Automated cardiac state diagnosis from hybrid features of ECG using neural network classifier

Abstract

Accurate non–invasive Electrocardiogram (ECG) analysis has a significant emerging role in automated cardiac state diagnosis. Conventional ECG analysis techniques such as linear and second–order spectra fail to retain Fourier phase relationship and suppress random variations in non–linear, non–stationary and non–Gaussian ECG signals. This may provide misguided results. A highly accurate algorithm utilising statistics of fourth–order spectra (trispectrum) is introduced to capture clinically significant variations in ECG that can cater to these limitations. Five temporal interval and three trispectral entropy features are extracted from individual beat of ECG signals loaded from MIT–BIH arrhythmia database and quantified using box plots. A three–layer feedforward neural network classifier with 50 hidden layer neurons is configured using the Levenberg–Marquardt algorithm to yield an average accuracy of 95.10% while classifying six cardiac states. Significant enhancement in the performance with reduced computational complexity has been observed using a low–dimensional hybrid ECG feature set and a simple classifier showing the effectiveness of the proposed man–machine interface for automated cardiac state diagnosis.