One Class Support Vector Machine Classification based on High-Order Statistical Features of ECG Signals

Abstract

In this paper, we begin with the point that the distinct difference between normal heart beats and abnormal heart beats gives significant statistical features. 1st derivative of the input signals is applied to extract primitive features of the ECG signals. Before obtaining the histogram of the 1st derivatives, normalization of the 1st derivative signals and quantization of the normalized 1st derivative signals to keep the same dimension are applied. Moreover, more important high-order statistical features such as variance, skewness and kurtosis, can be extracted from the histogram of 1st derivative signals. Because of considering the classifier to be trained using only normal ECG signals, an OCSVM is applied to the extracted feature spaces. In this classifier, Gaussian kernel is used for kernel transformation. The sampling rate of all the ECG signals is 128Hz. The histograms of the normalized 1st derivatives for normal and abnormal ECG signals are 35 levels in the experiment. Therefore every ECG signal based on the histograms is represented by a 35-dimensional feature vector. 36 normal ECG signals are used for training and 94 ECG signals are used for testing the proposed OCSVM classification model. The model shows significantly improved performance for the arrhythmia testing data with 94.9% correct classification rate.