Arrhythmia Classification System Based On ResGC-Net

The fast-paced life has overwhelmed more and more people,and the number of patients suffering from diseases,especially cardiovascular disease,has also increased year by year.Cardiovascular disease is the world’s largest death disease,one of the most serious diseases is arrhythmia,which kills a large number of patients every year.However,due to the influence of population,education and other factors,the number of institutions carrying out arrhythmia diagnosis and treatment is far from enough to cover the number of patients,in order to save more people with limited medical resources,the early warning of cardiovascular disease is very important.At this time,it is necessary to realize the automatic diagnosis of arrhythmia by computer to assist doctors in diagnosis and treatment.At present,the mainstream arrhythmia classification algorithms mainly use ECG for classification after preprocessing and feature extraction.Nowadays,most researchers in the world have put forward many effective classification algorithms,but some problems of these algorithms need to be improved:Firstly,due to the inevitable mixing of various noises in the process of signal acquisition,it is necessary to denoise the collected ECG signal in the process of classification,and the cumbersome denoising algorithm and parameter adjustment will cause the loss of the original signal characteristics.On the other hand,due to the large number of normal beats and less pathological beats in the arrhythmia classification system,the data skew problem also restricts the accuracy of the classification model.In order to solve the above problems,this paper makes the following research:（1）An automatic arrhythmia recognition algorithm based on gram angle field and ResGC-Net is built.Firstly,gram angle field is an imaging method to transform one-dimensional signal into two-dimensional matrix,which can transform one-dimensional arrhythmia signal into two-dimensional matrix representation with more features while retaining time correlation features.Then,through the comparative experiments of various attention mechanisms,GCblock module with less parameters and high accuracy is selected.GCblock is a global context modeling module.Using the method of global attention pooling,we can establish effective long-distance dependence and better extract the dependence characteristics of timing signals in neural network,it can deeply extract the characteristics of different kinds of arrhythmia signals.ResGC-Net introduces the combination of gram angle field,Resnet and GCblock for the first time,fully extracts and retains the time correlation features and long-distance dependence,and then classifies the arrhythmia signals better,and achieves the classification accuracy of 99.37%.Compared with other literatures using the same database,the recognition accuracy of this paper is better than other algorithms.（2）Aiming at the problem that ResGC-Net is prone to misclassification when processing skewed data sets,this paper introduces OHEM（Online Hard Example Mining）module to give attention mechanism to ResGC-Net model,repeatedly trains the beat types with large loss value after calculation by softmax classifier（i.e.the types that the model cannot correctly classify,mostly unbalanced data）,and gives more training and weight to the class unbalanced data,so as to improve the class imbalance problem of the model,Improve the recognition accuracy of unbalanced data.At the same time,this paper innovatively improves the OHEM algorithm and proposes two parameter transformation methods（the proportional parameter of the data set and the arithmetic mean of the whole data set）to replace the parameters to be adjusted of the original algorithm（the number of heart beats that are difficult to identify）.Replacing the original parameters with the arithmetic average value not only saves the long process of parameter adjustment,but also improves the upper limit of the accuracy of the system.The overall classification accuracy in MIT-BIH database reaches 99.47%.In order to verify the ability of the model to process tilt data,the data attenuation experiment is carried out on MIT-BIH database.In addition,in order to verify the improvement effect of each module on the model,we designed ablation experiments and discussed the results.Finally,the anti-noise robustness of the model is verified on the MIT-BIH noise stress test noise data set.Based on the above comprehensive performance,the system lays a foundation for the effective clinical monitoring of arrhythmia.