Statistical filtering methods for feature selection in arrhythmia classification.

In this study we demonstrate the appropriate use of statistically based filtering methods for feature selection and describe the application to Heart Rate Variability (HRV) features used to distinguish between arrhythmia and normal sinus rhythm electrocardiogram (ECG) signals. The initial set of HRV features is evaluated using both correlation and statistical significance tests. Normality assumption is assessed for each feature in order to select appropriate correlation methods and significance tests. In addition, the impact of outliers on the statistical test results is illustrated by an explorative analysis of correlation before and after outlier removal. Finally, a reduced set of features is selected, and the decision process guided by correlation and statistical significance test results is described and discussed.