Research On ECG Classification Method Based On Multi-Task Deep Neural Network

Electrocardiogram(ECG)is a commonly used examination method for screening and diagnosing cardiovascular diseases,and automatic and accurate classification of ECG signals is of great significance for early detection and timely diagnosis and treatment of cardiovascular diseases.In recent years,data-driven methods such as deep learning have made significant progress in the field of ECG classification.However,most of these classification models lack the guidance of prior information,which easily lead to overfitting for a specific single classification task on the training set.It is difficult to learn the deep representation characteristics of ECG signals for these models,and their generalization performance is not high.To address above problems,this study presents an ECG classification method based on multi-task learning and deep neural networks,mainly including the following contents:(1)A multi-task learning and convolutional neural network-based ECG classification model is proposed to address the issue of ignoring the correlation between different ECG categories when using a single task deep learning model for ECG signal classification.Among them,residual structure is used to solve the gradient vanishing and network degradation problems caused by a deep network structure,which is required by convolutional neural networks for extracting deep level electrocardiogram signal features.Considering that an ECG signal is a continuous sequence,this paper introduces a selfattention mechanism and a long short-term memory network on the basis of a multi task residual network to model data dependencies in the sequence.The multi-task residual network model is implemented through hard parameter feature sharing,which includes two parts: feature sharing and independent feature extraction.This thesis adopts two different strategies to create auxiliary tasks and conducts experiments on two twelve-lead ECG datasets.The final accuracy on the two datasets is 0.823 and 0.843,respectively,improving the accuracy of ECG classification and verifying the effectiveness of the proposed model.(2)To address the issue of deep neural networks neglecting the unequal importance of feature information between leads in feature extraction of twelve-lead ECG signals,a squeeze-and-excitation residual network is proposed as a hard parameter feature sharing network.It re-assigns weights to each lead based on the importance of the information contained in the lead,and pays greater attention to the leads with higher weights.This study also proposes using contextual self-attention model to model and process attention mechanism and ECG sequence information simultaneously,achieving information exchange at different positions in the ECG sequence,and thus obtaining static local sequence information and dynamic global sequence information.Considering the characteristic of information exchange between adjacent positions in the ECG sequence,a bidirectional gated recurrent unit mechanism is used to further extract sequence information and enhance the connection between the positions before and after the sequence,achieving further extraction of feature information in the ECG sequence.This thesis validates the proposed method on two twelve-lead datasets and proves the effectiveness of the proposed model by comparing it with existing methods.The final accuracy,F1 score,and AUC of this method on the 2018 China Physiological Signal Challenge dataset are 0.824,0.837,and 0.976,respectively,and the final accuracy,F1 score,and AUC on the PTB-XL dataset are 0.849,0.833,and 0.925,respectively.