High‐Fidelity Atrial Fibrillation Electrocardiogram Signal Generation Based on Mamba‐WGAN‐GP

ABSTRACT Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia, characterized by highly irregular temporal dynamics and complex pathological morphologies in electrocardiogram (ECG) signals. The development of robust, intelligent diagnostic systems for AF is critically hampered by the scarcity of high‐quality, annotated clinical data. To address this data bottleneck, we propose a novel generative model, Mamba‐WGAN‐GP, which integrates the Mamba state space model with a Wasserstein generative adversarial network incorporating a gradient penalty (WGAN‐GP). This model is specifically designed to capture the long‐range temporal dependencies inherent in AF ECGs. By leveraging Mamba's selective state mechanism and linear‐time complexity, our model efficiently captures the defining characteristics of AF, including irregular R–R intervals and complex f‐wave morphology. We conducted extensive experiments on the MIT‐BIH Arrhythmia Database, demonstrating that Mamba‐WGAN‐GP achieves state‐of‐the‐art performance with a peak signal‐to‐noise ratio (PSNR) of 20.07 dB, a dynamic time warping (DTW) distance of 0.10, and a Fréchet distance of 0.56. Compared to a baseline CNN‐based WGAN‐GP, our model improved PSNR by 5.6%, while reducing DTW and Fréchet distances by 80.4% and 13.8%, respectively. The synthesized AF signals exhibit high clinical fidelity, with an R–R interval coefficient of variation () and f‐wave spectral characteristics (4–9 Hz) that are statistically indistinguishable from real signals. Mamba‐WGAN‐GP provides a powerful and reliable data augmentation solution, serves as a robust data augmentation tool and research methodology to facilitate the training of AF diagnostic algorithms in data‐scarce scenarios.