The accuracy of Electrocardiographic Imaging (ECGI) is limited by the inverse problem's ill-posed nature, making it highly sensitive to residual noise from myogenic activity and electrode instability. While previous research addressed single-lead synthesis, realistic multi-lead residual noise generation for Body Surface Potential Mapping (BSPM) remains largely unexplored.
This study introduces EDEN (Electrocardiographic Deep Emulator of Noise), a Generative Adversarial Network (GAN) that learns the underlying patterns of real-world data to generate highly realistic synthetic samples. Noise was isolated from 373 clinical recordings using Principal Component Analysis (PCA). To maintain spatial information by considering the geometry of the patient's torso, signals were projected onto Laplace-Beltrami (LB) eigenfunctions. To ensure training stability and avoid issues such as mode collapse, a Wasserstein GAN with Gradient Penalty (WGAN-GP) was implemented to synthesize the LB coefficients. Performance was evaluated using Cosine Similarity and Wasserstein Distance.
EDEN successfully replicated clinical noise characteristics, achieving a Cosine Similarity of 0.57 ± 0.11 and a Wasserstein Distance of 2.61 ± 1.40, significantly outperforming Gaussian baselines (Cosine Similarity: 0.09 ± 0.02, Wasserstein Distance: 121.08 ± 2.10). Qualitative validation via Local Activation Time (LAT) maps confirmed that EDEN emulates realistic clinical distortions, whereas standard models produce unrealistic uniform blurring. This framework represents a relevant advancement in synthesis algorithms, providing a robust tool for validating ECGI solvers and training machine learning models without the privacy constraints associated with clinical data.