A novel Dual-Branch Generative Adversarial Network for electrocardiogram data generation

Computer-assisted electrocardiogram (ECG) analysis is vital for the clinical diagnosis of cardiovascular diseases. However, the performance of many ECG classification methods is adversely affected by data imbalance. Generative Adversarial Networks (GANs) have recently emerged as a promising approach to address this challenge. Nevertheless, existing GAN models for ECG generation typically utilize a single generator structure, which limits their ability to generate complex ECG waveforms. To address this limitation, this paper proposes a novel dual-branch GAN model that integrates the strengths of Transformer and Long Short-Term Memory (LSTM) networks, along with self-attention mechanisms, to enhance the quality of ECG generation. The effectiveness of the proposed method was validated using the internationally recognized MIT-BIH Arrhythmia Database (MIT-BIH-AR) and the Chinese Cardiovascular Disease Database (CCDD). By incorporating the ECG data generated by the proposed model into the training set, the classification accuracy on the MIT-BIH-AR database for four diseases improved from 90.98 % to 96.66 %. Similarly, the classification accuracy for ventricular premature beats on the CCDD increased from 98.51 % to 99.34 %. The experimental results demonstrate that the proposed generative model can produce more realistic ECG data, thereby enhancing the performance of subsequent classification tasks.