摘要
作为模式识别研究的典型应用,计算机辅助心电图分析在可穿戴终端及面向基层的"云"服务平台具有重要价值.本文首先说明了面向临床的心电图分类模型的复杂性,接着从非线性函数拟合能力角度分析了现有的特征提取算法和分类算法,并采用深层学习构造心电图分类模型.针对多导联心电图这种特殊的二维结构,提出导联卷积神经网络,并利用"平移起始点"和"加噪"增加训练样本数.通过15万多条记录的测试数据,取得了准确率为83.66%和AUC为0.9086的成绩.最后我们还移植分类模型到移动终端设备,其实时分析结果满足应用需求.
As one of the classical applications of pattern recognition research, electrocardiogram（ECG） classification has important application values for wearable ECG devices and ＂cloud＂ service platforms. In this paper,first of all, the complexity of the ECG classification model for clinical application is illustrated. Consequently,the approximation ability of a nonlinear function in an existing feature extraction and classification algorithm is analyzed, and deep learning is employed for ECG classification. Then, lead convolutional neural networks（LCNN）is presented considering the special two-dimensional structure of multi-lead ECG, in which ＂translating starting point＂ and ＂adding noise＂ are two of the main strategies to increase the training sample. Tests conducted using more than 150,000 ECG records show that the proposed method has an accuracy of 83.66% and 0.9086 AUC.Finally, the classification model is implemented on a mobile terminal, where its real-time analysis performance is shown to meet application requirements.
出处
《中国科学：信息科学》
CSCD
北大核心
2015年第3期398-416,共19页
Scientia Sinica(Informationis)
关键词
深层学习
心电图(ECG)
非线性拟合
卷积神经网络
数据分布与分类
deep learning
electrocardiogram（ECG）
nonlinear approximation
convolutional neural networks
data distribution and classification
