ECG Classification Based On Feature-extraction And Neural Network

In this study, according to the feature extraction and neural network classification, an ECG classification method simulating the real world situation is presented: One hand, a modified approach of linear approximation distance thresholding (LADT) algorithm has been studied and complemented. This complement avoid the case that in the common LADT algorithm, as the endpoints of the approximating segment sometimes may not be properly determined on the ECG signals, the saw-tooth like approximation is sure to come out. Thus the quality of the approximation is enhanced. And then, through analyzing the segments, the QRS complexes are detected and then the features of the ECG signals are obtained. The other hand, as the ECG waves have different frequencies, the Mexican-hat wavelet-transform is adopted to detect the character points. According to the particularity of the Mexican-hat wavelet-transform, the ECG character points are just corresponding to the local extremes of the signals transformed. It overcomes the complexity that if transformed with spline wavelet, the character points are only corresponding to the zero-crossing points of the modulus maximum pairs. And thus the processes of detections can be simplified and this also improves the detection performance that the correct rate of R detection achieves to 99.9%. After detecting the character points of the ECG signals, according to the theory of the ECG analysis and the real world situation, with the splendid capability of classification of RBF network, the ECG signals are classified in a high dimensions space. The tests with some ECG signals of MIT-BIH show that after training by the features extracted first, the correct rates of classification are excellent. As the correct rates from the article [6] are less than 97% to the training waves and 54% to the untrained waves, the classification correct rate of the training waves is 100% in this study. As to the untrained waves, the correct rates are a little different. To the waves whose features are extracted by the LADT, the correct rate is 78.2%; to the waves whose features are extracted by the wavelet-transform, the correct rate is 86.6%.