| Abstrakt: |
In this paper, we propose a method of extracting features from ECG signals based on the sparse representation of double-threshold Stagewise orthogonal matching Pursuit. First, the ECG signal was de-noised and then the K singular value decomposition algorithm (KSVD) was used to iterate the ECG data set to obtain a supercomplete dictionary of ECG signals. Then, through the double-threshold Stagewise orthogonal matching Pursuit, a sparse atomic matrix with the best reconstruction effect is finally selected to obtain the characteristics of ECG signals. Normal, atrial fibrillation, and ventricular fibrillation ECG signals from MIT ’ s ECG Signal Database (MIT-BIH) were used to evaluate the proposed method. Experimental results show that in the binary classification experiment, the comprehensive recognition rate of the proposed algorithm for normal ECG signals, ventricular fibrillation and atrial fibrillation is 93.96%, 1.26% higher than that of CNN, and 3.99% higher than that of wavelet transform. Meanwhile, the five parameters of the proposed algorithm, such as comprehensive Kappa and comprehensive root relative square error, are the best. In addition, this paper carries out three-classification recognition of normal, atrial fibrillation, and ventricular fibrillation signals. The average recognition rate of the proposed algorithm among the three classifiers is 89.17%. The average recognition rate is 3.2% and 6.67% higher than that of convolutional neural networks and wavelet transform, respectively. The experimental results show that the features extracted in this paper have high recognition performance for arrhythmia ECG signals. [ABSTRACT FROM AUTHOR] |
