Higher order statistics and ECG arrhythmia classification

For cardiologists, the problem of ECG arrhythmia is to discriminate different kind of arrhythmia from a normal cardiac rhythm. Our goal is to discriminate ventricular fibrillation (VF) and ventricular tachycardia (VT) from a normal sinus rhythm (NSR). VT and VF are fatal arrhythmia for patients and are treated by electroshock. The nature of the shock depends on a VF or a VT condition. An automatic discrimination between these two conditions may help medical personnel, with varying level of skills, to give the proper treatment. The cardiac rhythm is modeled by a auto regressive (AR) model. A code book using a learning vector quantization (LVQ) from different ECG segments with and without different abnormalities is constructed. Using different criteria, we show that the optimum order for the classical AR model is five. Classification using distance measures (Itakura and euclidian) between feature vectors and the code book vectors detect NSR from other arrhythmia, VF and VT conditions are detected with an error less than 10%. Since ECG signals are mainly nonGaussian and non linear, the AR coefficients are sensitive to noisy data. Using a higher order AR model, the third and fourth cumulants are non zero and noise insensitive. Higher order AR modeling can be used advantageously for classification and discrimination between fatal arrhythmia. A code book is constructed and classification using a distance measure can be performed. We show that the approach using higher order model outperform the previous approaches since it allows a better discrimination between VF and VT conditions.