Effects of the Window Size and Feature Extraction Approach for Arrhythmia Classification

Abstract

Arrhythmia, a common form of heart disease, can be detected from an electrocardiogram (ECG) signal. This research work presents a comparative study between five feature extraction methods applied separately on two window sizes for detecting three ECG pulse types, namely normal and two arrhythmia variations. The library support vector machine (LIBSVM) was used to classify the three classes of the ECG pulses. The ECG signals were obtained from MIT-BIH database. The ECG dataset was normalized and filtered to remove any noise and after that the signals were windowed into two window sizes (long window and short window). Five approaches were used to extract the features from the ECG signals. These approaches are scalar Autoregressive model coefficients, Haar discrete wavelet transform (DWT), Daubechies (db) DWT, Biorthogonal (bior) DWT, and principal components analysis (PCA). Each approach was applied separately on the two window sizes. The results of the classification show that scalar Autoregressive model coefficients, Haar, db, and bior are better approaches to catch the ECG features for short window than the long window. However, PCA gave the closest and highest results for the two window sizes than other approaches. That mean the PCA is the better feature extraction approach for both window sizes.