Due to the difficulty of collecting electrocardiogram (ECG) data during emergency situations, ECG data generation is an efficient solution for dealing with highly imbalanced ECG training datasets. However, due to the complex dynamics of ECG signals, the synthesis of such signals is a challenging task. In this paper, we present a novel approach for ECG signal generation based on Generative Adversarial Networks (GANs). Our approach combines GANs with statistical ECG data modeling to leverage prior knowledge about ECG dynamics in the generation process. To validate the proposed approach, we present experiments using ECG signals from the MIT-BIH arrhythmia database. The obtained results show the benefits of modeling temporal and amplitude variations of ECG signals as 2-D shapes in generating realistic signals and also improving the performance of state-of-the-art arrhythmia classification baselines.