摘要
提出一种基于卷积神经网络(CNN)的心电特征提取方法,以心拍R点前后75ms内的信号为QRS波群,提取QRS波群的180Hz下采样和RR间期,把QRS采样作为CNN的卷积层输入,把CNN的全连接层输出作为形态特征。对CNN特征进行分类能力测试,与文献方法对比结果表明,CNN特征在平均Se上有1.7%的提升,在平均P^+上有2.9%的提升,对于S类的Se提升7%,P^+提升9%,对V类的Se提升4%,P+提升2%;该方法在F类上性能低于文献最佳值,因为神经网络对训练样本数要求比支持向量机(SVM)高。
A method of using convolutional neural networks to extract ECG heartbeat features was proposed. 150 ms long signal segment around the fiducial point of heartbeat was used as the QRS complex, 180 Hz downsamples of QRS complex and RR in- tervals were extracted, QRS downsamples were fed to convolutional layer of a trained CNN, and the outputs of CNN＇s full-con- nected layer were extracted as CNN feature. The classification performance of CNN features was evaluated. Comparing with reference features, CNN features obtains average improvement of 1.7% in Se and 2.9% in P＋, and improvement of 7% in Se and 9% in P＋ for S class, 4% in Se and 2% in P＋ for V class. The proposed method loses performance for F class, because neu- ral networks demand more training instances than SVM.
出处
《计算机工程与设计》
北大核心
2017年第4期1024-1028,共5页
Computer Engineering and Design
基金
国家自然科学基金项目(61271423)
关键词
卷积神经网络
心电节拍分类
特征提取
模式识别
形态特征
convolutional neural networks
heartbeat classification
feature extraction
pattern recognition
morphology features
