计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (8): 331-340.DOI: 10.3778/j.issn.1002-8331.2201-0385

• 工程与应用 • 上一篇    

视频脉搏特征的非接触房颤检测

张姁,杨学志,刘雪南,方帅   

  1. 1.合肥工业大学 计算机与信息学院,合肥 230009
    2.安徽省工业安全与应急技术重点实验室,合肥 230009
  • 出版日期:2023-04-15 发布日期:2023-04-15

Non-Contact Atrial Fibrillation Detection Based on Video Pulse Features

ZHANG Xu, YANG Xuezhi, LIU Xuenan, FANG Shuai   

  1. 1.School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China
    2.Anhui Key Laboratory of Industrial Safety and Emergency Technology, Hefei University of Technology, Hefei 230009, China
  • Online:2023-04-15 Published:2023-04-15

摘要: 房颤的早发现对心脑血管疾病的预防至关重要。提出一种面部视频房颤检测方法。该方法通过人脸跟踪和集成经验模态分解从面部视频中提取脉搏信号,根据房颤发作时的脉搏特性,从视频脉搏信号中提取房颤判别特征。设计一种改进的递归特征消除特征选择方法,筛选出对房颤检测较重要的特征,基于以上特征采用机器学习方法实现房颤检测。在采集的122例房颤病人和139例正常窦性心律面部视频上实验,最优特征为相邻RR间期大于50?ms的心搏百分比(PNN50)、RR间期的最大值(maxRR)、庞加莱图水平半径(SD2)等。基于以上最优特征集,房颤检测的准确率为92.31%,特异度是90.24%,灵敏度是94.59%,AUC是0.920 5。

关键词: 机器学习, 房颤检测, 面部视频, 特征选择

Abstract: Early detection of atrial fibrillation is very important for the prevention of cardiovascular and cerebrovascular diseases. This paper proposes a facial video atrial fibrillation detection method. The method extracts the pulse signals from the facial video by face tracking and improved complete ensemble empirical modes decomposition(ICEEMD), and extracts atrial fibrillation discriminative features from video pulse signals according to the pulse characteristics during atrial fibrillation episodes. An improved recursive feature elimination feature selection method is designed to screen out the more important features for atrial fibrillation detection. Based on the above features, machine learning methods are used to achieve atrial fibrillation detection. Experiments are conducted on 122 cases of atrial fibrillation patients and 139 cases of normal sinus rhythm facial videos. The optimal features are the percentage of heartbeats with adjacent RR interval greater than 50 ms (PNN50), the maximum value of RR intervals (maxRR), and the horizontal radius of poincare diagram (SD2) etc. Based on the above optimal feature set, the accuracy of atrial fibrillation detection is 92.31%, the specificity is 90.24%, the sensitivity is 94.59%, and the AUC is 0.920 5.

Key words: machine learning, atrial fibrillation detection, facial video, feature selection