ECG Signal Processing Based On Deep Learning

Cardiovascular disease has always been one of the major threats to human health.Among the main tools used to assist in the diagnosis of cardiovascular disease,ECG is indispensable.However,ECG acquisition is often accompanied by noise,such as baseline wander,electrode motion artefacts,muscle artefacts and other noise.It is therefore important to pre-process ECG signals for noise reduction.This thesis is dedicated to the study of ECG signal processing techniques to improve the quality and accuracy of ECG signals and help doctors to better diagnose cardiac diseases.This thesis focuses on noise reduction methods for ECG signals based on auto-encoder and flow-based models.(1)This thesis proposes a disentangled autoencoder based ECG number noise reduction method.By introducing a disentangled mechanism and focusing on shunt potential variables,a new disentangled autoencoder network model is proposed,which can effectively separate signal features from noise features and retain useful detailed features of ECG.In the MIT-BIH arrhythmia database,the average improved signal-tonoise ratio of this method for the three noises is 27.45 d B for baseline drift,25.72 d B for muscle artifacts and 29.91 d B for electrode motion artifacts,respectively.compared with the denoising autoencoder based on a fully convolutional neural network,the signal-tonoise ratio is improved by 12.57% on average.(2)The deep flow model ECG signal noise reduction method proposed in this thesis is a new noise reduction framework based on flow model,which uses a series of flow model transformations to process the noise distribution extracted from the complex ECG distribution and inverts back to the original clean ECG signal distribution to achieve efficient noise reduction of the ECG signal.In particular,a new deep coupling layer and a reversible convolutional neural network based on singular value matrix decomposition are designed in this method to enhance the model for signal feature extraction and further noise reduction processing of ECG signals.Through extensive experimental validation,it is demonstrated that the method achieves excellent performance on both the MIT-BIH arrhythmia database and the European ST-T database.The overall design of this thesis meets the expected objectives and has good practical and applied value to help improve the quality and accuracy of ECG signals.