Neural Networks Based Diagnosis of Heart Arrhythmias Using Chaotic and Nonlinear Features of HRV Signals

In this paper, an efficient novel algorithm is presented for classification of the most important heart arrhythmias. By computing Heart Rate Variability (HRV) signal form electrocardiogram (ECG) signals, 14 carefully selected time domain, frequency domain, nonlinear and chaotic features are extracted and used to train MLP neural networks. Before applying those features to neural networks, we reduce the order of feature space by generalized discriminate analysis (GDA). Furthermore, to improve the learning performance of MLP network, training set is filtered by deleting confusing data. This is done by means of a self organized map which categorizes train data set and indicates data which are not representative to be ignored. HRV signal is known to be less sensitive to noise as compared to ECG and has higher chaotic and nonlinear characteristics. 7 arrhythmias (i.e. PVC, AF, CHB, LBBB, NSR, VF, and VT) have been classified with accuracies ranged from 95% to 100% on MIT-BIH dataset.