Study Of The Algorithm For ECG Beat Classification Based On Convolutional Neural Network

Arrhythmia is a disease of abnormal cardiac conduction.The automatic identification of different types of heartbeats is important for electrocardiogram(ECG)analysis,and the results can be used as the basis for the diagnosis of arrhythmia.With the vast database of ECG signals,it is possible to develop an accurate computer-aided diagnostic system to speed up the cardiologist's diagnostic process.In this paper,based on the self-learning convolutional neural network(CNN),the following researches on the heartbeat classification algorithm were carried out.1.The basic theory and application of one-dimensional convolutional neural network(1D CNN)were studied.Aiming at the characteristic that ECG signals were one-dimensional time series,the typical structure of 1D CNN suitable for sequence analysis was explored.Through the detailed introduction of its forward propagation and back propagation processes,as well as activation function and regularization two key optimization techniques,the application process of heartbeat classification based on 1D CNN was analyzed.2.A hybrid convolutional neural network(HCNN)that fuses physical information was studied.In order to solve the problem that the traditional CNN tended to focus only on the heartbeat and ignored other clinically meaningful physical information,a HCNN structure was proposed.It contains two inputs,the main input and physical information input.In this paper,the physical information was spliced into the input layer,convolution layer and full connection layer of traditional CNN respectively,and the three substructures of HCNN were realized by deducing their forward and back propagation processes.The structure not only retains the self-learning ability of the traditional CNN,but also introduces physical information which is beneficial to further classification.In addition,two schemes for applying dropout on HCNN were proposed,one was to use dropout before splicing,the other was to use dropout before splicing.3.The application of HCNN in heartbeat classification was studied.Aiming at the problem that most CNNs have the poor recognition of atrial period contraction(APC),the HCNN combined with helpful RR intervals was applied.In this paper,the traditional CNN and the three HCNNs were designed,and their performance was verified by 2970 clinical patients.The results showed that the average accuracy of HCNN on the 12 leads was87.56%,which was 3.72% higher than that of traditional CNN.And this was mainly benefits from the improvement of APC recognition ability.Among them,the substructure of HCNN obtained by splicing the RR intervals into the input layer of traditional CNN had the greatest improvement.And the APC sensitivity of its V1 lead increased from 56.18% to 80.15%.At the same time,the two dropout schemes of HCNN were experimented,showing that using dropout before splicing could acheieve higher accuracy.