Publication

Abstract : Electrocardiogram (ECG) signals are essential for diagnosing and monitoring cardiac conditions. However, the scarcity of diverse and high-quality ECG data and the presence of class imbalance pose significant challenges for developing robust diagnostic models. Existing generative approaches for ECG signal synthesis often fail to capture the essential domain-specific features, such as PQRST waveforms and long-range temporal dependencies, which are clinically significant and vital for diagnostic accuracy. To address these limitations, this study presents a Transformer-based Generative Adversarial Network (TransGAN) approach that incorporates a hierarchical multi-scale patch discriminator and a PQRST-constrained generation module for realistic and clinically valid ECG signal synthesis. The generator integrates a temporal self-attention mechanism to capture long-range dependencies, while the PQRST feature extraction module ensures the preservation of clinically significant morphological characteristics, thereby improving the physiological relevance of the generated signals. Unlike prior GAN-based biomedical synthesis models, the proposed architecture is specifically adapted to ECG’s three-level temporal hierarchy (beat, rhythm, sequence), enabling domain-aware morphological and rhythmic fidelity. Furthermore, the multi-scale patch discriminator ensures the local consistency of the ECG signals by classifying overlapping segments at multiple temporal resolutions, preserving fine-grained waveform details. Moreover, to address data imbalance, the framework uses a weighted Binary Cross-Entropy loss (WBCE). Extensive experiments demonstrate that the proposed work outperforms the state-of-the-art methods, achieving Mean Squared Error (MSE) of 0.042, dynamic time warping (DTW) distance of 2.29, Wasserstein distance (WD) of 0.120, and PQRST feature accuracy of 94.1%. Notably, the generated ECG signals improve downstream classification accuracy by up to 7.2% with data augmentation. The results confirm that the synthesized ECG signals exhibit high clinical fidelity, temporal coherence, and enhanced diagnostic relevance, highlighting the effectiveness of the proposed domain-specific integration for ECG synthesis.