Classification Of Arrhythmia Based On ResNet-GBDT Fused Model

The phenomenon of arrhythmia is often ignored by people in daily life.Most disease prevention and medical enhancement are put into large-scale diseases such as heart disease and hypertension,but the early manifestation of these diseases is arrhythmia.Therefore,strengthening the diagnosis of arrhythmia is helpful to the research and diagnosis of complex diseases and improve the level of medical technology.Single lead ECG signal data is the key to the performance of cardiac rhythm.Starting from low-cost data,this study screened individual arrhythmias end-to-end and divided them into ventricular,atrial and sinus arrhythmias according to different parts of the origin of ECG rhythm.Many researches on ECG signal classification focus on feature extraction and classification model design of ECG waveform.In terms of feature extraction,most studies process sample information through key feature extraction algorithms of obvious significance in the field of ECG,and enhance data quality with prior theories.However,there are often problems such as excessive annotation information required and imperfect wave group feature extraction algorithm.In terms of model design,previous studies have used a variety of model combinations,and the results show that the combined model is usually superior to a single model.However,the nonlinear integrated classifier as the last link remains to be studied,which provides a basis for the subsequent model fusion.To sum up,in order to solve the problems of difficult extraction of key features and imperfect model design,this paper starts from easily accessible single lead data,uses variational mode decomposition to enhance data information,and combines deep neural network feature extractor with powerful nonlinear ensemble classifier to improve the recall effect of abnormal categories.In terms of data processing and model setting,the sample information is first processed by denoising or decomposition of signal data,and then the class imbalance problem is improved.In order to ensure that the sample information does not have deviation,the deviation is processed by low-intensity resampling technology and loss function weighting,and then the model is established.Based on the establishment of a single model,firstly,the integrated model with high accuracy and strong explanatory advantages of gradient lifting decision tree fitting denoising data,combined selection among comprehensive characteristic data,the overall accuracy of more than 95% can be achieved,ventricular,atrial and sinus arrhythmia can reach about 50%,80%,90% recall rate respectively.Secondly,the long and short term memory network,cyclic neural network and deep residual neural network were used to design the model respectively,and the high dimensional data after variational mode decomposition was used to learn the model.Compared with the random guess gain of LSTM and RNN,the effect of Res Net was better,with f1＿weighted up to 98%.The recall rate of ventricular,atrial and sinus arrhythmias is about 80%,90%,90%.In conclusion,compared with the circulatory neural structure and the integrated model,the deep residual neural network has improved the recall rate of abnormal categories,significantly improving the diagnostic effect.Based on the effect of the above single model,it can be seen from the recall rate of the North Star index of abnormal diagnosis that the optimal single model can make the recall of abnormal categories reach the acceptable range.However,due to the unbalanced sample information,there is still a certain short-plate effect on the least number of ventricular arrhythmias.Therefore,in order to enhance the sample analysis of ventricular arrhythmias,the above models are integrated to make up for the defects of the single model.That is,cyclic neural network or deep residual neural network is used as the base learner and gradient lifting decision tree is used as the meta-learner for stacking fusion.The effects of multiple fusion models are compared.The final conclusion is that the optimal model in this data set is the deep residual neural network and gradient lifting decision tree fusion model(Res Net-GBDT).The overall recognition rate is as high as 99%,and the recall rates of three abnormal categories are all over 90%.The greatest advantage is reflected in the improved effect on ventricular arrhythmias,which are the most difficult to distinguish.It is significantly better than the single application of the two kinds of learning devices,and better than other fusion models,so that the model can achieve high quality in the diagnosis performance of abnormal categories.Secondly,the second conclusion is drawn during the research process.After comparing the single and fusion models,the circulating neural structure is not sufficient to extract the time series information.In the abnormal recognition scenario of arrhythmia,the most attention is paid to the recall performance of abnormal categories.In order to achieve the purpose of high recall application,the variable mode decomposition is carried out based on the simple and easily accessible single lead ECG data,and the Res Net-GBDT fusion model is used as the research and judgment mechanism to realize the abnormal diagnosis of possible patients,improve the medical efficiency and strengthen the diagnostic effect.Assist in reviewing the doctor’s diagnosis.