Multi-Noise Removal from ECG Signals Using Residual Block-Based Generative Adversarial Networks with Customized Loss Function

The field of biomedical signal processing has witnessed significant advancements, particularly in the domain of electrocardiogram (ECG) denoising. Existing methods often lack consideration of both local and global feature differences, have limited applicability to various noise types, and can introduce signal distortion after denoising. In our research, we propose and investigate a multi-noise removal model based on Generative Adversarial Networks (GANs), specifically using residual networks. This enables the generator to learn the distribution of ECG noise through adversarial training against a stacked-layer discriminator, thereby addressing persistent challenges in ECG noise reduction—crucial for accurate diagnosis of cardiac conditions. Extensive experiments were conducted on both intra- and inter-patient paradigms using an ECG arrhythmia dataset from the PhysioNet Database, ensuring a comprehensive evaluation of the model’s performance. While many studies—including those evaluated on the intra-patient setting, which tends to overestimate results—report high accuracy, our method demonstrated superior generalization and reliability in the challenging inter-patient paradigm. Specifically, our approach outperformed recent techniques, achieving the highest Signal-to-Noise Ratio (SNR) average of 36.21 and the lowest Root Mean Square Error (RMSE) scores across most SNR input levels. In addition to these standard metrics, our approach demonstrated a maximum improvement in Approximate Entropy (ApEn) of 0.0307 relative to the original clean signal under the most challenging noise condition. This improvement underscores the model’s ability to enhance signal quality after denoising, with promising generalization to new patients. These results highlight the effectiveness of our GAN-based approach for real-world ECG denoising and its potential to advance medical diagnostics.