Research On Classification Method Of Imbalanced ECG Samples Based On Deep Learning

Arrhythmia is a kind of cardiovascular disease caused by abnormal heartbeat,which is a great threat to the life of patients.How to diagnose arrhythmia accurately and effectively is an important problem to be solved in the medical diagnosis of cardiology.Electrocardiogram(ECG)is a standard technique used to detect cardiac activity,and it is the main basis for clinical diagnosis of arrhythmias at present.In the current clinical diagnosis,doctors mainly judge the ECG with the help of the naked eye.Because human eyes are prone to fatigue and the judgment is easy to be affected by subjective factors and other reasons,this way may lead to misclassification of ECG.Automatic classification of ECG with the help of deep learning technology can overcome the shortcomings of naked eye diagnosis and improve the classification accuracy,which is of great significance to solve the problem of arrhythmia diagnosis.The existing deep learning methods are generally better than the traditional methods in ECG heartbeat feature extraction.However,due to the serious category imbalance in most of the existing ECG databases,the existing deep learning methods cannot extract beat features effectively,resulting in poor classification performance of heartbeat samples.At the same time,the existing classification methods solving the category imbalance cannot effectively improve the overall classification performance and the classification performance of individual classes at the same time.This paper focuses on the above two problems.The main contents of this thesis are as follows:(1)Based on the existing deep learning network model,a beat classification method based on the combination of CNN and GRU based on autoencoder is designed.This method uses autoencoder model for unsupervised learning.In the encoder,the combination of CNN and GRU model is introduced to learn the input features.In the decoder,the CNN model is introduced to reconstruct the input features,so that the input before coding is the same as the output after decoding as much as possible.This method not only reduces the amount of model calculation and model training time,but also enables the encoder to extract more beat features,which can effectively improve the accuracy of classification and solve the problem of insufficient feature extraction ability in class unbalanced ECG database.(2)In order to solve the problem of imbalance of heartbeat samples,in the aspect of ECG data processing,a method of data augmentation based on domain-specific transformation is proposed,including random cropping,translation and noise.To a certain extent,this method enhances the diversity and representativeness of samples and the number of heartbeat samples of a few classes,and alleviates the over-fitting phenomenon that may occur in the classification of a small number of heartbeat samples.(3)A new batch weighted loss function is proposed to effectively quantify the loss.On the basis of the above data augmentation,the classifier is improved,and a new batch weighted loss function is used to better quantify the loss,so that the weight of each batch of heartbeat samples introduced into the model is in a process of dynamic nonlinear change,which effectively solves the problem of class imbalance of heartbeat samples.This experiment is carried out on the MIT-BIH database,which is verified by intra-patient and inter-patient experiments,and the classification results are better than the existing methods.In the classification of inter-patient,the overall accuracy of the method proposed in this paper reached 99.30%,where F1 score reached 94.69%and 98.08%for SVEB and VEB,respectively.In the classification of intra-patient,the overall accuracy of the method proposed in this paper reached 97.37%,where F1 score reached 81.64%and 92.08%for SVEB and VEB,respectively.It is proved that the method proposed in this paper has high classification performance and pathological heartbeat detection performance.