Analysis of sEMG signal for KOA classification

Journal of Harbin Institute of Technology ( New Series)，Vol. 18, No. 6, 2011Analysis of sEMG signal for KOA classificationLI Yu-rong',2, LIAO Zhi-wei'2, DU Min2李玉榕，廖志伟，杜民(1. College of Electrical Engineering and Automation, Fuzhou University , Fuzhou 350108 , China, liyurong@ fzu. edu. en;2. Fujian Key Laboratory of Medical Instrumentation & Pharmaceutical Technology, Fuzhou 350002，China)Abstract: The sEMG signals are ollected from the vastus lateralis , vastus medialis , biceps femoris, and semi-tendinosus of lower extremnity during level walking among control subjects and knee osteoarhritis (OA) patients , the latter including mild , moderate and severe degree. The 5-fold cross-validation is used to measure theaccuracy of the proposed analysis algorithm on collected sEMG recordings. For comparison, the more classicalfeature vectors of form factor, degree of skewness, kurtosis, and wavelet entropy are also tested. In experiment ,the normalized energy ratio and marginal spectrum ratio achieve larger accuracy than the other features for allthe four muscular groups. Moreover the accuracy of vastus medialis and biceps femoris are larger than that ofvastus lateralis and semitendinosus. These results suggest that the normalized energy ratio and marginal spec-trum ratio via the analysis of knee sEMG signals by HHT can server as characteristic parameters to easily classi-fy ostcoarthritis with noninvasive method. The more important muscular groups for maintaining the knee jointfunction are medialis and biceps femoris; as a result of that they should be exercise especially for rehabilitation.Key words: osteoarthritis ( OA) ; noninvasive diagnosis; surface electromyography ( sEMG) ; Hilbert-HuangTransform ( HHT); neural network classifierCLC number: Q81Document code: AArticle ID: 1005-9113(201 1)06-0113-07Osteoarthritis (OA) is a highly prevalent chronictraction can be recorded with electrodes placed over thehealth condition that causes substantial disability in lateskin. The result for surface electromyography ( sEMG)life. The knee is commonly affected and is the most fre- is the sum of the action potentials generated by the mo-quent site of joint pain in elderly patients. Knee osteo-tor units and filtered by the volume conductor-10. Be-arthritis ( KOA) impacts upon activities of daily livinging a noninvasive method, sEMG signals is achievingwhich consequently lead to a loss of functional inde-the increase of attention in several fields, such as phys-pendence-2. Noninvasive detection and analysis ofiological muscle assessment， rehabilitation, and sportdifferent signals around knee joint can provide a simpleand geriatrie medicine[17-18].and effective method to classify and cure KOA in earlyBy applying machine learming algorithms, such asstage, which is important for improving life quality.kernel, neural network and linear discriminant algo-Various signals around knee joint have been used forrithms to the analysis of surface electromyographic paanalyzing knee pathology, such as dynamic electricalterns, a recent study found the differences betweenimpedance signals--acoustical signals'sl ，terahertzsymptomatic arthritic patients and healthy subjects dur-pulsed imagingacceleration da-ing gait， The study involved subjects suffering fromtal8l ，vibroarthrographic signals', knee X-Ray im-two forms of arthritis , namely，rheumatoid arthritis andage 10) and gait data"" . They have gotten some inter-hip osteoarthritis. In this analysis algorithm, however,esting conclusions， namely these signals around thewithout the feature extracting， raw surface electromyo-knee joint are changed because of KOA.graphy recordings are input to the machine learning al-Meanwhile, KOA is believed to result from degen-gorithms，entailing substantial computations. With theerative changes in the joint cartilage by deleteriously in- development of signal analysis，various techniques havecrease in joint loading, which will cause weak muscle ,been employed for sEMG signals processing, amongand in turn further aggravate joint degeneration, form-which the Fourier and wavelet transform are the twing the vicious cycle' 2，Much research has illuminat-most important and widely used classes. However ,:d the changes of muscle activityKOA subthere are some crucial restrictions for the Fourier spec-tral analysis: the system must be linear, and thee datatremity is a good alternative for discriminating peoplemust be strictly neriodie or_ stationarv; otherwise, thewith KOA. The electrical activity in muscle during con-spectrum W中国煤化工The stationarityYHCNMHGRecceived 2011 -01 -17.Sponsored by the International Science and Technology Cooperation Project of China ( Grant No. 2009DF A32050)..Journal of Harbin Insitute of Technology (New Series), Vol. 18, No. 6, 2011The sEMC signals is sampled with 1 kHz frequen-d0;(I)w,(t) =cy and converted into a digital stream of data. All theddata are exported and analyzed by MyoResearch andHaving obtained the IMF components, the HilbertMatlab software，respectively.transform is applied to each of these IMF components.After performing the Hilbert transform to each IMF2 Principle of Hilbert-Huang Transformcomponent, the original data can be expressed as thereal part in the following form :The sEMG signal is nonstationary as its statisticalproperties change over time. The complexity and non-x(1) = Re E a()lacoulinearity of neuromuscular system lead to the nonlinearHere we have left out the residue rn(t) on pur-property of sEMG signals.pose, for it is either a monotonic function or a con-The combination of empirical mode decompositionstant.( EMD) and Hilbert spectral analysis is designated byEq. (1) can represent the armplitude ( or the ener-NASA as the Hilbert-Huang Transform ( HHT) forgy)，and instantaneous frequency as functions of timeshort. It is based on an adaptive basis , and the frequen-in a three dimensional plot. The HHT time-frequencycy is defined through the Hilbert transform. Conse-spectrum can be obtained as follows:quently, there is no need for the spurious harmonics torepresent nonlinear waveform deformations as in any ofH(w,t) = Re 2 a()el0o(1)由(1)the a priori basis methods, and there is no uncertaintyThen the marginal spectrum can be defined asprinciple limitation on time or frequency resolution fromthe convolution pairs based also on a priori bases. Inh(w) =[ H(w,t)di(2)Ref.[ 20]，there is a summary of the comparison bewhere T is the total length of the signal.tween Fourier， Wavelet and HHT analyses ，indicatingthat HHT is a super tool for time-frequency analysis of3 Data Analysisnonlinear and nonstationary data, and is widely used inthe analysis of earthquake signal, voice signal ，weatherThe sEMG signal analysis involves several stages:data and mechanical engineering detection.sEMG signal detection, feature extraction, and sEMGThe EMD technique can decompose the originalclassification. Correct classification depends on the ex-signal into a set of intrinsic mode functions ( IMFs).tracting distinctive features. According to the nonsta-The obtained IMFs must satisfy two conditions: 1) intionarity and nonlinearity of the sEMG signal, the Hil-the whole data set, the number of extrema and thebert-Huang transform is adopted in the paper to extractnumber of zero crossings must either equal or differ atthe features, which are then feed to a neural networkmost by one; 2) at any point, the mean value of thefor classification. Of course, before the Hilbert-Huangenvelope defined by the local maxima and the envelopetransform, some proper preprocesses ，including normali-defined by the local minima is zero. Thus, the originalzation and noise reduction, are needed. Fig. 2 shows thesignal x(t) is the sum of the IMFs plus the residue:block diagram of proposed system.x(1)= Ec(1) +r,()wherec;(t) (i = 1,2,,n) are the IMFs; r。(t) issEMGsENGFeaturethe residue, which can be either the mean trend or arecording preprocessingextracting| CI assificationconstant.The Hilbert transform is applied to each IMF ，Fig. 2 Block diagram of the sEMG signal classification sys-c,(I), to obtainy;(t)y() =二p[°Rdr3.1 Time and Amplitude Normalizationwhere P indicates the Cauchy prineipal value. WithIf one does not use standardized gait speeds, it isthis definition, c;(t) and y;(t) form a complex conju-difficult to undertake the intra-and inter-individualgate pair, so we can have an analytic signal,z(t),ascomparisons and define the true changes in sEMG af-z;(1) = c(t) +y;(t) = a;(t)i10(0)fected by K0A. Furthermore, age can also have influ-where .ence on gait speed. So in the paper, the time and am-x(1)中国煤化工lift and drop ofa;(t) =[c?(t) +y?(I)] ,0(t) = arctantoleg occurreide cycle respec-The instantaneous frequency of the dynamic signaltively，anMHCN MH Gused the linearmethod, i. e.，for each muscle its maximum activitycan be obtained as●115..Journal of Harbin Institute of Technology (New Series)， Vol. 18, No. 6, 2011was set to one and its minimum activity was set to neg-ratio M of ith IMF componentare defined asl24) : .ative one. Such a normalization procedure was adoptedto give an equal importance to all the muscular groupsE 1 a,(1)12E;and also to preserve information concerning differencesE; =-,M =‘,i = 1,2,..n .in sEMG amplitude between patients and control.ZE3.2 Noise ReductionThe sEMG signals is usually contaminated withwhere n is the number of IMF components, andN is therandom noises ，among which electromagnetic interfer-points of every IMF component. In our study N is theence from power lines is the most serious one. The in-normalized time points, equal to 1000. The normalizeddependent components analysis ( ICA) is an effectiveenergy ratio M; represents the energy of ith IMF compo-blind source separation technology. Firstly ICA is spe-nent in the whole energy, which can reveal the energycially adopted to reduce power lines interference for thedistribution of the sEMG signals.collected sEMG signal dataset.Then for the rest noise, the wavelet denoising is agood alternative. The wavelet coeffcient w of the hard巴(threshold is discontinuous at threshold λ, leading thereconstructed signals to be fluctuant. The soft thresh-old, however, having a constant bias before and after8001000processing, will affect the approximation degree of re-Normalized timeconstructed signals. An improved wavelet denoising is(a) Original signalproposed. It can not only reduce oscillation of hardthreshold, but also avoid constant deviation of softthreshold effectively. The improved wavelet ceofficientw\ can be represented as follows:sign(w) *、(@3 +(1 w1-A),1 w1≥λ200w、={(b) Signal afer noise reduction by ICAsign(w)0-,1 ol < λ(3)In Eq. (3), in order to ensure the integrity of sig-nal in denosing process, the wavelet coefficient W、isexpressed as the mean square error ( MSE) of w and w40060100- λ，which makes the processed wavelet coefficientcloser to original ’s. For the case of wavelet coefficient(c) Signal after noise reduction by improved waveletless than threshold, the wavelet coefficient is transitedFig.3 Time and amplitude normalized sEMG signal of va-to zero smoothly other than directly set to zero , makingstus medialis of subject 5 when level walkinginput-output curve continuous.A time and amplitude normalized sEMG signals ofTab.2 Performance of noise reductionvastus medialis of subject 5 when level walking, signalwaveletWavelet package Improved waveletafter noise reduction by ICA and improved wavelet are14. 25626.554shown in Fig. 3.RMSE .0.972440.0562480. 031549In order to quantify the performance of noise re-duction, two quality measures，Signal-to-Noise ( SNR )The HHT time-frequency spectrum and marginaland Root Mean Square Error ( RMSE)223], are used.The higher SNR and lower RMSE suggest better per-spectrum of ith IMF component are obtained accordingformance. The SNR and RMSE after noise reduction byto Eqs. (1) and (2) as follows:wavelet, wavelet package and improved wavelet of Fig.H(w,) = Re[ a(t)aekwa],h;(w) ={ H(w,t)dt3(b) are shown in Tab. 2. The noise reduction by im-The marginal spectrum offers a measure of totalproved wavelet achieves the best performance.3.3 Hilbert-Huang Transform and Spectrum A-amplitude ( or energy ) contribution from each frequen-cy value. It represents the cumulated amplitude overnalysisthe entire data中国煤化工se. The con-After the empirical mode decomposition, a set oftribution of theancy is meas-intrinsic mode functions will be gotten for every sEMGCNMHGsignal. The average energy E; and normalized energyured by the maIn our study, the frequency band between 0 and●116●.Journal of Harbin Instiute of Technology (New Series), Vol. 18, No.6, 2011500 Hz of marginal spectrum is divided into five equalsubsamples used exactly once as the validation data.segments，so length of every segment is 100 Hz. ByThek results from the folds are then averaged to pro-the division of frequency band, the marginal spectrumduce a single estimation. The advantage of this methodover individual band will be highlighted.over repeated random sub-sampling is that all observa-The normalized marginal spectrum ratio SH; acrosstions are used for both training and validation， andthe jth 100 Hz segment of ith IMF component is calcu-each observation is used for validation exactly once. 5-lated as:fold cross-validation is used in the paper. Then theroutine is repeated 10 times, and the 10 results fromSH;= S h,(k)/E S h;(k)each routine are then averaged to give the final value.声k=(j-1)x100 .4.2 Performance MeasureThe performance of classifier is determined usingwherei = 1,..n,j = 1,2,3,4,5.accuracy，calculated for the four classes. The accuracyof a class is the ration between the number of test sam-4 Results and Discussionples that the classifier identifies correctly and the num-ber of all test samples in the class.4.1 Experiment DesignWith the sEMC signals recorded from vastus late-The BP and RBF neural network classifiers areralis, vastus medialis, biceps femoris, and semitendi-used, output of which are four classes: control, mild,nosus, the average accuracy for different feature vectormoderate and severe. After the HHT, the normalizedachieved by BP and RBF classification respectively ,energy ratio M; and marginal spectrum ratio SHg(i = 1,are shown from Tab.3 to Tab. 6. All results are deno-2,.n, where n is the number of IMFs, andj = 1,2,ted in the form of mean土standard deviation.3,4,5) are chosen as two kinds of feature vector,We can see that after HHT, two features , normal-feeding to classifer respectively. Furthermore anotherized energy ratio and marginal spectrum ratio, performfour kinds of feature vector， form factor， degree 0better than the other features , and the normalized mar-skewness, kurtosis, and wavelet entropy, are chosenginal spectrum ratio performs the best, indicating thatHHT is a super tool for time-frequency analysis of non-for performance comparison.A total of fifty objects shown in Tab. 1 are tested.linear and nonstationary sEMG data.The k-fold cross validation is adopted in the classfica-It can also be found that the classification accura-tion routine. The original sample is randomly parti-cy of vastus medialis and biceps femoris are larger thantioned into k subsamples. Of the k subsamples, a singlethat of vastus lateralis and semitendinosus. It suggestssubsample is retained as the validation data for testingthat vastus medialis and biceps femoris play a more im-the model, and the remaining k - 1 subsamples areportant role in maintaining the knee joint function, andused as training data. The cross-validation process isthe damage of these two muscular groups will be morethen repeated k times ( the folds)， with each of the hlikely lead to KOA.Tab.3 Average accuracy and standard deviations with vastus lateralis ( mean土s. d.)Feature vectorClassifiercontrolMildModerateSevereBP64. 83 +9.2172. 62 +6. 3372. 21 +4. 2955. 00+11.83Form factorRBF67. 027.78.73.54+4.9373.81士4. 2860.00+13. 70Degree of69. 20+4.9276. 16+5. 0176.88 +5.52.65. 00+13.69skewness71.40+6.03.74. 36 +6.0978. 50 +6.6468. 20+8.2358. 30+3. 8364. 70+3.9360.00+13. 69Kurtosis70. 40 +4. 9859.31+4.63.66. 10 +3.9865. 00+13. 69Wavelet71. 40 +6. 3560. 30 +5.0767. 50 +5.7955. 00+13.69entropy73. 60+6.3561.11 +4.2869.00 +4.953P75. 80 +4.9262. 10+4. 2574. 80 +6.2275. 00+17. 68ratio80.20+6.03.63. 00+2.2中国煤化工00x11.18 .Normalized marginal82. 40+6.0364. 60+3.7:YHC N M H G00+13.69spectrum ratio84. 60 +6.0368. 20 +5.3182. 40+6.6985. 00+22.36●117●.Journal of Harbin Institute of Technology (New Series)， Vol. 18, No. 6, 2011Tab.4 Average accuracy and standard deviations with vastus medialis ( mean士s. d.)%Feature vectorClassifierControlMildModerateSevereForm factorBP75. 80 +4. 9276. 20士4.3869. 40+6.3855. 00+11.83BF80. 20 +6.0372. 00 +9.9571. 40 +7.6760.00 +13.7069.20+4. 9269. 60+8.5973. 40 +6.076s. 00+13.69Degree of skewnessRBF71. 40+6. 3071. 20+9. 1774.80+5.59.70.00+13. 6971.40 +6.3072.80 +9.3975. 00+5.6670. 00 +22.36Kurtosis73. 60 +6.0371. 28 +9.2576. 60 +4.5675. 20 +9.2075. 20 +4.2580. 40+6. 5470.00 +20.92Wavelet entropy77. 40 +6.3076. 54 +6. 3682. 80 +6.2275. 00+17.68Normalized energy ratio3P85.00+11. 0083. 00+6.6085. 00+4.1280. oo +20.9289. 20+12.0584. 80 +5.7286. 60 +3.7295. 00+11.18Normalized marginal87. 60 +6.0388. 40+2.1991. 40+5.8495. 00+11.64spectrum ratio90. 60 +9.8490.20 +3.1992.00+5. 6795. 00+13.69Tab.5_ Average accuracy and standard deviations with biceps femoris ( mean :ts.d. )71. 40 +6.3576.80 +5.3171. 20土10.0155. 00 +11.8373. 60+6. 3577. 60 +5.6472. 40士10.9055. 00+11.83.69.20+12. 0575. 80+7. 1275.20 +8.8465. 00土I3.6971. 40 +9.8475. 00+6.5973. 60+9. 9965. 00+13.6967. 00+7.7874. 20 +8.6774.00 +9.2760.00+13. 7069. 20+12.5076. 00+7.9774. 60+9. 9275. 00 +17.6880. 02 +9.3181.40 +4.5176. 20+7.5370. 00 +20.9278. 00 +11.0082. 20+3. 8377. 40+6.51.75. 00 +17.68.75. 40 +9.8484. 23 +2.7482. 40 46.7380.00 +20.9281. 60 +9.8483. 20 +2.7485. 00 +22.368P85. 20+4. 2085. 80+3. 1984. 20 +5.3190.00+13. 6988.40+6. 0387. 60 +2.1985.80 +6.3495. 00+12.36Tab.6 Average accuracy and standard deviations with semitendinosus ( mean土S. d. )Forn factor68. 20 +8.2366.62 +5.8174. 60 +9.9270.40土4.98.67. 40+5.6475. 80 +9.4264. 83+9.2169. 20+7.1271. 80+11. 6169. 20 +4.9270. 00 +6.5970. 20 +9.9465.00+13. 6964. 83 +9. 2165. 60+2. 1973.40+4.0420. 00 +20.9267.02 +7.7868. 20+5. 3175. 00 +6.2825. 00+17.6873. 60+12. 5470. 80+5.8973. 40 +8.7440. 00+22.36Wavelet entopy75. 80+12. 0573. 40+4.9374. 80+9. 9445. 00 +20. 92Nornalized energy ratio75. 80+9.2176. 40 +5.3250. 00+17.68.RBE78. 00 +7.7877. 60+5. 6479. 40 +4.:4. 2255. 00 +20. 9280.02 +9. 2379. 40+2. 1985. 60士6.6985. 00 +13.69.82. 40+6.0380.20+1.7984.00+8.0090.00+20. 92cause of the nonlinearity and nonstationarity of sEMG5 Conclusionssignals, the Hilbert-Huang transform is used to processthe sEMG signals and normalized energy ratio and mar-Due to the reduced knee joint stability, muscleginal spectrum ratio are extracted as the features. Theactivation pattern at the knee is altered in different se-BP and RBF neural network classifiers are then adopt-verity degree KOA compared with healthy subjects.ed to classify the subjects respectively. ExperimentalSurface electromyography ( sEMG) signals, which isresults by the 5-fold cross-validation indicate that the :the spatio-temporal bioelectric activity of muscles re-proposed meth中国煤化工altemnative to .corded by surface electrode, is important to quantita-classify the KDHc N M H Ggsical rehabil-tively evaluate skeletal muscle activity. So it is chosentation training 10r non, vastus mealais and bicepsas the signal to classify the KOA in this paper. Be-femoris should be exercised specially.●118..Journal of Harbin Institute of Technology ( New Series)，Vol. 18, No. 6, 2011In order to further understand the disorder, aceedings of the 5th International Conference on Machinemore detailed examination of the noise characteristic,Leaming and Applications. Orlando: IEEE, 2006. 225 -230.its corresponding noise reduction algorithm, and more[12]Yux, Yi Y, Hu Y. Surface electromyography of patientswith knee osteoarthritis. Chinese Jourmal of Physical Medi-effective classification algorithm are currently under-cine and Rehabilitation, 2006, 28(6): 402 - 405.way. How to mask the movement artifacts is another13 ]AstephenJ L, Deluzio K J, Caldwell G E, et al. Gait andproblem to be necessarily handled.neuromuscular patterm changes are associated with differ-References:ences in knee osteoarthritis severity levels. Jourmal of Bio-mechanics, 2008, 41(4): 868 - 876.[1] ChildsJ D, Sparto P J, Fitzgerald G K, et al. Alterations[14]Likavainio T, Bragge T, Hakkarainen M, et al. Cait andin lower extremity movement and muscle activation patternsmuscle activation changes in men with knee osteoarthritis.in individuals with knee osteoarhitis. Clinical Biomechan-The Knee, 2010, 17 (1): 69 -76.ics, 2004, 19(1): 44 -49.[ 15 ]Hubley -Kozey C L, Deluzio K J, Landry sC, et al. Neu-[2] von Tscharner V, Valderrabano V. Classification of multiromuscular alterations during walking in persons with mod-muscle activation patterms of osteoarthritis patients duringerate knee osteoarthritis. Journal of Electromyography andlevel walking. Jourmal of Electromyography and Kinesiolo-Kinesiology, 2006, 16 (4): 365 - 378.gy, 2010, 20(4): 676 -683.[ 16]Farina D, Holobar A, Merltti R, et al. Decoding the neu-3] Gajre s s, Singh S U. Novel method of using dynamic elec-ral drive to muscles from the surface electromyogram. Clini-trical impedance signals for noninvasive diagnosis of kneecal Neurophysiology, 2010, 121 (10): 1616 - 1623.osteoarthritis. Proceedings of the 28th IEEE EMBS Annual[17] Jiang Chingfen, Kuo Shoulong. A comparative study ofInternational Conference. New York: IEEE, 2006. 2207 -wavelet denoising of surface electromyographic signals. Pro-ceedings of the 29th annual International Conference of the4] Neves E B, Pino A V. Objective assment of knee os-IEEE EMBS. Lyon: IEEE, 2007. 1868 - 1871.teroarthritis in parachuters by bioimpedance spectroscopy.[18] Kaplanis P A, Pattichis C s, Hadjileontiadis L J, et al.Proceedings of the 30th IEEE Annual Intermational Confer-Surface EMG analysis on normal subjects based on isomet-ence of Engineering in Medicine and Biology Society. Can-ric voluntary contraction. Journal of Electromyography andada: IEEE, 2008. 5620 - 5623.Kinesiology, 2009, 19 (1): 157 -171.5] KimKS, LeeSO, SeoJ H, et al. Classification of joint[ 19]Nair S s, French R M, Larche D, et al. The application ofpathology using an acoustical analysis of knee joint sound.machine learning algorithms to the analysis of electromyo-Proceedings of the IEEE Biomedical Circuits and Systemsgraphic patterns from arthritic patients. IEEE TransactionsConference. London: IEEE, 2006. 65 -68.on Neural Systems and Rehabilitation Engineering, 2010,6] Kan K W C, Pickwell-MacPherson E, Cheung W H. Tera-18(2):174- 184.hertz pulsed imaging of osteoarthritis. Proceedings of the[ 20 ]Huang N E, Wu Zhaohua. An adaptive data analysis meth-5th International Summer School and Symposium on Medi-od for nonlinear and nonstationary time series: the empiri-cal and Biosensors. Hong Kong: IEEE, 2008. 209 -212.cal mode decomposition and hilbert spectral analysis. Pro-[7] Mezghani N, Boivin K, Turcot K, et al. Automatic casif-ceedings of 2007 Intermational Conference on Wavelet Anal-cation of asymptomatic and osteoarthritis knee gait patternsysis and Applications. Macau: IEEE, 2007. 363 - 376.using kinematic data features and the nearest neighbor clas-[21 ]HuangN E, Shen z, LongS R, et al. The empirical modesifier. IEEE Transactions on Biomedical Engineering,decomposition and the Hilbert spectrum for nonlinear and2008, 55(3): 1230 - 1232.non-stationary time series analysis. Proceedings of the Roy-[8] Taylor P E, Almeida G J M, Kanade T, et al. Classifyingal Society of London, 1998, 454( 1971) :903 -995.human motion quality for knee ostcoarthritis using acceler-22] Panoulas K I, Hadjileontiadis L J, Panas S M. Hilbert-omelers. Proceedings of Annual International Conference onHuang spectrum as a new field for the identification of EEGEngineering in Medicine and Biology Society ( EMBC).event related De-/ Synchronization for BCI applications.Buenos Aires: IEE, 2010. 339 - 343.Proceedings of the 30th annual Intemnational Conference of[9] Rangayyan R M, Wu Yunfeng. Analysis of knee-joint vi-the IEEE EMBS. Vancouver: IEEE, 2008. 3832 - 3835.broarthrographic signals using statistical measures. Pro-23 ] Phinyomark A, Phukpattaranont P. EMG denoising estima-ceedings of International Symposium on Computer- Basedtion based on adaptive wavelet thresholding for mulifunc-Medical Systems. Maribor: IEEE, 2007. 377 - 382.tion myoelectric control. Proceedings of Conference on In-[10]Shamir L, LingS M, Scott W w, et al. Knee X-ray imagenovative Technologies in Intelligent Systems and Industrialanalysis method for automated detection of osteoarthritis.Applications. Malaysia: IEEE, 2009. 171 - 176.IEEE Transactions on Biomedical Engineering, 2009， 56[ 24]Tavakkoli F, Teshnehlab M. A ball bearing fault diagnosis(2): 407 -415.method based on wavelet and EMD energy entropy mean. Pro-[11]Koktas N S, Yalabik N,avuzer G. Ensemble elassifiers forceedings of International Conference on Intelligent and Ad-medical diagnosis of knee osteoarthritis using gait data. Prvanced Systems. Kuala Lumpur: IEEE, 2007. 1210 - 1212.中国煤化工MHCNMHG●119●