基于弱监督学习卷积神经网络的心脏按压评估  

Assessment of chest cardiac compression in convolutional neural network based on weak supervised learning strategy

作　　者：鲍宇[1,2] 殷佳豪 刘世杰 杨轩 朱紫维 Bao Yu;Yin Jiahao;Liu Shijie;Yang Xuan;Zhu Ziwei(College of Computer Science and Technology, China University of Mining and Technology, Xuzhou 221116, China;Mine Digitization Engineering Research Center of Ministry of Education of the People's Republic of China, Xuzhou 221116, China)

机构地区：[1]中国矿业大学计算机学院,徐州221116 [2]矿山数字化教育部工程中心,徐州221116

出　　处：《仪器仪表学报》2019年第5期203-212,共10页Chinese Journal of Scientific Instrument

基　　金：国家重点基础研究发展计划(973计划)(2013CB227900);徐州市应用基础研究项目(KC17073)资助

摘　　要：传统的基于加速度波形积分的心脏胸外按压评估方法受噪声和积分时延影响,在计算距离时存在较大误差,评估效果不理想。为此,在弱监督学习策略与波形分割的基础上,提出了一种基于一维卷积神经网络的心脏按压加速度波形识别算法,实验结果表明,一维卷积神经网络达到了99.4%的正确率,明显优于传统的积分方法和BP神经网络算法。进一步采用Grad-CAM方法对评估结果进行可视化分析,发现卷积神经网络所关注的特征集中于开始按压至按压到达平衡位置,以及此次按压松手后反向加速度达到最大值至下一次按压开始这2个阶段的加速度波形变化情况。此外该评估模型不再需要对按压距离进行精确测距,因而不受按压遮挡、电磁波干扰等因素的影响,可以实时检测按压是否规范有效,在复杂环境中也具有较高的鲁棒性,在医疗急救领域中具有一定的实用价值。The traditional evaluation method of cardiac chest compression is based on acceleration waveform integration, which is affected by noise and integral delay. A large error is introduced in calculating the distance and the evaluation effectiveness is not ideal. Therefore, based on the weak supervised learning strategy and waveform segmentation, this paper proposes one kind of acceleration waveform recognition algorithm for cardiac chest compression based on one-dimensional convolutional neural network. Experimental results show that the one-dimensional convolutional neural network achieves 99. 4% accuracy, which is significantly better than the traditional integration method and BP neural network algorithm. Further, the Grad-CAM method is adopted to visually analyze the evaluation results. The features of convolutional neural network focus on the acceleration waveform changes in the two compression stages of starting to press until the pressure can reach the equilibrium position. The reverse acceleration can achieve the maximum value after the pressing of the hand to the next press start. In addition, the evaluation model does not need to accurately measure the pressing distance. Thus, it is not aflected by factors such as pressing occlusion and electromagnetic wave interference. Its effectiveness can be checked in the real-time manner. It also has the feature of high robustness in complex environment and has certain practical value in the field of medical emergency.

关 键 词：加速度波形评估 一维卷积神经网络 脉冲识别与波形分割 弱监督学习策略 一维Grad-CAM可视化 

分 类 号：TP306[自动化与计算机技术—计算机系统结构] TH825[自动化与计算机技术—计算机科学与技术]