RDiffGAN-ECG: High-quality specific R-peaks conditional ECG generation using denoising diffusion generative adversarial network with state space models

The electrocardiogram (ECG) artificial intelligence faces challenges in collecting ECG data, imbalanced training data, and data privacy issues. Deep generative models can generate high-quality synthetic ECGs to alleviate these challenges, but current models often require extensive conditional information, which limits their effectiveness. This study introduces RDiffGAN-ECG, a personalized R-peaks conditional denoising diffusion GAN that integrates structured state space sequence model (S4) and selective state space model (Mamba) within its generator. The generator of RDiffGAN-ECG uses S4 to capture the temporal dynamics across diffusion time steps and R-peaks information for ECG reconstruction. Mamba integrates diagnostic annotations with S4 features, enhancing the generation of high-quality 12-lead ECGs with accurate R-peak position and diagnostic relevance. This model is trained to generate 12-lead ECGs for five classes diagnoses with specific R-peaks information. RDiffGAN-ECG was compared with several existing generative models through qualitative and quantitative analysis, and heart rate analysis. On the PTB-XL test set, it achieved better quantitative results (RMSE = 0.2245, FD = 5.6795, DTW = 68.1109) than existing models. On the CPSC2018 dataset, it demonstrated superior generative performance for different R-peaks, showing its generalization capability, with the results of RMSE = 0.2293, FD = 4.3552, and DTW = 52.2489. Ablation studies confirm the importance of the GAN framework and the combined use of S4 and Mamba in enhancing ECG generation quality. In summary, the RDiffGAN-ECG exhibits significant advantages in improving the quality of the generated ECG, and achieving specific R-peaks ECG generation. It provides substantial value for ECG-related research and clinical applications.