Statistical-based HRV Feature Importance Evaluation for Arrhythmia and Atrial Fibrillation Classification
The paper evaluates statistical significance of the differences in the feature values necessary to differentiate the signals corresponding to cardiac arrhythmia (AR) and atrial fibrillation (AF). The initial set of heart rate variability (HRV) features includes time and frequency domain metrics, as well as geometric metrics based on the Poincare diagram. Due to non-uniformity of the heart rate signal, frequency domain features are calculated using two approaches: the Lomb-Scargle method for spectral analysis for non-uniform signals, and Welch method for uniform signals, but after the signal interpolation and resampling. Selection of an appropriate statistical test was depending on the distribution of feature values. Normal distribution allowed use of parametric ANOVA test and otherwise non-parametric Wilcoxon–Mann–Whitney test were used. The statistical tests indicated statistically significant difference between the two observed groups of signals of interest with respect to the evaluated feature. The success of the classification depends on the well-chosen features according to their importance. In the paper, statistical tests resulted in selection of 27 features out of the initial 51. The proposed set of features could be used for the classification between the AR and AF signals to assist diagnosis of the mentioned heart diseases.