Analysis to Evaluate Levels of Sleepiness_论文


2023年12月20日发(作者:rev是什么意思)

Journal of Communication and Computer 10(2013)585-592 aLI5HIN羁 C A ustomization of Wavelet Function for Pupil Fluctuation nalysis to Evaluate Levels of Sleepiness Ruben.Dario Pinzon—Morales and Yutaka Hirata Department ofComputer Science,Chubu University,Kasugai 1200,Japan Received:May 3i,2012/Accepted:July 4,2012/Published:May 31,2013 Abstract:This paper proposes a method to customize a wavelet function for the analysis of pupil diameter fluctuation in the detection of drowsiness states under a driving simulation.The methodology relies on a genetic algorithm-based optimization and lifting schemes,which are a flexible and fast implementation of the discrete wavelet transforl/1.To customize the wavelet function a clustering separability metric is employed as a fitness function so that the feature space created by the wavelet analysis exhibits the maximum class separability favorable for classification.Therefore,a completely new wavelet function is createdhaving unique ,characteristics customized to pupil diameter fluctuation analysis.It is demonstrated that the customized wavelet function own distinguished frequency and temporal responses suitable speciifcally for pupil diameter fluctuation analysis(namely, application-dependent),and in the classiifcation they outperform classical wavelet families including Daubechies,Coilfet and Symlet. which are assumed to be application-independent.Thus the proposed method is useful for analysis of pupil fluctuation in evaluating sleepiness levels.as has been demonstrated in other applications. Key words:Wavelet function customization,evolutionary optimization,pupil diameter fluctuation,drowsiness detection. 1.Introduction Wavelet analysis,owning to its capability for During time——frequency analysis with the wavelet transform,the input signal is projected into an orthogonal or bi-orthogonal space created by the scaling and translation of one single wave——like function,called wavelet mother.It is comprehensible representation of non-stationary biological signals, has inspired several methodologies in applications such as recognition of the nuclei of the basal ganglia from their neural spike trains【1],identification of epileptic seizures and classification of drowsiness that the successful representation of the signal mainly depends on the proper selection of a wavelet mother. Although there are plenty of wavelet prototypes in the literature,there is not an established rule that states which wavelet should be used for each application. Instead.it is a usual task for the researcher to test more or less arbitrary different wavelet functions to states out of EEG[2,3],recognition of hand movements by using EMG activity【4]among others. More recently,particular attention has been given to customized wavelet analysis that encloses the design of new wavelet mother functions[1,5-10].Here, taking advantage in recent advances in wavelet customization, a methodology or designing fifnd the one suited orf the data to be analysed[2].For this reason several methods have been proposed to create new wavelet with desire features with either deterministic or evolutionary approaches.On one application-dependent wavelet functions via GAs (genetic algorithms)are applied for evaluation of PD (pupil diameter)time series aiming at drowsiness assessment[1,5]. COrresp0nding author:Yutaka Hirata,Ph.D.,professor hand,deterministic methods have been shown to be burdensome due to the complexiy of matthematical conditions on orthogonality,symmetry,compactness, and smoothness【6].On the other hand,evolutionary research ifeld:neurosciences.E—mail:yutaka@isc.chubu.ac.jp. 

586 Customization of Wavelet Function for Pupil Fluctuation Analysis to Evaluate Levels of Sleepiness Table 1 Equations involved in the proposed method = :i=1….,Ⅳ}∈RnxN 人= ,:i=1,…,N}∈Z = Input set with observations each ofN samples(1) Label vectorink classes. indicatesmembership ofx (2) ^, 。, = (2 1) : w 『_1) Even and odd samples ofx.(0)indicates the original space of X(3) Multiresolutional analysis・Even and odd samples of w (4) ?=w w = (i- 一p= ,: 1,…,N p j = ,:i=1,…,Ⅳ }Lifting scheme prediction and update filters(5) ∑ +1 .A t-,t+) ,)) ,WaveIet detail c。eif nts at 1eVe1 l(6) waVeI1 aelpproximation coefifcients at 1eVe1 1(7) Symmetrical and linear phase constraint(8) /2 w (1- +∑N u/2 2+l“,w P,=P(.J+1),u,= (_,+1),i=1 .,N p∑ ~p,=1/2,∑ 一“ =1/4 Z= :i=1,..., )∈R (人)= = : ( )=f,i:1….,c)∈R加 N。ITIIalization constraint(9) Matrix。ffeatureswithddimensi。ns(10) Mat 。fcluster centres(11) s,= ∑ lIz II2] d = 一V 1 12'i≠ 。舢(人,z)= 1 k Withinj-th cluster scatter value(12) Mink。wski distance between clusters(13) max ,:.: )Davies-Bouldin separability index(1 4) Avcrage n。nlinear energY,whereE{.)isthe expectati。n05) Standard deviati。n。fthe abso1utes Values,whcr。∥=E{X)(16) l=E 一(x(f_1) ( ))} :=√E{(7 J.∥) } =√E } Root mean square value,or quadratic mean(17) intra-cluster compactness. To evince the benefits of the customized wavelet approaches,such as GAs,CAs(cultural algorithms), and AS(ant systems),have been presented to have newsworthy features for wavelet design.Therefore.in the present study,an evolutionary methodology for creation of a customized wavelet function is employed. function,the application under consideration in this study is the analysis of fluctuation of PD.The pupil that determines eye’s optical properties including depth of focus,and spatial frequency of the retinal image is controlled by two kinds of antagonistic muscles,the dilator and the sphincter,which are each innervated by two kinds of the autonomic nervous system,the sympathetic and the parasympathetic. The main motivation behind.additiona1 to avoiding the use of a generic function.is to create a wavelet function that achieve maximum class separability.In other words,given a feature space Z∈R ,where d is the number of features and the number of signals analyzed that are discriminated in k classes or clusters,and generated by wavelet decomposition These two muscles interact with each other to secure the proper acquisition of visual images and therein information to the brain.This co.action of the two (1)-(7),the methodology attempts at generating a wavelet function that maximize the separability of kinds of smooth muscles has been described as highly classes under a desired cost function D,so that the feature space yielded by the wavelet function presents non—stationary as well as non-linear[1 l】.What is more,the autonomic nervous system whose activity is characteristics bound towards classiicatfion such as known to change in parallel with drowsiness states 

Customization of Wavelet Function for Pupil Fluctuation Analysis to Evaluate Levels of Sleepiness 587 may be also evaluated in PD time series【1 2]. This paper is organized as follows:In Section 2 the theory basis of the methodology is introduced;Then in Section 3.2 a novel wavelet function is produced for PD fluctuation analysis;Section 3.3 evinces classiifcation performance with the newly customized wavelet function and for the sake of comparison classical wavelet functions are also evaluated;Finally, conclusions and ending remarks are drawn in Section 4. 2.Methods 2.i wavelet Customization The precise procedure for optimization of wavelet functions has been reported elsewhere[1].Therefore, only a summarized version is presented here referring to equations given in Table 1. First,extract a subset of training signals Xe∈X. It has been shown that a 30%sample of the patterns available provides enough info・rmation for the methodology to converge[1].Second,each pattem is decomposed with the wavelet transform up to level, by using cascade litfing schemes(3 to 7).Litfing schemes are a fast and flexible implementation of the discrete wavelet transform comprised by a prediction iflter P and an update filter with order N口 and ,respectively Eq.(5).Prediction filter is associated to the wavelet function in the classical wavelet scheme,and is responsible for extracting high rfequency components in the signal by eliminating low order polynomials,while the update operator is related with the scaling function and the extraction of approximation coefifcients.Multiresolutional analysis up to level,is achievable by taking sequential decomposition on the approximation coefifcients Eq. (4),similar as in the wavelet discrete transform case. Once the wavelet transform is computed,a matrix W of wavelet coefifcients is yielded.The dimension of such matrix happens to be heavy load for classiifcation.Indeed,for a signal with N samples, performing a,scale wavelet analysis produces a matrix W discriminated in,+1 frequency bands.To reduce the dimension of the wavelet representation a quantitative vector Eq.(1 0)is proposed and constructed by a feature extraction transformation,i.e.,Q: Z.The set of metrics that assemble = :i=1….d} is entirely application ̄lependent.Such metrics are to be introduced for the application under consideration in Section 3.2 At this point a major step of the methodology must be set,namely,the cost function in the optimization procedure.Owing to the focus of the present content,classiifcation of drowsiness states,the cost function is a measure that allows the evaluation (_of classification potential of Z without depending upon the classiifer itself,which could increase considerably the procedure time and by which the methodology may also decrease generalization among classifiers.Consequently,the Davies-Bouldin index is employed(10 to 14)as the cost function because the B is a measure of classes separability that assesses not only the compactness of each class Eq.(1 2)but also the global dispersion between classes Eq.(1 3).Therefore,by employing the Davies-Bouldin index given a crisp partition of the data人.i.e. DDB(人,z)the iflters P and are expected to be 0 > o J ∞ ∞ 0 C 0 ∽ Time(minutes) Fig.1 PD fluctuation signal during the monotonous driving situation simulation of atypical subject(top)and reported self-sleepiness level(bottom). guI .Q. 

588 Customization of Wavelet Function for Pupil Fluctuation Analysis to Evaluate Levels of Sleepiness evolved to maximize the separability among classes in 2.3 Database Using custom made scripts,PD was calculated from the feature space Z. The terminating step of the procedure is the the video recordings.and errors in the PD detection due to blinks and saccades were eliminated evolutionary optimization of the lifting scheme filters Eq.(5)by using GAs under constraints Eqs.(8)and(9) The constrains secure the linear phase,symmetry, compact support,and normalization of lifting filters, in other words,the associated wavelet and scaling functions.Furthermore,due to Eqs.(8)and(9)the GAs only evolve N /2+N /2—1 values. 2.2 Driving Simulation Experiment In a custom made driving simulator,a subject sat comfortably on the driver’s seat equipped with a steering wheel,brake and accelerator pedals(Logicool PRC—l 1000、in a dark room.The subject wore a goggle(Newopto ET-60一L)with 2 CCD cameras each of which takes infrared images of each eye at the rfame rate of 29.97 fps(NTSC).Brightness and contrast of the projector that presents driving simulation images were adjusted for each subject so that the initial PD became in its intermediate size (around 6 mm).The simulation image generated by a PC is projected onto the screen at 2.64 m away from the subject’s eye.The horizontal and vertical visual angles of the image are±12.1 and±10.7 deg. respectively.Subjects were instructed to follow the car in front,driving straight on the straight road at the maximum speed(monotonous driving situation).Only the white lane markings and texture of lawn on the shoulder of the road move radially backward depending on the speed of the car during the simulation.Every two minutes the subj ect was asked to inform his/her sleepiness level according to the following three scales 0:awake.1:not sure if sleepy or not. 2: sleepy. Subject recruitment and experimental procedures for this study conformed to the Declaration of Helsinki and all approved by the human research committee of Chubu University.All subjects had given their informed consent. automatically[1 2].The eliminated periods of data were interpolated by a cubic function.Then 2.minutes segments were extracted and categorized according to the self_sleeDiness leve1. 3.Results 3.1 PD during Driving Simulation A total of 3 1 healthy subjects fmean SD:22.8± 7.1 years old)participated in the experiment.Fig.1 illustrates the change in PD during the monotonous driving situation simulation of a typical subject with a relatively steady state until 1 0 min followed by gradual decrease in average PD until 1 3 min and large low frequency fluctuation after that.In total there are 1 50 segments belonging to state 0 labelled as class A, 98 of state 1 labelled as class B.and 131 of state 2 referred to as class C.The classificati0n task under consideration is a two-class problem given by class A and C(七=2 classes,N=3596 samples,n:281 segments).Class B has not been included in the present study because the boundaries of awake and sleepy are sometimes hard to decide subjectively. 3.2 New Wavelet Functionfor PD Analysis In the customization of the new wavelet function ofr the PD data a major step is the GA——based optimization.Regarding the GA,the following five parameters must be selected:(1)the arithmetic operator; (2)the mutation operator;(3)the population scale;(4)the number of generations and (5)the bounds of iteration.Parameters(1)to(4)can be set directly from literature【l】.Besides,the arithmetic crossover and no uniofITR mutation operators are employed as recommended in Ref.[1]. For the sake of simplicity,the population scale is set tO be 30.whereas the number of generations is set 

Customization of Wavelet Function for Pupil Fluctuation Analysis to Evaluate Levels of Sleepiness 589 equal to 20.The iteration parameters of the GA are selected to range within the interval[-0.5,0.5],which are possible values for the predictor and update coefifcients that meet the normalization constraints Eq. (9).The number of parameters to be optimized during the GA procedure for Np= =6 four The wavelet decomposition level is set to six so that the frequency band spanned by ,w and w (approximation coefifcient at level six,detail coefifcients of level six and five),reside within the interval【0,1]Hz where discriminant information related with respiratory components has been identiifed for pupil diameter[13].The set of metrics that composed Z are three,and given that only three wavelet spaces are considered,i-e.,w ’w ’,,and ( =9):the average non-linear energy,also known as Teager’s operator Eq.(1 5),the root mean square value Eq.(1 6),and the variability of absolute coe师cients Eq.(1 7).Such metrics were selected among a set of statistical moments employed in previous works【1】.Only three metrics are employed to allow the construction of threc}_ imensional feature spaces and hence to allow its visual inspection.To avoid local minima in the GA convergence,above steps are repeated ten times with random initializations.Thus,ten new wavelet functions are generated.The best mother wavelet is selected as the one with the largest value Eq.(1 4),for the current application the best one corresponded to the operators “:f.0,033.0,039 0,322 0,322—0,039-0,033】,and p:[0,013—0,027 0,514 0,514-0,027 0,013】・ Fig.2 depicts the new wavelet function(a and b) that was calculated from u and P,respectively by using the reverse decomposition method[1 41,along with two classical wavelets,Daubechies and Symlet of order four and six,respectively(c).As can be seen the main advantage of the proposed methodology is that the symmetyr and compact support of the customized wavelet can be secure by employing the constrained optimization in the GA.Additionally,it is important to notice that since the training set X e is composed of actual PD data,the temporal and rfequency response of u and P must have been customized to extract features in PD fluctuation, whereas the classical wavelets are non-symmetrical and do not exhibit major specialization on temporal and rfequency characteristics ofPD fluctuation[5]. 3.3 Classiifcation Performance To evaluate the performance of the new wavelet function,classiifcation scores of sleepy and awake states classes C and A respectively,are calculated. Since a classiifer method such as support vector machines or neural networks may increase the discrimination aptitude of z,here a basic linear Bayes classifier is used so that the classification rates will not be enhanced by the classifier.Four classical wavelet families with diferent supports are considered for comparison:Daubechies of order two, four and six(db2,db4,db6),Symlet of order two,four and six(sym2,sym4,sym6),Biorthogonal of order two and four(bior2,bior4),Coilfet of order one and four(coifl,coif4).The results yielded by the best.customized wavelet function are also shown in Table 2.Outcomes are presented in terms of the Type I and Type II classiifcation error.The former evaluates the number of sleepy states,which are misclassified,while the later evaluates the number of awake states classified as sleepy states. Classification results showed that the best customized wavelet outperformed the classical wavelets for the current application in PD fluctuation analysis.In particular Type I error achieved with the new wavelet is lower than the classical ones.Indeed, the classifier is able to reach low Type I error rate (20%),in other words,of 40 sleepy states samples fed to the classifier,8 were erroneously labelled as awake 

590 Customization of Wavelet Function for Pupil Fluctuation Analysis to Evaluate Levels of Sleepiness states(1ower than reported scores[1 5】).It is important to notice that the bior2 wavelet achieved similar Type 3 4 6 DataPoints 牡罄嚣lll王∞l} 童啦孵0^ 浒 聪 嚣廿0蚕 毒l1 拍霹 警罅l_l0d 舶 || ;|'篓 Dala Points Fig.2 Temporal(upper)and frequency response(1ower)of the wavelet functions created for PD lfuctuation analysis(a,b), and two classical wavelet functions Daubechies wavelet of order 4(doted lines),and Symlet wavelet of order 6(solid lines)(d, e).Shadows represent the mean square error(MSE N=10)in the GA procedure,while the bold line indicates the wavelet with the largest fitness function.Note that frequency axis has been normalized by兀. Table 2 Type I and Type II classification errors with classical and customized wavelet functions {囊C1ass A O.8 豢Class A 1。5 承Class C 露Class C O. 5 Fig.3 Two—dimensional feature spaces constructed with a classical wavelet function(bior2)(right)and with the customized wavelet functions n and p(1eft). 

Customization of Wavelet Function for Pupil Fluctuation Analysis to Evaluate Levels of Sleepiness 591 I rate,however the Type II error is higher than the customized one.Regarding the customized wavelet function that yielded the lower fitness function。the classification error Type I and Type I1 were 2 1.5% and 34.2%.respectively。which are lower than some of the classical wavelets. References 【1】 R.P.Morales,A.O.Gutierrez,G.C.Dominguez,Novel signal dependent filter bank method for identification of multiple basal ganglia nuclei in parkinsonian patients, Journal Neural Engineering 8(201 1)26-36. 【2】 Gandhi,B.K.Panigrahi,S.Anand,A comparative study of wavelet families for eeg signal classification 3.4 Generated Feature Space Fig.3 depicts the two-dimensional feature spaces generated by plotting metric 2,1 against 3,2,that is, metric Eq.(1 6)computed for the wavelet space w6 a against the metric Eq.(1 7)evaluated in wavelet space w5 d.Results for both the db6 classical wavelet and the best customized wavelet are rendered.As shown, the feature space generated with the customized function exhibited higher compactness than that yielded by the classical one. 4.Conclusions In the present study a new wavelet function has been designed speciifcally for the analysis of pupil lfuctuation signal aiming at identifying subjective sleepiness levels.The designed wavelet outperformed classical wavelets in terms of less misclassification of sleepiness states,suggesting that it acquired unique features suitable to classify the pupil fluctuation data in different sleepiness conditions.Although the new wavelet still misclassiifed 20.4%of sleepy data as awake,the error rate should be signiifcantly improved by employing more sophisticated classifier such as support vector machines,instead of a basic linear Bayes classifier currently employed.Moreover,the self-sleepiness level is somewhat variable in different subjects and even within a subject[12],and does not always correlate with physiological indicators of drowsiness[1 6].Thus in the next study,the authors will try to classify gradual myosis and large low frequency fluctuation of PD exemplified in Fig.1, which allows us not only to detect driver’s sleepiness but to predict it before the driver perceives own sleepiness[12]. Neurocomputing 74(20l1)305l-3057. 【3】 R.Khushaba,S.Kodagoda,S.Lal,G.Dissanayake, Driver drowsiness classiifcation using fuzzy wavelet.packet-based feature extraction algorithm,IEEE Transactions on Biomedical Engineering 58(201 1) l21.131. 【4】 M.F.Lucas,A.Gaufriau,S.Pascual,C.Doncarli,D. Farina.Multi-channel surface emg classiifcation using support vector machines and signal-based wavelet optimization,Biomedical Signal Processing and Control 3(2008)169-174. [5】P.Arpaia,C.Romanucci,A.Zanesco,Optimal design of wavelet filters based on lifting scheme by means of cultural algorithms,in:Proceedings of the IEEE Conference on Instrumentation and Measurement Technology,Playa Del Carmen,Mexico,2006. 【6】 R.C.Guido,J.F.W.Slaets,R.Kberle,L.O.B.Almeida,J. C.Pereira,A new technique to construct a wavelet transform matching a speciifed signal with applications to digital,real time,spike and overlap paRern recognition, Digital Signal Process 1 6(2006)24—44. 【7】D.Chen,T.Zhang,Neville-lagrange wavelet family for lossless image compression,Signal Process 88(2008) 2833-2842. 【8】 G.Quellec,M.Lamard,P.Josselin,G.Cazuguel,B. Cochener,C.Roux,Optimal wavelet transform for the detection of microaneurysms in retina photographs, IEEE Transactions on Medical Imaging 27(2008) 1230.1241. [9】 A.Maitrot,M.F.Lucas,C.Doncarli,Design of wavelets adapted to signals and application,in:Proceedings of the IEEE International Conference on Acoustics,Speech and Signal Processing,Philadelphia,US,2005. 【1 0】L.Zhen,H.Zhengj ia,z.Yanyang,W.Yanxue, Customized wavelet denoising using intra-and inter・scale dependency for bearing fault detection,Journal of Sound and Vibration 3 13(2008)342—359. 『l】1 S.Usui,Y.Hirata,Estimation of autonomic nervous activity using the inverse dynamic model of the pupil muscle plant,Annals of Biomedical Engineering 23 (1995)375-387. 【1 2】J.Nishiyama,K.Tanida,M.Kusumi,Y.Hirata,The pupil as a possible premonitor of drowsiness,in:29th 

592 Customization of Wavelet Function for Pupil Fluctuation Analysis to Evaluate Levels of Sleepiness Annua1 International Conference of the IEEE Industrial and Applied Mathematics f 1 992). Engineering in Medicine and Biology Society,Lyon, France,2007. 【l5】B.Shi,K.P.Moloney,Y.Pan,V.K.Leonard,B. Vidakovic,J.A.Jacko,etc.,Wavelet classification of high frequency papillary responses,Journal of Statistical A.Kaltsatou,E.Kouidi,D.Fotiou P.Deligiannis.The [13】 use of pupillometry in the assessment of cardiac autonomic function in elite different type trained athletes. Computation nd aSimultiaon 76 f2006)43 1.445. 【16】M.Nakayama,K.Yamamoto,F.Kobayashi,Estimation of sleepiness using frequency components of pupillary response.in:IEEE Proceedings of the Conference on Biomedical Circuits and Systems.Baltimore.2008. European Journal of Applied Physiology 1 1 1(20 1 1) 2079-2087. I.Daubechies,Ten lectures on wavelets Society for 【14】 


本文发布于:2024-09-22 20:27:26,感谢您对本站的认可!

本文链接:https://www.17tex.com/fanyi/16891.html

版权声明:本站内容均来自互联网,仅供演示用,请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系,我们将在24小时内删除。

标签:
留言与评论(共有 0 条评论)
   
验证码:
Copyright ©2019-2024 Comsenz Inc.Powered by © 易纺专利技术学习网 豫ICP备2022007602号 豫公网安备41160202000603 站长QQ:729038198 关于我们 投诉建议