Feature Representation And Classification Of ECG Data Based On Deep Learning Quantified Weights

Remote wearable monitoring devices can continuously monitor heart activity and improve medical services for patients with cardiovascular disease.These devices record a large number of original Electrocardiogram(ECG)data,providing a data basis for deep learning technology for classification and recognition.The core of deep learning technology is to map the discrete category data of each domain into an expression suitable for neural network input,and extract deep abstract features by automatic mining of original information.In essence,its automatic extraction mechanism is to combine low-level features to form more abstract high-level representative attribute category features and discover distributed feature representations of data.Such causal uncertainty is the key to the improvement of deep learning techniques.ECG interpretation using deep learning has achieved high recognition results.However,the reasons why deep learning technology can achieve effective ECG classification and recognition remain to be explored.In this paper,based on ECG timedomain features,frequency-domain features,sequence features and spatial features,deep learning technology is used to learn,and a classification method with feature expression ability is explored.The main content of this paper is as follows:(1)Based on the time-frequency characteristics of ECG,a hybrid model of the Gathering Decomposition Tree and GRU(Gathering Unit)network,abbreviated as TGRU(Tree and Gathering Unit),was designed for classifying and recognizing a variety of arrhythmia diseases.ECG has multi-domain characteristics.According to this characteristic,Gini value was used to analyze the contribution of each power spectrum entropy feature to the classification results,and GRU model,which is good at learning the timing characteristics of ECG,was used to classify and recognize the original ECG.Finally,CNN network was used to learn the correlation between GRU model and tree model and sample labels.The proposed algorithm was verified by using the data from MIT-BIH arrhythmia database.The experimental results show that in the ECG classification task,compared with the high frequency band,the frequency domain features of the low frequency band have a greater impact on the classification results of the fusion model.(2)In order to further improve the recognition rate of T-GRU fusion model,based on ECG time domain characteristics,SA-GRU and MA-CNN integration framework was designed to improve the GRU model in the T-GRU fusion model.Firstly,the sequence attention vector was introduced into the two-way GRU network,and the importance weight of each sequence feature to the model learning was obtained.Then,the two-layer convolution network is used to carry out multi-scale convolution on the output feature vectors of the bidirectional GRU.Meanwhile,the spatial attention vector is introduced into the fusion features obtained by multi-scale convolution to learn the importance of ECG spatial features.It is found that it is necessary to consider the sequence of feature learning when analyzing the global and local characteristics of ECG.(3)In the process of decision tree training,when the frequency band attribute judgment standard value is close to the critical value,the attributes will be randomly selected to split the nodes.An incremental learning algorithm is designed to change the model training method and actively eliminate abnormal samples.Through the experiment,the F1 value in the result is increased by 1.78%,which indicates that changing the training method of the model and eliminating the samples in the bad frequency domain can improve the generalization ability of the model to a certain extent.