基于CNN和频率切片小波变换的T波形态分类

心电实时监控是心血管疾病防治的重要手段.心电图中T波的变化是心肌缺血和心脏猝死等疾病的重要表征,T波形态自动识别是心电远程监控中一个重要问题.由于实时监护用心电的强噪声背景影响,传统的T波特征提取与分类算法遭遇瓶颈.提出一种结合切片频率小波变换和卷积神经网络的T波形态识别算法,包括:自动定位R波波峰位置与T波终点位置,...

T-Wave Morphological Classification Based on CNN and Modified Frequency Slice Wavelet Transform

Real-time monitoring of ECG is one important means of cardiovascular disease prevention. T-wave is an important characteristic of diseases such as myocardial ischemia and sudden cardiac death. The automatic identification of T-wave is a challenging task in ECG remote monitoring. Due to the influence of high noise background of real time monitoring ECG,the conventional T-wave feature extraction and classification algorithm encounters a bottleneck. In this paper,a T-wave morphological recognition algorithm combining slice frequency wavelet transform and convolutional neural network was proposed. The algorithm included: locating automatically the R-wave peak’s position and the T-wave ends’ position to identify a segment containing the Twave;the frequency slice wavelet transform was performed,and the generated time-frequency image was input into the convolutional neural network to complete the classification of the T-wave. The frequency slice wavelet transform transformed the signal to the time-frequency domain,which accurately presented the time-frequency energy distribution characteristics of the ECG signal. The hidden layers of the convolutional neural network completed the three features’ extraction of the time-frequency image by convolving,activating and pooling the time-frequency image three times. These features have translation and scaling invariance. In this paper,12 830 fragments in European ST-T database was used. The convolutional neural network model was trained and tested by the 3-fold cross validation method. The classification accuracy of experiment based on heart beats reached97. 34%,and the F1 measure reached 96. 97%. The classification accuracy of experiment based on samples was84. 80%,and the F1 measure was 83. 29%. The classification accuracy of the model tested in QT database was87. 83%,F1 measure was 85. 38%,and the generalization performance was good. Compared with other T-wave classification algorithms(such as decision tree,support vector machine,etc.),the classification accuracy based on heart beat experiments was improved by 1 ~ 5%. The results demonstrated that the algorithm designed for the classification of six types of T-wave improved the accuracy and performed well in terms of robustness and generalization performance. In addition,the algorithm model was also applicable to the analysis of other physiological signals and has certain guiding significance in the field of medical image analysis.