Abstarct: Idiopathic ventricular arrhythmias (IVAs) are among the most common cardiac disorders, encompassing a spectrum from benign to life-threatening conditions. A considerable proportion of these arrhythmias originate from the cardiac outflow tract—a small and structurally complex region in which morphological similarities of electrocardiogram (ECG) signals make precise localization of the source challenging and increase the likelihood of errors in distinguishing between right and left origins. Since accurate localization of the site of premature ventricular contractions (PVCs)—corresponding to clinical source-baxsed classification—is the first and most critical step toward catheter ablation therapy, developing noninvasive, rapid, and accurate approaches for this purpose holds significant clinical and research value. In this thesis, using premature beats extracted from twelve-lead ECG recordings of 334 patients and employing one-dimensional deep learning architectures, novel approaches were proposed for discriminating the origins of arrhythmias, particularly within the highly prevalent RVOT and LVOT regions. The proposed models, in comparison with baxseline methods, demonstrated the superiority of one-dimensional convolutional neural networks in achieving effective spatial discriminability for the target regions. All implementations were validated using an inter-patient paradigm, and independent assessments were also conducted on an Iranian dataset. The optimization strategies introduced in this research—including the use of PVC-stream, multi-rate analyses, and extended vectorcardiography-baxsed approaches—significantly enhanced the performance of the proposed models compared with baxseline analyses. Accordingly, the CNN-BiLSTM model achieved an accuracy of 91.53% and an F1-score of 0.9445. Among the models implemented within the PVC-stream approach, the proposed CNN-Attention model achieved the best performance, with an accuracy of 92.16% and an F1-score of 0.9503. Moreover, the extended vectorcardiography-baxsed method, developed upon the proposed frxamework, yielded an accuracy of 93.09% and an F1-score of 0.9551, while effectively reducing the volume of data to be processed. These results represent notable improvements over the reference methods and underscore the significance of the proposed approaches. The findings indicate that deep learning–baxsed methods can noninvasively and accurately predict the origin of premature ventricular beats without the need for additional equipment.