基于半监督卷积神经网络的心律不齐分类

Arrhythmiaclassification based on semi-supervised convolutional neural network

心律不齐早期的检测和分类有十分重要的价值。然而,由于心律不齐特征不够明显、不同人体心电信号存在较大的差异性等问题,心律不齐分类技术在现实中的应用存在很大困难。针对上述问题,提出了一种基于半监督卷积神经网络的心律不齐分类方法。首先,对心电信号进行马尔可夫变迁场变换将其变为二维图像;然后,迁移二维卷积神经网络训练得到普通模型;最后,使用半监督训练的方法得到特定模型,使用得到的特定模型对特定病人的心电信号进行分类。该方法从二维图像中提取更高层次的特征来提高分类性能,并有效地克服不同人体心电信号之间的差异性,增强了模型的泛化能力。遵循美国医疗器械促进协会标准,使用麻省理工学院和波士顿贝丝以色列医院心律不齐数据库中的数据,完成了心律不齐五分类,总准确率达到了97.9%,提升了心律不齐分类技术的现实应用价值。

Early detection and classification of arrhythmia is of great value.However,due to the lack of obvious characteristics of arrhythmia and large individual differences,the application of arrhythmia classification technology in reality is very difficult.To solve above problems,a method for arrhythmia classification based on semi-supervised convolutional neural network was proposed.First,markov transition field transformation was used on the ECG signal to transform it into two-dimensional image;then,a common model was obtained by transfer learning;finally,a specific model was obtained by the semi-supervised training,which was used to classify the ecg signal of the specific patient.This method extracted higher-level features from two-dimensional images to improve classification performance,and overcome the differences between different human ecg signals effectively,the generalization ability of the model was enhanced too.Followed the standards of the association for the advancement of medical instrumentation,and used the data from the arrhythmia database of the Massachusetts Institute of Technology and Boston Beth Israel Hospital.The data was classified to five classes,with a total accuracy rate of 97.9%,The practical application value of arrhythmia classification technology is enhanced.