ECG Beat Classification Based On Frequency Domain Features

Cardiovascular disease is one of the dangerous diseases that seriously endanger the human health, while Electrocardiogram (ECG) is an effective method used in clinic to detect cardiovascular disease. Automatically and correctly analysis for ECG with computer program has been a hot issue among scholars at home and aboard for many years, and ECG beat classification has always been a difficult problem.Researches in this field can be divided into time domain method and frequency domain method. Though time domain feathers have obvious physiology significance, they are affected by noise in human body and instrumentation, and show variation in different patients and different period. These increase the difficulty of classification. In frequency domain method, mathematical transforms are applied to the original ECG beat to generate new features which have further information. Fourier transform (FT) and wavelet transform (WT) are tipic transforms. Wavelet transform gradually attracts more and more attention because of its ability of giving prominence to partial characteristics and detecting singular point.In this paper, coefficients of WT on ECG beats are used as features, and BP neural network and support vector machine (SVM) are trained by the feature vectors to classify four types of ECG beats (normal beat, left bundle branch block beat, right bundle branch block beat and paced beat).Preprocessing is applied on samples of ECG beats to remove the affection caused by individual difference and signal excursion. Then multi-level detail and approximation coefficients of db2 WT are used as features to generate the high dimensional feature space. Feature searching algorithm based on maximal divergence, which uses Euclidean distance to measure the distance between classes, is applied to get the optimized feature combinations at different dimensions. Searching result shows that eight dimensional combination gets the high divergence with low dimension. To study the classification performance by different wavelet, eight dimensional optimized feature combinations of db1~db10, coif1~coif5 and sym1~sym10 are extracted, respectively. It is observed that the divergence of coif2 and sym8 are obviously bigger than that of db2.BP neural network, which uses eight dimensional feature vector of db2 as input, is trained many times to confirm the nodes number in hidden layer and net weights.