The contribution of the phase spectrum in automatic multiple cardiac arrhythmias recognition in wearable systems

In this paper we implement an automatic procedure that is to be embedded in a wearable system in order to discriminate five arrhythmic classes of QRS complexes from normal ones. Due to the limited hardware resources offered by the wearable system, several requirements such as low computational cost, memory usage, reliability and real-time have to be addressed. In order to better comply with these requirements, the classification process is performed using features that can easily be extracted from the signals, i.e. magnitude and phase of the Fourier Transform (FT) applied to the QRS complexes. The ECG signals, from which QRS complexes are extracted, are gathered from the MIT-Arrhythmias Database. More specifically, three datasets of features are created: the first (alpha) is obtained from the magnitude, the second (beta) from the phase, and the third (gamma) from the union of the two. According to the results of the Royston Multivariate Normality Test, which verifies the gaussianity of the distribution of the three sets of features, a parametric, Nearest Mean Classifier (NMC), or non-parametric, MultiLayer Perceptron (MLP) classifier is used. The comparative performance evaluation is showed in terms of a confusion matrix obtained from twenty steps of cross validation. The matrices report the percentage of successful recognition of the six classes.