Research On Arrhythmia Classification Based On Deep Learnin

The detection of arrhythmia disease is one of the research hotspots at present.According to the public data set provided by the 2018 physiological signal challenge and the characteristics of ECG signals,this paper proposes effective signal denoising algorithm,data balance method and arrhythmia diagnosis algorithm,realizes the multi label classification of 9 arrhythmia categories,verifies the effectiveness of the algorithm on the Georgia ECG database,and finally uses a variety of evaluation indicators to verify the advantages of the algorithm proposed in this paper.The main research contents and achievements of this paper are as follows:First,through analyzing the features of ECG noise,a denoising method of the multi noise ECG signal is firstly proposed,in which the baseline drift noise can be filtered by the median filter that has low computational complexity and high realtime and the noise of the power frequency interference and EMG interference can be filtered by the wavelet analysis method.By comparing the signal-to-noise ratio and mean square deviation,it is better to filter the noise in the order of baseline drift,power frequency interference and EMG interference,so as to make the obtained ECG signal more smooth.Aiming at the imbalance of ECG data samples,the combined sampling technology of under sampling and grouping and manual clipping is adopted to greatly reduce the impact of the model on the classification performance of a few ECG samples.Second,according to the recognition characteristics of 12 lead ECG signals,the DNNs model,which is currently more authoritative in the field of ECG classification,is improved for arrhythmia classification.A larger convolution kernel is selected at the first layer of neural network.With the reduction of the size of the feature map,the size of the convolution kernel is gradually reduced,and the dilated convolution is used in the residual block to replace the traditional convolution,so that the model can obtain a larger receptive field,Then,at the bottom of the network,the two feature extraction schemes of Global Average Pooling and Long Short-Term Memory network are compared.At the same time,combined with the advantages of dense connection network,a parallel combination model based on residual network and dense connection network is proposed to further improve the classification effect.Third,aiming at the shortcomings of practicability and classification performance of parallel composite model,an improved residual dense network model is further proposed.Firstly,the residual module is designed.The correlation between 12 leads is obtained by depthwise separable convolution instead of traditional convolution,and the common features between leads are extracted;At the same time,through the channel attention mechanism,we pay more attention to the valuable and differentiated ECG segments while paying attention to the correlation of ECG leads,so as to realize the feature selection and improve the weight distribution of important features.Three schemes are designed to explore the combination mode of channel attention mechanism;Finally,through the global feature fusion strategy,the model can obtain both shallow and deep features at the same time,so as to prevent the over fitting of the model.Fourth,parameter optimization and result analysis of algorithm experiment.Build neural network models with different depths,reasonably configure and optimize the number and size of convolution kernels,select the appropriate optimizer and loss function,select different batch sizes and model training rounds for experimental verification,and finally use a variety of evaluation indexes to get the advantages of this algorithm.