Hybrid feature vector creation for atrial fibrillation detection improvement

Two stages system consisting of feature extraction and selection part followed by neural classifier dedicated for atrial fibrillation (AF) detection, with preliminary ventricular activation cancellation is presented. According to proposed in this paper method the set of parameters obtained from time-frequency signal analysis mixed with features characterizing these signals in separately time and frequency domains was created. As a efficient tool for non-stationary signals analysis the discrete wavelet transform was used to obtain the T-F signal representation and then new parameters based on energy and entropy measure were computed. Features selected based on discrimination measure are the input to neural ECG classifier, where both supervised learnt multilayer perceptron and unsupervised Kohonen maps (SOMs) were tested on the set of 20 AF and 20 patients from control group divided into learning and verifying subsets.