A study on arrhythmia classification based on local maximum scalogram using convolutional neural network

The purpose of this study is to investigate whether it can be used for arrhythmia detection as a wavelet transform and another feature extraction method using the probability distribution of LMS(Local Maximum Scalogram). The SVM(Support Vector Machine) model uses two kernels: Polynomial and Radial Basis Function(RBF). Three types of data are used: standard deviation of the row (ηv), standard deviation of the column(ηv/2), and standard deviation of the row and column (3ηv/2) of the basic LMS matrix according to the sample (ηv = 90,180, 270, 360, 450), The training data of the CNN model uses two LMS matrices when fixed to and. The trained model is divided 5 times, and K-fold cross-validation is performed and evaluated using ROC, AUC, and confusion matrix. Finally, the filtered ECG data is compared with the confusion matrix result graph to consider the types of arrhythmia that are difficult to classify. CNN is evaluated to show good overall performance when is. The results of the SVM model show the possibility that the standard deviation values of the LMS's rows and columns can be used as a feature of arrhythmic bit detection. Since it is simple but very efficient, it is expected to be used in various ways as a single feature extraction method. Copyright © 2021 The Korean Institute of Electrical Engineers.