Research On ECG Signal Classification Algorithm Based On Convolutional Neural Networ

Cardiovascular diseases(CVDs)are serious threats to human life and health.The early stage of the CVDs is often accompanied by arrhythmia,which can be screened by Electrocardiogram(ECG).With the increasing volume of ECG data,the shortage of medical resources cannot guarantee the effective diagnosis of arrhythmias.Therefore,the automatic detection of ECG signals is of great significance for clinical diagnosis and follow-up treatment of arrhythmias.With the rapid development of deep learning,the dilemma of traditional classification methods relying on manual feature extraction has been broken,and deeper feature information of ECG signals has been dug out to realize more efficient and accurate classification of arrhythmias.In this paper,two arrhythmia classification models are proposed based on the Convolutional Neural Network.The original ECG signals and the ECG signals with additional noise are respectively classified by them,and the proposed models are compared with other classical network models.The main research work is summarized as follows:(1)Currently,most studies use filtering to eliminate the effect of noise in ECG signals on the model.However,this can lead to the partial loss of the heartbeat information,resulting in poor classification accuracy.Therefore,in this paper,the ECG signal is directly converted into two-dimensional image without filtering,which preserves the data information to identify arrhythmias more accurately.In addition,the ECG waveform acquired in the real scene is more complex and contains more noise than the signals in the dataset.In this paper,three kinds of common noises in ECG are manually added to the original ECG signal to simulate the real scene and explore a more useful ECG signal detection algorithm.(2)Given the problem that noise filtering may lead to the decline of classification accuracy,the SE-CNN model based on Convolutional Neural Network(CNN)and Squeeze-Excitation Networks(SENet)is proposed.In the preprocessing stage,the one-dimensional ECG signals are directly converted into two-dimensional grayscale images.And then the data volume of minority classes is expanded by a specific data augmentation method to achieve data balance to a certain extent.Finally,the images are input into the SE-CNN model for classification detection.In the convolution part of the SE-CNN model,the continuous small convolution kernels are adopted for feature extraction,which reduces the number of parameters and thus accelerates the training process.The SE module embedded in the convolutional layer can assign higher weight to important information and ignore the effect of noise.The SE-CNN model is used to classify eight heart rhythm types in the MIT-BIH arrhythmia database,and the accuracy reaches 99.53%,which shows obvious performance advantages compared with other classical models and provides a feasible scheme for accurate and efficient arrhythmia detection.(3)To further reduce the noise sensitivity of the classification model,the DS-ECGNet model based on a deep shrinking network is proposed.Firstly,the ECG signals are segmented into 3-second fragments and converted into two-dimensional spectrograms by short-time Fourier transform.Then the spectrograms are input into the DS-ECGNet model for classification.The soft thresholding subnetwork is embedded in this model and automatically calculates the threshold through the attention mechanism to filter out the noise.In addition,the Focal Loss function is employed to optimize the model to eliminate the effect of data imbalance.The DS-ECGNet model achieved 99.74%classification accuracy in the original ECG signal detection.In addition,the stability of DS-ECGNet is verified by experiments on ECG signals with additional noise.Compared with other classical models,it has better anti-noise performance and broad application prospects in the auxiliary diagnosis of arrhythmia.