摘要
目的建立结合心率变异性(heart rate variability,HRV)和呼吸频率(respiratory rate,RESP)的驾驶疲劳识别机器学习模型并明确最优特征子集。方法2021年6-12月,从陆军军医大学招募20名年龄在20~30岁之间的健康男性志愿者参加疲劳驾驶试验。记录正常睡眠和睡眠剥夺后驾驶员在驾驶任务中的心电信号,提取18维疲劳相关HRV特征值,选择清醒与疲劳状态下存在差异的HRV特征结合RESP作为特征集。比较支持向量机(support vector machines,SVM)、K最邻近(k-nearest neighbor,KNN)、朴素贝叶斯(naive Bayes,NB)、决策树(decision tree,DT)和逻辑回归(logistic regression,LR)这五种经典机器学习方法,筛选最优特征子集,并建立疲劳识别模型。结果低频功率(low frequency,LF)与高频功率(high frequency,HF)的比值LF/HF、RESP、平均RR间隔(mean RR interval,Mean RR)、样本熵(sample entropy,SampEn)、去趋势波动分析(detrended fluctuation analysis,DFA)短期斜率DFAα1五个特征是疲劳识别的有效特征子集,在SVM中分类效果最好,疲劳识别的准确性、敏感性、特异性分别为87.03%、87.07%、87.13%。其中LF/HF和RESP是最为重要的驾驶疲劳识别指标,两个维度下各个模型的准确性均能达到80%以上,SVM与LR整体表现更好,准确性、敏感性、特异性分别为84.99%、85.13%、82.65%和84.43%、86.49%、82.02%。结论LF/HF、RESP是驾驶疲劳识别的有效特征,在基于HRV特征和RESP的驾驶疲劳识别中SVM与LR的整体表现优于其他模型。
ObjectiveTo establish a machine learning model for driving fatigue recognition based on heart rate variability(HRV)and respiratory rate(RESP),and identify the optimal subsets of features.MethodsFrom June 2021 to December 2021,20 healthy male volunteers between the ages of 20 and 30 were recruited from the Army Medical University to participate in the fatigue driving experiment.The electrocardiographic(ECG)signals under normal sleep and sleep deprivation during driving tasks were recorded,and 18-dimensional fatigue-related HRV feature values were extracted.The HRV features that differ between awake and fatigue states were selected,which were further combined with RESP as the feature set.Moreover,5 classic machine learning methods were adopted and compared,including support vector machines(SVM),k-nearest neighbor(KNN),naive Bayes(NB),decision tree(DT)and logistic regression(LR).The optimal feature subset was screened,and then a fatigue recognition model was established.ResultsThe ratio of low frequency power to high frequency power(LF/HF),RESP,mean RR interval(Mean RR),sample entropy(SampEn),and detrended fluctuation analysis short-term slope(DFAα1)constituted the effective feature subsets of fatigue recognition,which achieved the best classification effect in SVM,with an accuracy of 87.03%,a sensitivity of 87.07%and a specificity of 87.13%in fatigue recognition.Among them,LF/HF and RESP were found as the most important indicators of driving fatigue identification,with an accuracy of each model in the above 2 dimensions reaching more than 80%.In addition,SVM and LR showed better overall performance,with an accuracy,sensitivity and specificity of 84.99%,85.13%,82.65%for SVM and 84.43%,86.49%,82.02%for LR,respectively.ConclusionLF/HF and RESP are the effective features for driving fatigue recognition.In the driving fatigue recognition model based on HRV features and RESP,the overall performance of SVM and LR models is better than the other models.
作者
向洪义
张琼敏
王俊杰
王思平
廖志康
孙丽璐
李奎
赵辉
XIANG Hongyi;ZHANG Qiongmin;WANG Junjie;WANG Siping;LIAO Zhikang;SUN Lilu;LI Kui;ZHAO Hui(Department of Military Traffic Injury Prevention and Treatment,Army Medical Center of PLA,Chongqing,400042;College of Computer Science and Engineering,Chongqing University of Technology,Chongqing,400054,China;College of Management,Chongqing University of Technology,Chongqing,400054,China)
出处
《陆军军医大学学报》
CAS
CSCD
北大核心
2022年第13期1299-1306,共8页
Journal of Army Medical University
基金
军队医学科技青年培育计划拔尖项目(20QNPY031)
重庆市在渝院士牵头科技创新引导专项(cstc2019yszx-jscxX0002)。
关键词
军用车辆
驾驶疲劳
心率变异性
呼吸频率
机器学习
military vehicles
driving fatigue
heart rate variability
respiratory rate
machine learning
作者简介
通信作者:赵辉,E-mail:box.zhaohui@163.com。
