ECG Signal Denoising Algorithm Based On Generative Adversarial Network And Performance Analysis

The electrocardiogram(ECG)is a frequent tool of diagnosing and detecting cardiac disease in both health monitoring and telemedicine,where abnormal changes in the ECG signals can be analyzed to determine the type and location of the disease.However,ECG signals are extremely susceptible to various types of noise during acquisition,which causes some important pathological signal features to be frequently interfered with by noise,resulting in reduced diagnostic accuracy or even misdiagnosis.As a result,it is critical to reduce noise from ECG signals.Although there are many methods for ECG noise reduction,the effective ECG signal spectrum often overlaps with the noise spectrum in the transform domain,making it difficult to enhance the noise reduction performance of traditional methods.With the development of big data medical technology,the massive amount of patient data provides a data source for deep learning methods to study,and deep learning-based research for ECG noise reduction has emerged in recent years.However,these methods rely on sample selection and the denoising effect still needs to be improved.Generative adversarial network(GAN)is an emerging method that has been effectively employed in the domain of image processing.In this thesis,we introduce it to the task of ECG noise reduction and investigate how to build a better model for this method.And the denoising performance of the method is also analyzed through a series of experiments.The essential points of the thesis are as follows.(1)A GAN-based noise reduction algorithm for ECG signals,AE-GAN(Auto-encoder GAN),is proposed.The method has the advantage of adversarial training,where the generator and discriminator can automatically complete the update of the whole network parameters through gaming.After training is completed,the optimal generator model is saved as the noise reducer.The noisy ECG signal is processed by this noise reducer to output the denoising signal that is fundamentally consistent with the raw ECG signal,so that the noise can be reduced.(2)To address the problem of over fitting caused by too deep hidden layers in the AE-GAN algorithm,two optimization models are proposed to improve the noise reduction effect.These two models are based on recurrent networks and convolutional networks,namely RAE-GAN and CAE-GAN,respectively.Complex non-linear units based on RAE-GAN can construct larger deep neural networks,making the generator more suitable for processing ECG data and improving the accuracy of noise reduction.The CAE-GAN-based denoising network maintains spatial locality and input neighborhood relations in the potential high-level feature representation of the convolutional layer so that low-dimensional samples contain as much noise and ECG signal information as possible.Therefore,the number of nodes at the network layer is simplified,and the method’s computational complexity is lowered.By comparing with the start-of-the-art methods,the noise reduction algorithm based on CAE-GAN proposed in this thesis can achieve a better noise reduction effect.(3)In order to validate the method’s usefulness,a number of experiments are conducted to further investigate the performance of CAE-GAN.The R peak detection experiment demonstrates that this strategy efficiently preserves R peak information.The experimental results of classification of diseases by SVM show that ECG signal after noise reduction is still of certain medical value.Experiments on new leads and new record datasets further validated the method’s generalizability.In conclusion,the proposed CAE-GAN performs exceptionally well in terms of denoising.In summary,this method not only removes single noise and mixed noise but also retains the relevant information of the ECG signal after noise reduction.Although this method achieves good noise reduction results,further verification of the performance of this method in real systems is required.