Research On ECG Abnormal Events Detection Method Based On Convolutional Neural Network

Nowadays,lots of people have cardiovascular disease,and the detection of abnormal ECG events is particularly important.The existing abnormal ECG multi-classification detection accuracy needs to be improved,and there are problems such as imbalanced data categories,too little label data,and abnormal ECG waveform distortion.In view of these problems,the main research contents of this thesis are as follows:1)Aiming at the problem that the existing models cannot effectively extract multi-scale spatial information from the ECG signal input with two channels,this thesis proposes a multiscale prediction residual CNN model based on two channels.By analyzing the advantages and disadvantages of existing convolutional networks,combined with the characteristics of ECG signals,a dual-channel multi-scale prediction residual convolutional neural network is proposed to extract ECG features.Because the ECG data has the characteristics of unbalanced categories,the classifier is inclined to large categories of samples,so a hierarchical classification model is adopted to classify the large categories of samples first,and then perform secondary classification for the minority samples.The experimental results showed the detection of Se,Sp,and F1 of N-type is significantly higher than that of existing methods,and the detection of Se and Sp of S-type exceeds that of existing methods by 3%.2)ECG feature point location can obtain the timing information of ECG signal,which is helpful for abnormal ECG detection,but it is difficult to accurately locate abnormal ECG due to waveform distortion.Therefore,this thesis proposes an abnormal ECG detection model based on fusion ECG feature points.The high-precision positioning of ECG feature points is realized by the method of target detection,and multiple feature points of the ECG signal are extracted,including P wave,QRS wave,and T wave,and the extracted feature points are integrated as important features.into the subsequent classification model.The experiments showed the positioning error of the start and end points of the P wave is reduced by 0.1ms compared with the existing method,the positioning error of the QRS end point is reduced by3 ms compared with the existing method,and the location error of the T wave start point is reduced by 1ms compared with the existing method.The located ECG feature points are integrated into the abnormal ECG detection model.Through comparative experiments,the Se,Sp,and F1 of N-type detection are higher than the existing methods,and the detection F1 of minority-type samples S exceeds the existing method by 20%.3)Aiming at the problem that too little labeled ECG data may easily lead to overfitting in detection model learning,semi-supervised abnormal ECG model with GAN is uses.The original data is expanded by the way of enhancing the unlabeled data through the generative adversarial network,and the semi-supervised model is used to combine the labeled data with the unlabeled data for consistent training,which finally makes the model more robust.Through the visualization of intermediate results and comparative experiments,for N-type detection Se,F1 is 4% higher than existing methods,S-type detection Se,F1 exceeds existing methods by4%,and V-type detection Sp is 9% better than existing methods.The validity of the model is verified.