摘要
针对心电信号中的室性早搏检测,利用经验小波变换获取心拍时频二维数据作为神经网络的输入,以避免传统一维数据输入造成的信息损失问题.通过二维卷积神经网络进行心拍的特征提取,在网络代价函数中引入了布雷格曼散度进行优化,以增强小样本训练时的特征识别能力.该算法模型在MIT-BIH心电数据库的网络训练与性能评估表明,代价函数情况下的总体敏感度与总体阳性检测率分别为96.39%、97.25%,均优于用传统交叉熵代价函数所检测的结果.通过实验验证该算法模型获得了良好的检测率与鲁棒性,具有较强的临床应用意义.
For the detection of ventricular premature beats in ECG signals,the empirical wavelet transform was adopted to obtain the time-frequency two-dimensional data of the heartbeat.Taking the data as the input of the neural network,it could avoid the information loss caused by the traditional one-dimensional data input.The feature extraction of heartbeat was performed by two-dimensional convolutional neural network.We introduced Bregman divergence optimization in the network cost function to enhance the feature recognition ability of small sample training.According to the established algorithm model,network training and performance evaluation were carried out in MIT-BIH ECG database.The results show that the overall sensitivity and the positive detection rate are 96.39%and 97.25%respectively in the condition of cost function,which are superior to detection results of traditional cross-entropy cost function.The experiment proves that detection algorithm has good detection rate and robustness,which is of strong clinical significance.
作者
吴义满
徐瑶瑞
Wu Yiman;Xu Yaorui(School of Medical Imaging,Jiangsu Vocational College of Medicine,Yancheng 224000,Jiangsu,China;Department of Equipment,Yancheng First People s Hospital,Yancheng 224000,Jiangsu,China)
出处
《计算机应用与软件》
北大核心
2019年第11期275-279,共5页
Computer Applications and Software
基金
江苏省科技厅自然科学基金项目(BK20151293)
2019年盐城市医学科技发展计划项目(深度学习在心电信号室性早搏检测中的应用)
关键词
室性早搏
经验小波变换
时频谱
卷积神经网络
布雷格曼散度
Ventricular premature beat
Empirical wavelet transforms
Time-frequency spectrum
Convolution neural network
Bregman divergence
