On the Effects of Feature Selection in Atrial Fibrillation Detection

Atrial fibrillation is the most prevalent cardiac arrhythmia, and its automatic detection is an important task. In this research, segments that presented atrial fibrillation were separated from segments with normal sinus rhythm. Dozens of parameters were computed from RR intervals, and each one of them was separately used as input to linear classifiers. Then, the ten parameters that provided the highest accuracies–-maximum over minimum ratio, median of the absolute deviation; sum of the magnitude of spectral components divided by the mean RR interval; mean value of the absolute angles and mean value of the sines in Poincaré plots; standard deviation of the differences between consecutive RR intervals; sample, fuzzy and Shannon entropies; and coefficient of sample entropy–-were standardized and presented as inputs to a support vector machine. Accuracy rates of up to 99.81% were attained.