Higher order statistics towards cardiac state diagnosis using neural network classifier

Electrocardiogram (ECG) signal analysis is an established clinical tool for the cardiac state diagnosis. Being a physiological signal, ECG is non-linear, non-stationary and non-Gaussian in nature. Most of the ECG signal analysis tools such as linear and power spectrum estimation ignore random variations and the Fourier phase relationship among signal components. This can lead to inaccurate classification of cardiac states. A highly accurate and efficient ECG signal analysis technique for cardiac state classification using higher order statistics is presented that caters to these limitations. It uses fourth order statistics i.e. trispectrum related features of ECG signal. The trispectrum dependent features are extracted from each QRS complex of normal and abnormal ECG signals selected from MITBIH arrhythmia database. These are used to train a feedforward neural network classifier using Levenberg Marquardt training algorithm. An overall classification accuracy of 98% is achieved in the classification of three cardiac states viz. normal, left bundle branch block and nodal (junctional) escape beat using proposed method. The experimental results reveal that higher order statistics is a promising cardiac signal analysis tool to capture significant features from ECG signals accurately and efficiently.