摘要
心电信号的ST段波形变化是心肌损伤等心血管类疾病临床诊断的重要辅助手段之一。针对ST段波形分类以及深度卷积神经网络过拟合问题,提出一种基于概率随机舍弃神经元建立子网络的Dropout深度卷积神经网络,通过心电信号数据去噪、ST段候选段筛选、神经网络卷积与下采样运算过程,实现ST段波形样本训练与测试。仿真实验对比分析了算法的波形分类准确率、卷积核个数影响和Dropout对算法泛化能力影响,与专家手工标注、BP、RNN和DCNN等方法进行比较,实验结果表明Dropout DCNN能够有效提高卷积神经网络泛化能力,提升算法的可用性。
ECG data analysis is one of the important auxiliary means for clinical diagnosis of cardiovascular related diseases. ST segment waveform changes can assist in judging myocardial injury and other diseases. Aiming at the problem of ST segment of ECG data waveform classification and deep convolutional neural network over-fitting,this paper proposes a Dropout deep convolutional neural network through ECG signal denoising,ST segments screening and down sampling to realize the waveform sample training with the test of ST segment. Waveform classification accuracy,the effect of convolution kernel and Dropout effect on the generalization ability are analyzed by the comparison of simulation results with expert manual annotation,BP,RNN and DCNN. The experimental results show that Dropout DCNN compared to the full depth of convolutional neural network can improve the generalization ability of neural network and enhance the availability of the algorithm.
作者
任晓霞
REN Xiaoxia(School of Mathematics and Information Science, Zhangjiakou University, Zhangjiakou Hebei 075000, China)
出处
《传感技术学报》
CAS
CSCD
北大核心
2018年第8期1217-1222,共6页
Chinese Journal of Sensors and Actuators
基金
江苏省政策引导类计划(产学研合作)-前瞻性联合研究项目(BY2016049-01)
