基于手机视频的帕金森病患者冻结步态的自动识别  

Automatic recognition of freezing of gait in Parkinson's disease based on mobile video

作　　者：李文丹 陈绣君 李蒙燕[3] 陈仲略 白红民 王嘉嘉 杜汉强 邹海强 Li Wendan;Chen Xiujun;Li Mengyan;Chen Zhonglue;Bai Hongmin;Wang Jiajia;Du Hanqiang;Zou Haiqiang(Graduate School,Guangzhou University of Traditional Chinese Medicine,Guangzhou 510006,China;Research Department,Gyenno Science Co.,Ltd.,Shenzhen 518057,China;Department of Neurology,Guangzhou First People's Hospital,Guangzhou 510180,China;Department of Neurosurgery,General Hospital of Southern Theatre Command of Chinese People's Liberation Army,Guangzhou 510010,China)

机构地区：[1]广州中医药大学研究生院,广州510006 [2]深圳市臻络科技有限公司研究部,深圳518057 [3]广州市第一人民医院神经内科,广州510180 [4]中国人民解放军南部战区总医院神经外科,广州510010

出　　处：《中华神经医学杂志》2022年第4期348-353,共6页Chinese Journal of Neuromedicine

基　　金：国家高技术研究发展计划(863计划)课题(2012AA02A514)。

摘　　要：目的通过手机拍摄PD伴冻结步态(FOG)患者的步态视频,构建基于手机视频的PD伴FOG患者的自动识别系统。方法选择广州市第一人民医院神经内科自2020年12月至2021年5月收治的49例PD伴FOG患者,收集患者的临床资料,使用手机拍摄患者"3 m往返"和"3 m往返通过窄道(长0.6 m)的过程,获得87个有效视频,标注每个视频的转身阶段、直行阶段及其中的FOG事件。提取视频中关键点的位置信号,对信号预处理后提取特征数据,由特征数据分别建立动作识别模型、直行FOG识别模型和转身FOG识别模型,最后组成端到端的FOG识别模型。采用留一法(LOSO)评估上述模型的性能。结果87个有效视频中25881个窗口样本,其中22066个非FOG窗口样本,3815个FOG窗口样本。LOSO法评估结果显示,动作识别模型的灵敏度为83.27%,特异度为91.38%,准确度为89.28%。直行FOG识别模型的灵敏度为57.69%,特异度为88.12%。转身FOG识别模型的灵敏度为61.54%,特异度为98.72%。端到端的FOG识别模型的灵敏度和特异度分别为85.71%、75.73%。结论基于手机视频的PD患者FOG自动识别系统具有较高的灵敏度和特异度,可实现远程识别PD患者的FOG,便于对PD伴FOG患者的筛查和随访。Objective To construct an automatic recognition system for PD patients with freezing of gait(FOG)based on mobile phone videos by recording the gait videos of PD patients with FOG.Methods Forty-nine PD patients with FOG,admitted to our hospital from December 2020 to May 2021,were chosen in our study.Their clinical data were collected.The processes of these patients accepted"3-meter-round trip"and"3-meter-round trip through narrow(0.6 m)"were recorded and 87 valid gait videos were extracted.Position signals of key points in the video were extracted,and featured data were extracted after signal preprocessing.From the featured data,action recognition model,straight FOG recognition model and turn FOG recognition model were established respectively,and finally end-to-end FOG recognition model was formed.Leave-one-subject-out(LOSO)method was used to evaluate the performance of the above models.Results A total of 22066 non-FOG window samples and 3815 FOG window samples were obtained from 87 valid videos,which constituted the training sample pool of this study.LOSO method showed that the motion recognition model enjoyed 83.27%sensitivity,91.38%specificity,and 89.28%accuracy;the straight FOG recognition model enjoyed 57.69%sensitivity and 88.12%specificity;the turn FOG recognition model enjoyed 61.54%sensitivity and specificity 98.72%;and the end-to-end FOG recognition model enjoyed 85.71%sensitivity and 75.73%specificity.Conclusion The automatic recognition system for PD patients with FOG based on mobile phone videos has relatively high sensitivity and specificity,which can realize remote assessment and is convenient for screening and follow-up of PD patients with FOG.

关 键 词：帕金森病 冻结步态 机器视觉 机器学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] R742.5[自动化与计算机技术—计算机科学与技术]