Feature Selection for Arrhythmia Classification Using Statistical Tests

In this paper, we evaluate the statistical significance of features enabling us to differentiate between the signals obtained from healthy patients and patients with some type of cardiac arrhythmia. The aim of our research is to obtain a unique feature subset from an original multi-domain feature set according to a filtering-based selection method, which selects the relevant features where the redundant and irrelevant features are removed. Feature selection was implemented us ing a statistical test appropriate for the feature distribution. When the normality assumption is satisfied, an unpaired t-test was performed, and an otherwise non-parametric Wilcoxon–Mann–Whitney test. Statistical based feature selection was performed by comparing ECG signals from the MIT-BIH Normal Sinus Rhythm (NSR) and MIT-BIH Arrhythmia (AR) Database.