ECG beat classification based on discriminative multilevel feature analysis and deep learning approach

Extraction of significant features from Electrocardiogram (ECG) signal is the primary concern for accurate diagnosis of cardiac arrhythmia. This work presents a novel approach of multilevel feature analysis and deep learning strategy for efficient ECG beat classification. The multilayer characteristics of ECG signal obtained from Empirical mode decomposition (EMD) are explored to extract discriminative feature vectors. The multilayer similarity coefficients are obtained by applying Dynamic time warping (DTW) metric and Pearson correlation coefficient (PCC) as diagnostic features. Furthermore, discrete orthonormal Stockwell transform (DOST) is employed for time-frequency representation of ECG data in multilayer aspect. The sublet changes in time --frequency spaces due to the presence of cardiac abnormalities are captured by estimating various nonlinear parameters. Interlayer deviations of these nonlinear parameters are estimated as the significant characteristics of arrhythmia detection. In addition, this study shows that the phase synchrony (PS) coefficients are prominent index for quantifying the crucial phase variation between normal and abnormal heart conditions. Hence multilevel PS coefficients are employed as the predictors of arrhythmia detection. Finally, the extracted feature vectors are fed to various classifiers to identify the heart anomalies. The proposed technique attains average accuracy of 98.82% and 98.14% using support vector machine (SVM) and k-nearest neighbors (k-NN) classifier respectively. The improved classification accuracy of 99.05% is obtained with the strategy of combining deep neural network (DNN) with the proposed feature extraction policy. Present work delivers satisfactory and su-perior performances for arrhythmia classification compare to other existing approaches.