Li Duan, Zhang Hongxin, Liu Zhiqing, Huang Juxiang, Wang Tian
School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, P.R.China;School of computer and communication Engineering, Zhengzhou University of Light Industry, Zhengzhou 450002, P.R.China.
School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, P.R.China;Beijing Key Laboratory of Work Safety Intelligent Monitoring, Beijing 100876,
Sheng Wu Yi Xue Gong Cheng Xue Za Zhi. 2019 Apr 25;36(2):189-198. doi: 10.7507/1001-5515.201712031.
Electrocardiogram (ECG) signals are easily disturbed by internal and external noise, and its morphological characteristics show significant variations for different patients. Even for the same patient, its characteristics are variable under different temporal and physical conditions. Therefore, ECG signal detection and recognition for the heart disease real-time monitoring and diagnosis are still difficult. Based on this, a wavelet self-adaptive threshold denoising combined with deep residual convolutional neural network algorithm was proposed for multiclass arrhythmias recognition. ECG signal filtering was implemented using wavelet adaptive threshold technology. A 20-layer convolutional neural network (CNN) containing multiple residual blocks, namely deep residual convolutional neural network (DR-CNN), was designed for recognition of five types of arrhythmia signals. The DR-CNN constructed by residual block local neural network units alleviated the difficulty of deep network convergence, the difficulty in tuning and so on. It also overcame the degradation problem of the traditional CNN when the network depth was increasing. Furthermore, the batch normalization of each convolution layer improved its convergence. Following the recommendations of the Association for the Advancements of Medical Instrumentation (AAMI), experimental results based on 94 091 2-lead heart beats from the MIT-BIH arrhythmia benchmark database demonstrated that our proposed method achieved the average detection accuracy of 99.034 9%, 99.498 0% and 99.334 7% for multiclass classification, ventricular ectopic beat (Veb) and supra-Veb (Sveb) recognition, respectively. Using the same platform and database, experimental results showed that under the comparable network complexity, our proposed method significantly improved the recognition accuracy, sensitivity and specificity compared to the traditional deep learning networks, such as deep Multilayer Perceptron (MLP), CNN, etc. The DR-CNN algorithm improves the accuracy of the arrhythmia intelligent diagnosis. If it is combined with wearable equipment, internet of things and wireless communication technology, the prevention, monitoring and diagnosis of heart disease can be extended to out-of-hospital scenarios, such as families and nursing homes. Therefore, it will improve the cure rate, and effectively save the medical resources.
心电图（ECG）信号很容易受到内部和外部噪声的干扰，并且其形态特征在不同患者之间存在显著差异。即使对于同一患者，在不同的时间和身体状况下其特征也会有所变化。因此，用于心脏病实时监测和诊断的心电图信号检测与识别仍然具有挑战性。基于此，提出了一种结合小波自适应阈值去噪和深度残差卷积神经网络算法的多类心律失常识别方法。利用小波自适应阈值技术对心电图信号进行滤波。设计了一个包含多个残差块的20层卷积神经网络（CNN），即深度残差卷积神经网络（DR-CNN），用于识别五种类型的心律失常信号。由残差块局部神经网络单元构建的DR-CNN缓解了深度网络收敛困难、调优困难等问题。它还克服了传统CNN在网络深度增加时的退化问题。此外，每个卷积层的批量归一化提高了其收敛性。按照医学仪器促进协会（AAMI）的建议，基于来自麻省理工学院-贝斯以色列女执事医疗中心（MIT-BIH）心律失常基准数据库的94091个两导联心跳的实验结果表明，我们提出的方法在多类分类、室性早搏（Veb）和室上性早搏（Sveb）识别方面的平均检测准确率分别达到了99.0349%、99.4980%和99.3347%。使用相同的平台和数据库，实验结果表明，在可比的网络复杂度下，与传统深度学习网络，如深度多层感知器（MLP）、CNN等相比，我们提出的方法显著提高了识别准确率、灵敏度和特异性。DR-CNN算法提高了心律失常智能诊断的准确性。如果将其与可穿戴设备、物联网和无线通信技术相结合，心脏病的预防、监测和诊断可以扩展到院外场景，如家庭和养老院。因此，它将提高治愈率，并有效节省医疗资源。
