Research On Arrhythmia Classification Methods Based On Deep Learning

Cardiovascular diseases now have the highest mortality rate of all non-communicable diseases,and this figure is increasing year by year.Arrhythmia is one of the most common cardiovascular diseases.If the condition is not detected in time,the patient may die suddenly due to a sudden attack or have persistent heart involvement.Therefore,accurate diagnosis of arrhythmia is crucial.According to the evaluation mechanism of ECG classification,there are two common deep learning methods for arrhythmia classification: arrhythmia classification methods based on intra-patient paradigm and arrhythmia classification methods based on inter-patient paradigm.Under the intra-patient paradigm,arrhythmia classification method does not consider the difference between individuals,which is easy to implement.Under the inter-patient paradigm,ECG of different subjects are used separately for training and testing.It ensures that the samples of testing and training do not overlap,which is in line with the actual medical application scenarios.We propose three arrhythmia classification methods.The following three aspects are carried out:(1)Considering the low accuracy of existing ECG classification methods,based on the intra-patient paradigm,a multi-branch feature classification method is proposed.The model extract local features in ECG by using multi-branch convolutional neural network(CNN).Then through the long and short term memory network(LSTM),the multi-branch feature information obtained in the previous step are fused.That can further extract the timing features of ECG.Finally,the classification of ECG is predicted through the full connection layer.Under the intra-patient paradigm,experimental results show that the model was able to achieve 98.89% accuracy with98.89% sensitivity,98.46% specificity,and 98.86% positive predictive value.(2)In order to improve the accuracy of arrhythmia classification based on intra-patient paradigm,a residual-coding model based on fusion attention mechanism is proposed.The residual structure includes a CNN module and Transformer.The CNN module can extract the local features of ECG.Transformer encoder is used to focus on the important feature information in ECG.The extracted features are connected optimized through a residual block to obtain more feature information hidden in ECG.Experimental results show that the model achieved 99.16% accuracy,99.14% sensitivity and 99.16% sensitivity under the intra-patient paradigm.Under the inter-patient paradigm,the model was able to achieve 94.42% accuracy with 94.42%sensitivity,84.14% specificity,and 91.71% positive predictive value.It proves that the residual-coding network can effectively improve the accuracy of arrhythmia classification.(3)Aiming at the problem of low classification accuracy of ventricular arrhythmia under the inter-patient paradigm,multi-level feature extraction network(CLSTM-Transformer)is proposed,which includes three parts: convolution module,timing module and coding module.Among them,the convolution module can obtain the local characteristics of ECG,the coding module can increase the extraction of important features of ECG by focusing on the important nodes of the beats,and the timing module can extract the dependence of the front and after the beats.Experimental results show that the model was able to achieve 98.56% accuracy with93.45% sensitivity,93.45% specificity and 98.54% positive predictive value.