融合CNN和BiLSTM的心律失常心拍分类模型

为更加准确地从动态心电中提取异常心拍,设计一种融合卷积神经网络(CNN)和多层双边长短时记忆网络(BiLSTM)的心律失常心拍分类模型。心电信号首先被分割成0.75 s和4 s两种不同尺度大小的心拍信号,然后利用11层CNN网络和3层BiLSTM网络分别对小/大尺度心拍信号进行特征提取与合并,并使用3层全连接网络对合并特征进行降维,最后利用softmax函数实现分类。针对MIT心律失常数据库异常心拍类型分布不均衡的问题,采用添加随机运动噪声和基线漂移噪声的样本扩展方法,降低模型的过拟合。采用基于患者的5折交叉检验进行模型验证。MIT心律失常数据库116 000个心拍的分类结果表明:所建立的模型针对4类心拍(正常、房性早搏、室性早搏、未分类)的识别准确率为90.42%,比单独使用CNN(76.45%)和BiLSTM(83.28%)的模型分别提高13.97%和7.14%。所提出的融合CNN和BiLSTM的心律失常心拍分类模型,相比单一基于CNN模型或者BiLSTM模型的机器学习算法,有更好的异常心拍分类准确率。

Arrhythmia Beat Classification Model Based on CNN and BiLSTM

In order to extract the abnormal beats more accurately from the dynamic electrocardiograph (ECG), a deep learning model combining convolutional neural network (CNN) and bi-directional long short-term memory network (BiLSTM) was proposed in this study. Firstly, ECG signals were segmented into two types of time window lengths: a small-scale length of 0.75 s and a large-scale length of 4 s. Then, features were extracted from the small- and large-scale length ECG segments using an 11-layer CNN network and a 3-layer BiLSTM network, respectively. Finally, the extracted features were combined and were then reduced using a 3-layer fully connected network. In addition, two data enhancement methods by adding random motion noise and baseline drift were used to attenuate the influence of over-fitting due to the unbalanced data distribution. The proposed model was tested on the MIT arrhythmia database using a patient-based 5-fold cross-validation method, and its accuracy for classifying the 4 types (normal, atrial premature, ventricular premature and unclassified) on 116,000 heartbeats was 90.42%, which was 13.97% and 7.14% higher than the CNN model (76.45%) and BiLSTM model (83.28%), respectively. This study validated that the proposed model with combining CNN and BiLSTM reports higher accuracy than only using CNN or BiLSTM model when performing the abnormal beat classification task.