Research On ECG Signal Classification Based On Deep Neural Networ

The Electrocardiogram(ECG)signal is the representation of the heart’s electrical activity.If a cardiac disease suddenly occurs,the ECG will be abnormal,and needs for timely and accurate diagnosis of arrhythmia category.When collecting ECG signals,it will generate high and low frequency noise,and because of the many types of arrhythmias,only relying on the clinician to observe the ECG not only time-consuming and labor-intensive,but also prone to subjective errors,thereby reducing the reading efficiency and accuracy of ECG signals.In order to reduce the pressure of doctors,improve the waveform quality of ECG and the classification efficiency of ECG signals,this paper based on the deep neural network algorithmia,deeply studying the automatic classification of ECG signals,the main research contents are as follows:1.Aiming at the high and low frequency noise of the ECG signal: using the Discrete Wavelet Transform,the Butterworth Low-pass Filter and the Moving Average Filter to filter out the high and low frequency noise,after noise reduction the ECG signal is smooth and stable,laying the foundation for improving the accuracy of signal classification;Aiming at the segment segmentation of long-term ECG signal: 1)For the two-lead 30-minute long time ECG signal,intercept 360 sampling points which including R peak points as a heart beat cycle,and annotate the category labels according to the individual heart beat signals.2)For the 12-lead indefinitely long time ECG signal,cut them into 24 ECG fragments by a fixed number of sample points and annotated as the same ECG label.Using the difference threshold method and Hamilton＿segmenter algorithm to detect the position of R peak,at the same time,the ECG signal is standardized by zero-mean and oversampled,which improves the stability of signal analysis and classification efficiency.2.Aiming at the automatic classification of the multi-channel ECG signal,designed the Convolutional Neural Network(CNN)model for multi-channel feature extraction and independent heart beat classification.On the basis of CNN,the network width and length are extended,in order to extract the important and correlated feature information in dual-lead ECG signals,compared with the effect of different feature extraction network layers on signal classification in the Inception structure,the 1D-CNN-Inception12 with the maxpooling layer and the dilated convolution has the highest F1 value for the automatic classification of 5 kinds of heart beats.3.Aiming at the waveform characteristics and timing characteristics of the12-lead ECG signal,a concatenated CNN with an attention mechanism,which is connected in series with Bidirectional Long Short Term Memory(BiLSTM)network is designed for each lead signal.In this arrhythmia classification model,the concatenated CNN is used to extract the local ECG features,the Squeeze-andExcitation module is used to increase the weight distribution of the abnormal band features,the BiLSTM is used to extract the ECG timing information,the XGBoost algorithm is used to fuse the 12 groups of model classification probabilities and Heart Rate Variability features.Compared with the model without feature fusion and attention mechanism,the fusion model has a higher accuracy for the automatic diagnosis of 9 arrhythmias.