Advanced ECG signal denoising using deep learning architectures: convolutional neural networks, autoencoders, and generative adversarial networks

ECG signal quality important for achieving high level of cardiac diagnose; however real ECG records are often affected by noise – baseline wander, powerline interferences, muscles noise. This paper describes an improved deep learning based solution for denoising of ECG signal employing the CNN, Autoencoder, and GAN models. In particular, the noise suppression is added at the last stage with each model trying to preserve relevant morphological features of the ECGs. The system is tested and trained using the Massachusetts Institute of Technology-Beth Israel hospital Arrhythmia Database in the presence of added noise to reflect real life. Based on the fusion approach, the outcome of the three models depends on a basic averaging method so that the benefits of all the three models can be harnessed. Quantitative performance evaluation is based on Signal-to-Noise Ratio where the proposed ensemble model shows better SNR of 17.3 dB than individual models where Autoencoder shows SNR 16.1 dB, CNN has SNR 15.4dB and GAN has SNR 14.7dB . This clearly shows that the selection of ensemble method fulfills the objectives of improving robustness and fidelity of the reconstructed ECG signal. This hybrid approach suggests the way in realtime, noise-resistant ECG monitoring and establishes the basis for developing new methods of using deep learning-based denoising for multi-channel and clinical ECG systems.