Conditional GAN-Based ECG Signal Denoising With Skin-Electrode Impedance Modeling

ECG signals from dry electrodes, commonly used in wearable devices, are vulnerable to noise due to high skin-electrode impedance. In this work, we propose a denoising method based on conditional generative adversarial networks, where the noisy signal is treated as a condition for the generation of a clean ECG signal. To develop deep learning models with high generalization capability, it is important to ensure diversity in the training dataset. However, most existing denoising methods rely on ECG data collected by wet-contact electrodes. To remedy this, we adopt an input-referred voltage noise model at the circuit level of the measurement device, incorporating three types of internal noise: amplifier, thermal, and flicker noise. In addition, we include baseline wander, muscle artifact, and electrode motion noise to construct a comprehensive noise set. This allows us to synthesize a dataset with diversity in both high- and low- frequency components. The proposed method employs adversarial loss framework where the discriminator network determines whether the denoised signal follows the distribution of the clean ECG signals. Meanwhile, the generator network minimizes additional two loss components, namely $L_{1}$ and an identity type loss, whose role is to preserve noise-free segments during denoisng process. We observe robust denoising results in all the examples in the results section. In one example, we propose a new classification method derived as a byproduct of our work. By leveraging both the noisy and denoised signals, the proposed method achieves improved classification accuracy for detecting abnormal ECG signals.