Multi-objective advisory system for arrhytmia classification

Abstract

The study proposes the best electrocardiography (ECG) arrhythmia classification features suited to application needs by using multi-objective approach. The wavelet transform (WT) is successful for ECG classification. Also, the combination of features obtained from different coefficients of different wavelets provides higher performance rate than individual wavelet. However, most of the feature selection algorithm focuses attention on one objective such as accuracy or to the number of features for a real time system. In this study, different solutions were proposed that will increase the classification performance on three different objectives such as positive predictive value (PPV), accuracy and number of selected features. The wavelet type and level that best reflect the 4 different ECG arrhythmia types were searched by using Multi-Objective Evoltionary Algorithm (MOEA). Multilayer perceptron (MLP) was preferred as a fitness function. The non-dominant sequencing genetic algorithm II (NSGA-II) was used and the algorithm ran many times with different seed values. The preferred solutions meeting the preference criteria were examined in detail. The highest accuracy and PPV rate obtained was 97,82% and 94.94%, respectively with 24 features. Moreover, it has been observed that some of the features obtained from an individual wavelet type have a low contribution to the classification performance and some of them can outweigh. To illustrate that the combination of features obtained from different level coefficients of different wavelet types provides more successful discrimination in ECG arrhythmias and to provide feature sets according to the requested metric values are the some of the main contributions of this study.