior diagnostic performance compared with conventional ECG indices. Gradient-weighted class activation mapping revealed that model attention was primarily focused on leads V3–V4 in iQRS, but was more diffusely distributed in PVC, notably the inferior limb leads and leads V2–V3.

Conclusions: The CNN-based prediction of PVC origin demonstrated clinical utility.