Accurate classification of ECG arrhythmia using MOWPT enhanced fast compression deep learning networks
Accurate classification of ECG arrhythmia using MOWPT enhanced fast compression deep learning networks
复制标题
DOI：
10.1007/s12652-020-02110-y
复制
发表时间：
2020-05-19
影响因子：
--
通讯作者：
Li, Yang
中科院分区：
文献类型：
--
作者：
Huang, Jing-Shan;Chen, Bin-Qiang;Li, Yang
心电信号的准确分类对于心脏病的自动诊断具有重要意义。为了实现心律失常的高精度智能分类，提出了一种基于快速压缩残差卷积神经网络（FCResNet）的智能心电分类器的准确分类方法。该方法利用最大重叠小波包变换（MOWPT）的时不变特性，将原始心电信号分解为不同尺度的子信号样本。随后，将五种心律失常类型的样本用作FCResNet的输入，以便识别和分类ECG心律失常类型。在FCResNet模型中，加入了快速下采样模块和多个残差块结构单元。所提出的深度学习分类器可以大大缓解计算效率低、难以收敛和模型退化的问题。通过单因素实验研究了FCResNet的参数优化。采用来自MIT-BIH心律失常数据库的数据集来测试所提出的深度学习分类器的性能。当快速下采样模块中的宽步长卷积个数为2，批量参数为20，选择MOWPT中低频子空间作为分类器的输入时，分类器的平均准确率达到98.79%。将这些分析结果与一些比较方法的分析结果进行了比较，验证了所提方法的优越性和增强性。
Accurate classification of electrocardiogram (ECG) signals is of significant importance for automatic diagnosis of heart diseases. In order to enable intelligent classification of arrhythmias with high accuracy, an accurate classification method based intelligent ECG classifier using the fast compression residual convolutional neural networks (FCResNet) is proposed. In the proposed method, the maximal overlap wavelet packet transform (MOWPT), which provides a comprehensive time-scale paving pattern and possesses the time-invariance property, was utilized for decomposing the original ECG signals into sub-signal samples of different scales. Subsequently, the samples of the five arrhythmia types were utilized as input to the FCResNet such that the ECG arrhythmia types were identified and classified. In the proposed FCResNet model, a fast down-sampling module and several residual block structural units were incorporated. The proposed deep learning classifier can substantially alleviate the problems of low computational efficiency, difficult convergence and model degradation. Parameter optimizations of the FCResNet were investigated via single-factor experiments. The datasets from MIT-BIH arrhythmia database were employed to test the performance of the proposed deep learning classifier. An averaged accuracy of 98.79% was achieved when the number of the wide-stride convolution in fast down-sampling module was set as 2, the batch size parameter was set as 20 and wavelet subspaces of low frequency bands in MOWPT were selected as input of the classifier. These analysis results were compared with those generated by some comparison methods to validate the superiorities and enhancements of the proposed method.