DEVELOPMENT OF AN EFFICIENT EPILEPSY CLASSIFICATION SYSTEM FROM EEG SIGNALS FOR TELEMEDICINE APPLICATION

Transcription

1 Interntonl Journl of Cvl Engneerng nd Technology (IJCIET) Volume 8, Issue 1, December 017, pp. 38 5, Artcle ID: IJCIET_08_1_005 Avlble onlne t ISSN Prnt: nd ISSN Onlne: IAEME Publcton Scopus Indexed DEVELOPMENT OF AN EFFICIENT EPILEPSY CLASSIFICATION SYSTEM FROM EEG SIGNALS FOR TELEMEDICINE APPLICATION Hrumr Rjguru Correspondng Author, Deprtment of ECE, Bnnr Ammn Insttute of Technology, Sthymnglm, Ind Sunl Kumr Prbhr Deprtment of ECE, Bnnr Ammn Insttute of Technology, Sthymnglm, Ind ABSTRACT: The most promnent tool used n the clncl dgnoss nd montorng of the brn s the Electroencephlogrph (EEG). Wth the help of severl electrodes, EEG sgnls re recorded nd for further processng t hs to be trnsmtted through verstle communcton chnnel. In ths pper, ntlly the dmensons of the rw EEG sgnl re reduced wth the help of Independent Component Anlyss (ICA) whch s employed here s dmensonlty reducton technque. It s then trnsmtted through the Dfferentl Spce Tme Bloc Coded (DSTBC) Multple Input Multple Output (MIMO) Orthogonl Frequency Dvson Multplexng (OFDM) system. For the DSTBC MIMO-OFDM System, Pe to Averge Power Rto (PAPR) reducton lgorthm clled Projecton onto Convex Sets Actve Constellton Extenson (POCS-ACE) Algorthm s employed to reduce the PAPR. At the Recever sde, low Bt Error Rte (BER) s med. The recever sde s ncorported wth Gussn Mxture Model (GMM) Clssfer to clssfy from the eplepsy from EEG sgnls. The Clssfcton Performnce Mesures re consdered n terms of Specfcty, Senstvty, Tme Dely, Qulty Vlues, Performnce Index nd Accurcy. Keywords: EEG, PAPR, ICA, GMM, POCS-ACE, MIMO Cte ths Artcle: Hrumr Rjguru nd Sunl Kumr Prbhr, Development of n Effcent Eplepsy Clssfcton System from EEG sgnls for Telemedcne Applcton, Interntonl Journl of Cvl Engneerng nd Technology, 8(1), 017, pp edtor@eme.com

2 Development of n Effcent Eplepsy Clssfcton System from EEG sgnls for Telemedcne Applcton 1. INTRODUCTION EEG records the electro-potentls whch re evoed from the cortcl sectons of the brn throughout the sclp [1]. Becuse of EEG dscovery, numerous reserch ctvtes re concentrtng on extrctng the most sgnfcnt nformton bout the condtons of the brn. In ntensve cre montorng of the ptents, EEG s hghly useful to mesure the neuronl ctvtes of the brn []. EEG s lso used n the tretment of epleptc sezures nd for the dgnoss of brn deth. Recordng vrety of electrcl potentls from humn brn s qute dffcult becuse EEG sgnls re non-lner nd non-sttonry n nture [3]. EEG sgnls re hghly prone to be dsturbed by the movement of eyes nd other muscle noses nd hence t s qute dffcult to dentfy nd dfferentte the vrous mentl clsses nd therefore clssfcton lgorthms, especlly mchne lernng lgorthms hold pvotl poston [4]. The mchne lernng lgorthms fnd t dffcult to del wth the enormous mount of rw dt whch s produced durng the recordng trls [5]. Thus the huge tme seres cn be represented usng dmnshed form of representton, tht s, dmensonlly reduced seres wth the help of sutble dmensonlty reducton technques. Becuse of the huge dt sze of the EEG, due to the longer recordng tmes nd lrger number of electrodes requred for processng t, dmensonlty reducton or dt compresson s requred for effcent trnsmsson nd so ICA s effectvely used here Relted Wors Compresson technques m to obtn the hghest dt volume reducton whle the vtl nformton s preserved for reconstructon purposes [6]. The dt compresson cn be ether lossy or lossless dependng on the fdelty of the sgnl wveform. In lossless condtons, the fdelty of the sgnl wveform s lost completely. In lossless compresson methods, due to the rndom nture of the EEG sgnl, hgher compresson rtes cn never be cheved. If the hgher compresson rte needs to be obtned, then the mthemtcl trnsforms re utlzed thoroughly. The perfect reconstructon of the sgnl cn never be cheved n lossy lgorthms but t lwys utlzes less computtonl sources nd ms to cheve hgh compresson rto. Severl technques le Wve Pcet Trnsform, Dscrete Cosne Trnsform, Dscrete Wvelet Trnsform nd SPHIT re dscussed suffcently n lterture [7]. Here we hve reduced the sze of the EEG dt usng ICA. The Fgure 1 shows the bloc dgrm representton of the entre wor. Fgure 1 Bloc Dgrm of the Entre Wor 39 edtor@eme.com

3 Hrumr Rjguru nd Sunl Kumr Prbhr 1. Mterls nd Methods Ths secton shows the cquston of the EEG dt n detl followed by the need for dmensonlty reducton. It lso emphszes the mportnce of ICA s dmensonlty reducton technque Dt Acquston Accordng to the Interntonl 10-0 system, the EEG s recorded by plcng electrodes on the sclp of the ptent. For both referentl montges, sxteen chnnels of EEG re recorded smultneously. All the electrodes re referenced nd plced to common potentl pont le er nd n the cse of bpolr montges, ech electrode s referenced to n djcent electrode. When the ptent s we nd restng, the recordngs re mde. The restng stge ncludes perods of hyperventlton, photonc stmulton nd ncludes perod of eyes open nd closed. Usng the EEG mchne Semens Mnogrph Unversl the mplfcton s esly provded. The sclp s clened well nd lghtly brded nd then between the electrode nd the sn, electrode pste s ppled nd then the electrodes re plced on t. The contct mpednce s less thn 10 Ω wth respect to the pplcton of electrode pste on the sn. The obtned EEG s broen down nto epochs or sectons for the purpose of dmensonlty reducton. An epoch of seconds s prmrly used becuse t s pretty long enough to cpture the most mportnt sttstcl fetures nd chrcterstcs of the EEG nd t s short enough to cpture the sezure evoluton. The EEG beng dgtzed t totl smplng rte of bout 00 Hz for n epoch of seconds ech whch contns 400 smples totlly, s very convenent length for computton purposes. For nlyzng the EEG dt, the softwre mplemented ws Mtlb 7.0. The EEG dt used n ths study were obtned from 0 dfferent epleptc ptents who were n the evluton nd tretment process n the Deprtment of Neurology, Sr Rmrshn Hosptl, Combtore, Ind. Through 10-0 nterntonl electrode plcng method, dgtl record of 16 chnnel EEG dt n bpolr method s cqured from clncl EEG montorng system n Europen Dt formt. Wth the help of neurologst, EEG records wth most dstnct fetures hd been selected. The totl number of rtfcts present n our dt s four types whch nclude Electromyogrm (EMG) rtfcts, eye blns, chewng nd moton rtfcts. The mn objectve of the ncluson of rtfcts s to hve spe nd non spe ctegores of wveforms. A sutble prt or segment of EEG dt hs to be selected n order to trn nd test the sgnl component extrctors nd clssfer. All the sgnls of EEG re exmned by good qulfed EEG technologst. The EEG records re over contnuous durton of bout 30 seconds nd these records re dvded nto n epoch whch hs two second durton. To detect the sgnfcnt chnges n ctvty, two second epoch s long enough. The two second epoch s lso long enough to vod unnecessry repetton nd redundncy n the sgnl. The EEG sgnl hs mxmum frequency of 50 Hz nd snce smplng frequency hs to be greter thn twce the mxmum frequency nd so ech epoch s smpled t frequency of 00 Hz. Ech nd every smple corresponds to the mpltude vlues of the sgnl whch re nstntneous n nture nd totlly 400 vlues for n epoch s obtned. Therefore ech chnnel hs 400 smples of EEG sgnls per epoch nd four such dt epochs comprses bn. For ptent, there re totlly 16 bns vlble. Therefore the dt volume for ech nd every ptent s round 5,600 smples. Hence for processng ths lrge mount of dt, dmensonlty reducton technques re ncorported. The dfferent prmeters used for the quntfcton of the EEG sgnls re computed usng mpltude vlues wth the help of sutble progrmmng codes. The representton of EEG smples n reduced smple spce s done ether by fetures or dmensonlly reduced components. So we hve chosen the dmensonlty reducton technques whch s dscussed n the followng sectons of the pper edtor@eme.com

4 Development of n Effcent Eplepsy Clssfcton System from EEG sgnls for Telemedcne Applcton 1.. Independent Component Anlyss (ICA) Assumng tht there re totlly 'n' lner mxtures s x 1... xn, where 'n' represents the ndependent components, t cn be wrtten mthemtclly s follows: x =... + j j1 s1 + j s + jnsn for ll j. The vector-mtrx notton s utlzed completely nd the bove equton cn be wrtten s follows [8] x = As Where A denotes the mtrx wth prtculr elements j, x s the rndom row vector of T x 1... x n or sometmes x s used whch denotes the trnspose of the row vector, s s lso the rndom row vector of s 1... sn. Emphszng the mportnce of columns of mtrx A, the model cn be wrtten s follows n x = = 1 s, where denotes the columns of mtrx A. It s consdered s genertve model where n observed dt s descrbed clerly. If the mtrx A s estmted, then the computton of ts nverse, sy P, s obtned esly nd then the ndependent component s obtned s follows S = Px The dmensonlly reduced vlues re then trnsmtted through the DSTBC MIMO OFDM System.. MIMO SYSTEM MIMO system s type where t employs multple ntenns for both trnsmtter nd recepton. Wthout the reducton n the spectrl effcency, ntenn dversty cn be cheved esly usng multple trnsmttng nd recevng ntenns [9]. MIMO beng one of the forms of smrt ntenn technology hs ttrcted lot of ttenton n the feld of wreless communcton. Wthout the ddtonl need of bndwdth nd trnsmt power MIMO offers gret ncrese n throughput of the dt..1 MIMO Spce-Tme Codes MIMO Spce-tme codes re two-dmensonl spce nd tme processng methods [10]. To cheve hgher codng gn, spce-tme codng (STC) s used nd t s generlly clssfed nto spce-tme codng (STC) nd Sptl Multplexng (SM). The Clsscl STC ncludes STTC nd STBC [11]. By performng very smple mxmum lelhood decodng lgorthm, STBC cheves full dversty gn. Wth the bsc codng schemes such s STTC nd STBC, t s ssumed tht the Perfect CSI vlble t the recever sde nd so the type of detecton employed s coherent one. However, f the moblty envronment s hgh, then t s dffcult nd qute expensve to ccurtely estmte the chnnel becuse of the rpdly chngng fdng condtons. So t s very mportnt to mplement spce tme codng technques tht do not consder the employment of chnnel estmtors n both trnsmtter nd recever sde nd such type s clled s dfferentl spce-tme codng. When MIMO s combned wth OFDM t becomes MIMO-OFDM systems [1]. Thus n comprson to the sngle crrer technques, OFDM s beng used wdely becuse t dvdes hgh dt rte strem nto number of low dt rte sub strems. These low dt rte sub strems re smultneously trnsmtted usng subcrrers whch re orthogonl n nture. A hgh spectrl effcency s provded becuse of the orthogonlty of subcrrers n tme domn edtor@eme.com

5 Hrumr Rjguru nd Sunl Kumr Prbhr. PAPR n MIMO-OFDM System wth Relted Wors In MIMO-OFDM System, the output comprses of superposton of multple-sub crrers. In such cse, certn nstntneous power outputs tend to ncrese to greter extent nd become hgher thn the overll men power of the system under the prerequste condton tht the phses of these crrers re sme. Ths condton s defned s hgh PAPR nd t s one of the serous problems ffectng MIMO-OFDM. Certn lgorthms le Clppng nd Flterng, Tone Reservton, Selectve Mppng nd Prtl Trnsmt Scheme were ncorported to combt the hgh PAPR n lterture [13]. In ths pper, POCS-ACE s used to combt the hgh PAPR for the DSTBC MIMO-OFDM System...1 Probblty Dstrbuton Functon of PAPR In ech nd every N T prllel OFDM trnsmtters, prtculr bloc of D dstnct complexvlued crrers sy, Aµ, v, µ = 1,..., N T, v = 0,... D 1, s trnsformed nto ts respectve tmedomn usng the Inverse Dscrete Fourer Trnsform, tht s, 1 1 D jπv / D µ, = Aµ, v. e, = 1,..., NT, = 0,..., D 1. v= 0 µ D On the combnton of the tme-domn smples nto specfc vector ], the respectve correspondence s wrtten s IDFT { } wreless pplctons, ll frequency-domn smples smlr constellton ponts wth vrnce σ. µ A µ µ = [ µ, 0,..., µ, D 1 =. Generlly n lmost ll the A µ, v re expected to be obtned from Snce the crrers re sttstclly ndependent, the tme-domn smples µ, re complex Gussn dstrbuted n n pproxmted sense. Ths leds to very hgh pe-to-verge power rto s PAPR mx = E µ { µ, }, mx µ = As verstle performnce mesure, the probblty tht the PAPR of n OFDM frme exceeds gven threshold,.e., tht the squred mgntude of t lest one smple over the N σ µ,, PAR0σ ntenns nd D tme steps s lrger thn tolerted: µ > s crred out n lterture. Wth ths consderton, the Complementry Cumultve Dstrbutve Functon (CCDF) of the PAPR P r PAPR > PAPR }, clppng probbltes cn be determned esly. { 0 In MIMO-OFDM, snce N T D nsted of D tme-domn smples re present nd the CCDF of the PAPR s represented mthemtclly s follows [14] PAPR 0 N P r { PAPR MIMO > PAPR 0} = 1 (1 e ) A lot of PAPR reducton technques re well dscussed n lterture. Gnesn et l. [15] proposed vrety of PAPR Reducton Technques for SISO OFDM Systems. The ntl extensons of reducton of PAPR to MIMO-OFDM Systems were gven by Bssem et l [16]. In STBC MIMO-OFDM, Mzhou Tn et l. [17] proposed PAPR reducton scheme usng Cross-Antenn Rotton nd Inverson. Bg et l. [18] proposed ZMCT Precodng Bsed STBC MIMO-OFDM System wth reducton n PAPR. Alhb et l. [19] proposed combned SLM nd closed loop QO-STBC for PAPR mtgton n MIMO-OFDM trnsmsson. Jngbo et l. [0] proposed PAPR reducton scheme n Blnd MIMO OFDM System bsed on Independent Component Anlyss. Hyen-Seng Joo et l. [1] proposed T D T 4 edtor@eme.com

6 Development of n Effcent Eplepsy Clssfcton System from EEG sgnls for Telemedcne Applcton Blnd SLM PAPR Reducton scheme usng cyclc prefx n STBC MIMO-OFDM System. Y- Sheng Su et l. [] proposed computtonlly effcent PAPR scheme for STBC OFDM System. P.S Kumr et l [3] showed the performnce nlyss of PAPR reducton n STBC MIMO OFDM System. 3. PAPR REDUCTION IN DSTBC MIMO-OFDM SYSTEM The followng flow chrt s mplemented for the PAPR Reducton Anlyss usng POCS-ACE for Dfferentl Spce Tme Bloc Coded MIMO-OFDM System. The Fgure shows the procedure of PAPR reducton n DSTBC MIMO-OFDM System. Fgure PAPR Reducton of DSTBC MIMO OFDM System Usng POCS-ACE 3.1. PAPR Reducton Technque for Dfferentl STBC MIMO OFDM System Usng PROJECTION ONTO CONVEX SETS-ACTIVE CONSTELLATION EXTENSION (POCS-ACE) Algorthm Projecton onto Convex Sets (POCS) [4] s method to fnd pont n the ntersecton of the closed convex sets. The cse when the sets re ffne spces s specl, snce the tertes not only converge to pont n the ntersecton, but n fct to the orthogonl projecton onto the ntersecton of the ntl terte [5]. Clsscl wor on the cse of two closed convex sets shows tht proof n most cses. Rte of the convergence of the tertes s lner lwys. Anlyss of POCS shows tht the lgorthm conveys nd whether t converges to the projecton of the orgnl pont [6]. These questons re lrgely nown for smple cses but topc of ctve reserch for the extensons. The Fgure 3 shows the ncorporton of POCS- ACE lgorthm for PAPR Reducton. Fgure 3 Bloc Dgrm for ncorportng the PAPR reducton technque usng POCS- Actve Constellton Extenson n DSTBC MIMO-OFDM system 43 edtor@eme.com

7 Hrumr Rjguru nd Sunl Kumr Prbhr The followng lgorthm steps re followed completely: Step 1: The progrm s strted nd then the nput bts re generted rndomly Step : The serl dt s converted nto prllel dt Step 3: The nput sgnls re modulted usng pproprte modulton schemes le DQPSK, 16 QAM, nd 64 QAM. In DQPSK, the symbol nformton s encoded s the phse chnge from one symbol perod to the next rther thn s n bsolute phse. In ths cse, the recever hs to detect phse chnges nd not the bsolute vlue of the phse, whch vods the need for synchronzed locl crrer. Step 4: Menwhle, the prllel dt obtned s lso drectly computed usng N pont Inverse Fst Fourer Trnsform (IFFT) Step 5: Then clppng process [7] s done nd t s converted nto frequency domn. The out of bnd crrers re removed nd then the Actve Constellton Extenson (ACE) lgorthm s ppled Step 6: The ACE lgorthm cn be ntlzed by selectng the prmeters nmely the trget clppng level nd the number of tertons, denoted by. Step 7: The ntl trget clppng level s ssumed s A nd then the clppng level s computed. Step 8: The clppng sgnl s trnsferred to the nt-pe sgnl subjected to the ACE constrnt. Step 9: Increse the totl number of tertons counter to ten. Step 10: The threshold vlue s clculted nd then t s checed tht whether PAPR>threshold vlue Step 11: As fnl step, the Complementry Cumultve Dstrbuton Functon (CCDF) plot versus probblty of PAPR s computed nd the plot s drwn Step 1: The bt error rte s lso computed for the DSTBC MIMO-OFDM system. Step 13: Stop the progrm 3.. PROJECTION ONTO CONVEX SETS ACTIVE CONSTELLATION EXTENSION ALGORITHM (POCS-ACE): Let the totl number of subcrrers be ssumed s K. Let nd H be the chnnel mtrx. W be the trnsmt weght mtrx Step 1: Assgn tme-user-domn representton constellton ponts, x, for ll the B users, where s the terton number. Construct the ntl P 1 tme-ntenn-domn 0 0 dt by Z = W K, nd set =0 ntlly. K K Step : Reconstruct the tme ntenn-domn dt by tng the Trnspose nd then pply zero-pdded IFFT on ech element p = nt( IFFT( Z T = [ 1 p ] Where nt represents the nterpolton functon Step 3: Clp the mgntudes of smples on MAX p )) for p {1,,... p}; tht exceed predetermned lmt, G Therefore `becomes â edtor@eme.com

8 Development of n Effcent Eplepsy Clssfcton System from EEG sgnls for Telemedcne Applcton Step 4: Project the chnges n the frequency ntenn domn dt onto the null spce of the chnnel mtrx H nd then updte the process s shown n the followng equton Where + 1 Z = ( I W W )( Zˆ Z ) + Z Pnv Pnv W s pseudo-nverse of W nd Ẑ s the frequency-ntenn-domn dt fter clppng Step 5: Return to step nd repet untl mnmum PAPR s cheved or the number of tertons s exhusted. Step 6: The smultons re crred out s per the system requrement by sutbly djustng the terton vlues PAPR for DSTBC MIMO-OFDM Systems Usng ACE Algorthm Intlly the PAPR reducton for the Dfferentl Spce Tme Bloc Coded-MIMO OFDM Systems re smulted ccordng to the smulton prmeters lsted n Tble 1 Tble 1 Smulton Prmeters for PAPR Reducton n DSTBC MIMO-OFDM System Modulton used DQPSK/16-QAM/ 64-QAM MIMO System nlysed x MIMO-OFDM Number of subcrrers 18 Mxmum Iterton Count (POCS- ACE) 100 Mxmum symbols loded 1e5 Symbol rte No of tme slots Wndow functon Blcmn-Hrs HPA Model SSPA No of frmes 10 No of OFDM symbols/ frme 4 Bndwdth 5 MHz Oversmplng fctor 4 4. GMM AS A POST CLASSIFIER th Consder dtset{ x }, where = 1,,3... N, where x s the vlue of the sgnl model nd N s the totl number of sgnls n dtset. Ths model lwys ssumes mxture model n generl [8]. It comprses of c Gussn densty components ntlly nd t s mxed wth the th prmeters θ = U Σ } n the component. {, For GMM, the probblty densty of x s determned s follows P( x π, θ ) = c = 1 π P( x θ ), Where θ = { θ1, θ,... θ c} s the prmeter of ll the components. π s the mxng of the component. th 45 edtor@eme.com

9 Hrumr Rjguru nd Sunl Kumr Prbhr The th Gussn s denoted s follows P( x θ ) = 1 (π ) Σ ( x exp T 1 u ) Σ ( x Where u nd Σ re the men nd covrnce mtrx respectvely. Expectton mxmzton s used to estmte the prmeters ( θ, π ) n n tertve fshon [9]. Thus the clncl trget vlues re mtched to the obtned clssfed vlues usng GMM. 5. RESULTS AND DISCUSSION 5.1 PAPR Results for DSTBC MIMO-OFDM Systems usng POCS-ACE Technque Intlly the PAPR reducton for the Dfferentl Spce Tme Bloc Coded- Multple Input Multple Output-OFDM Systems usng POCS-ACE lgorthm re smulted ccordng to the smulton prmeters lsted n Tble 1. u ) () (b) (c) Fgure 4 PAPR reducton usng POCS-ACE lgorthm for dfferent modulton formts () DQPSK b) 16-QAM (c) 64-QAM 46 edtor@eme.com

10 Development of n Effcent Eplepsy Clssfcton System from EEG sgnls for Telemedcne Applcton From the Fgures 4 (),(b) nd (c) t s observed tht the PAPR reducton lgorthm cn be successfully mplemented usng three dfferent modulton schemes,.e. DQPSK, 16-QAM nd 64-QAM. The fgure 4 shows the PAPR reductons for only ten tertons but n Tble II, the vlues hs been computed for dfferent vlues of tertons, mxmum of 100.The Bt Error Rte Anlyss s evluted under the three modulton schemes s shown n the Fgure 5. for x systems. DQPSK s chosen becuse ths formt provdes very promsng lterntve s t, le QPSK, trnsmts bts per symbol nd hence the symbol rte s hlf the bt rte wth somewht reduced complexty of the system. In DQPSK, the symbol nformton s encoded s the phse chnge from one symbol perod to the next rther thn s n bsolute phse. In ths cse, the recever hs to detect phse chnges nd not the bsolute vlue of the phse, whch vods the need for synchronzed locl crrer. For bnry formts, duo bnry nd dfferentl phse shft eyng (DPSK) re close to the theoretcl lmt, whch mes t dffcult to cscde multple dd drop multplexers long trnsmsson ln n rel tme. The nrrower spectrum of mult-level formts such s QPSK nd dfferentl qud phse shft eyng (DQPSK) therefore enbles both hgh spectrl effcency. For performnce comprson, the performnce of DQPSK hs been done wth the most wdely used modulton scheme for PAPR reducton, tht s, 16-QAM nd 64-QAM respectvely. Modulton Scheme used Tble PAPR Reducton for up to 100 tertons n DSTBC MIMO-OFDM System PAPR reducton n db (0 tertons) PAPR reducton n db (40 tertons) PAPR reducton n db (60 tertons) PAPR reducton n db (80 tertons) PAPR reducton n db (100 tertons) DQPSK QAM QAM From the Tble, t s shown tht the PAPR reducton vres for vrous types of terton vlues. When the tertons re reched upto 100 the mxmum PAPR reducton s cheved nd PAPR reducton s cheved by bout 5 db when DQPSK modulton scheme s cheved. If the opton s consdered to be more thn 100 tertons lso the reducton n PAPR s sturted. Fgure 5 Bt Error Rte Anlyss usng three dfferent modulton schemes for x System 47 edtor@eme.com

11 Hrumr Rjguru nd Sunl Kumr Prbhr From the Fgure 5, t s cler tht DQPSK modulton scheme produces low BER when compred to the other two modulton schemes. In DQPSK modulton scheme, the symbol nformton s encoded s the phse chnge from one prtculr symbol perod to the next symbol perod rther thn consderng t s n bsolute phse. Under such cse, the recever cnnot detect the bsolute vlue of the phse becuse t detects only the phse chnge. As result the need for the synchronzed locl crrer s voded nd hence t produces very less BER. Dfferentl qudrture phse-shft-eyng (DQPSK) modulton provdes promsng lterntve s t, le QPSK, trnsmts bts per symbol nd hence the symbol rte s hlf the bt rte wth somewht reduced complexty of the system. Ths type of DQPSK scheme s hghly tolernt to chromtc dsperson (CD) nd polrzton-mode dsperson (PMD). The hgh spectrl effcency of DQPSK mes t del for ultr long hul trnsmsson s well. 5. Clssfcton Results wth GMM t the Recever Sde For ICA s dmensonlty reducton technques nd GMM s Post Clssfer t the recever sde, bsed on the Qulty vlues, Tme Dely nd Accurcy the results re computed n Tble 3 respectvely. The formule for the Performnce Index (PI), Senstvty, Specfcty nd Accurcy re gven s follows PC MC FA PI = 100 PC Where PC Perfect Clssfcton, MC Mssed Clssfcton, FA Flse Alrm, The Senstvty, Specfcty nd Accurcy mesures re stted by the followng PC Senstvty = 100 PC + FA PC Specfcty = 100 PC + MC Senstvty + Specfcty Accurcy = 100 The Qulty Vlue Q V s defned s Q V C = ( R + 0.) *( T * P + 6* P ) f dly dct msd Where C s the sclng constnt, R f s the number of flse lrm per set, T dly s the verge dely of the onset clssfcton n seconds P dct s the percentge of perfect clssfcton nd P msd s the percentge of perfect rs level mssed The tme dely s gven s follows PC MC Tme Dely = The Specfcty nd Senstvty Anlyss for the pplcton of ICA s dmensonlty reducton technque followed by the pplcton of GMM s Post Clssfers s shown n Fgure 6. The Tme Dely nd Qulty Vlue Anlyss for the pplcton of ICA s dmensonlty reducton technque followed by the pplcton of GMM s Post Clssfers s shown n Fgure 7. Smlrly the Performnce Index nd Accurcy Anlyss for the 48 edtor@eme.com

12 Development of n Effcent Eplepsy Clssfcton System from EEG sgnls for Telemedcne Applcton pplcton of ICA s dmensonlty reducton technques followed by the pplcton of GMM s Post Clssfers s shown n Fgure 8. Averge Senstvty Mesures Epoch 1 Senstvty Epoch Senstvty Epoch 3 Senstvty Senstvty Mesures Number of Ptents Fgure 6 Averge Senstvty Mesures Averge Accurcy Mesures Accurcy Mesures Epoch 1 Accurcy Epoch Accurcy Epoch 3 Accurcy Number of Ptents Fgure 7 Averge Accurcy Mesures Averge Tme Dely Mesures Tme Dely Epoch 1 Tme Dely Epoch Tm Dely Epoch 3 Tme Dely Number of Ptents Fgure 8 Averge Tme Dely Mesures 49 edtor@eme.com

13 Hrumr Rjguru nd Sunl Kumr Prbhr Tble 3 Averge Vlues for ll the ptents usng GMM s Post Clssfer wth ICA s Dmensonlty Reducton Technque Prmeters Epoch-1 Epoch- Epoch-3 Averge PC MC FA PI Senstvty Specfcty Tme Dely Qulty Vlue Accurcy CONCLUSION Thus the EEG proves to be the most promnent tool used n the clncl dgnoss nd montorng of eplepsy nd other relted brn dsorders. In ths pper, ntlly the dmensons of the rw EEG sgnl were reduced wth the help of ICA nd t s then trnsmtted through the DSTBC MIMO OFDM system. For the DSTBC MIMO-OFDM System, PAPR reducton lgorthm clled Projecton onto Convex Sets Actve Constellton Extenson (POCS-ACE) Algorthm ws employed to reduce the PAPR. The recever sde ws ncorported wth Gussn Mxture Model (GMM) Clssfer to clssfy from the eplepsy from EEG sgnls. An overll ccurcy of 93.75% s reported n ths wor wth Perfect Clssfcton rte of 87.5%. The verge Qulty Vlue Reported s round 0.09 n ths wor. If the PAPR nlyss s consdered, then ths wor hs PAPR reducton of bout 5 db when DQPSK modulton scheme s mplemented for the system. Thus for EEG telemedcne pplcton ths wor s found to be hghly useful. Future wors my ncorporte the usge of dfferent dmensonlty reducton technques, dfferent clssfcton technques nd other PAPR reducton technques for the better trnsmsson nd recepton of EEG sgnls. REFERENCES [1] Al-Hmd, Hmd, nd Ry Chen. Redundncy mngement of multpth routng for ntruson tolernce n heterogeneous wreless sensor networs. Networ nd Servce Mngement, IEEE Trnsctons on 10, no. (013): [] Byun, Heejung, nd Junglo Yu. Adptve duty cycle control wth queue mngement n wreless sensor networs. Moble Computng, IEEE Trnsctons on 1, no. 6 (013): [3] Eswrmoorthy nd Uthyumr Anlyss of Bomedcl EEG Sgnls usng Wvelet Trnsforms nd multultfrctl Anlyss,IEEE EMB Mgzne vol. 30 (010): [4] Lehmnn, C, Koeng, T, Jelc, L & Ders T, Applcton nd comprson of clssfcton lgorthms for recognton of Alzhemers dsese n electrcl brn ctvty (EEG), Journl of Neuroscence Methods, vol. 161 (007), [5] Leon, D, Isemds, Deng-Shen Shu, Cho Vlt Wongse W, Scellres, JC & Prdlos, PN Adptve Epleptc Sezure Predcton System, IEEE Trnsctons on Bomedcl Engneerng, vol.50, no.5 (003), [6] Schuyler, R, Whte, A, Stley, K & Cos, KJ, Epleptc sezure detecton, IEEE Eng. Med.Bol. Mg (007), pp edtor@eme.com

14 Development of n Effcent Eplepsy Clssfcton System from EEG sgnls for Telemedcne Applcton [7] Wng, D & Hung, GB, Proten sequence clssfcton usng extreme lernng mchne, Proceedngs of IEEE Interntonl Jont Conference on Neurl Networs, Montrel, (005), pp [8] Howrd Loseless nd Lossy compresson of text mges by soft pttern mtchng, Proceedngs of Dt Compresson conference, Mrch/Aprl (1996) pp:10-19 [9] Shyh, Lossy Compresson of Nosy Imges, IEEE Trnsctons on Imge Processng, Vol.7, no: 1 (00), [10] McShrry, PE & Clfford, GD, A comprson of nonlner nose reducton nd Independent Component Anlyss usng relstc dynmcl model of the electrocrdogrm, Proc of SPIE Interntonl Symposum on Fluctutons nd Nose; (004),pp [11] Lozno A et l, Hgh-SNR Power Offset n Mult-ntenn Communcton. IEEE Tr.Comm, (005), pp [1] Troh V et l Spce-tme bloc codes form orthogonl desgns, IEEE Trns. J.Sel.Ares Commun., 18:7 (000), [13] Troh V, Spce-tme codes for hgh dt rte wreless communcton, IEEE Trns.Inform Theory, vol.44, no. (1998), pp [14] Xodong Zhu et l, Low-BER Clppng Scheme for PAPR Reducton n STBC MIMO- OFDM System Wreless Pers. Commun (01), 65: [15] Ltnovc Z, SFBC MIMO-OFDM Pe-to-Averge Power Rto Reducton by Polyphse Interlevng nd Inverson, IEEE Comm. Letters (006), pp [16] Krongold B, Jones D, An ctve-set pproch for OFDM PAR Reducton v tone reservton, IEEE Tr. Sgnl Proc., (004) pp [17] Gnesn G nd Stoc P, Dfferentl modulton usng spce-tme bloc codes, IEEE Sgnl Process. Lett., 9:, (00) [18] Bssem et l, PAPR Reducton method for STBC MIMO-OFDM Systems usng SLM n combnton wth sub bnd permutton, Thrd Interntonl Conference on Wreless nd Moble Communctons (ICWMC 07) (007), Gudeloupe, [19] Mzhou Tn et l, STBC MIMO-OFDM Pe-to-Averge Power Rto Reducton by Cross-Antenn Rotton nd Inverson, In IEEE Comm. Letters Vol.9, No.7 (005) [0] Bg et l, A ZMCT Precodng bsed STBC MIMO-OFDM system wth reduced PAPR, Ntonl Post Grdute Conference (NPC), Sngpore (011) [1] Alhrb F.S et l, A Combned SLM nd Closed-Loop QO-STBC for PAPR Mtgton n MIMO-OFDM Trnsmsson, Proceedngs of EURASIP (008) [] Jngbo et l, PAPR Reducton n Blnd MIMO OFDM Systems bsed on Independent Component Anlyss, 17 th Europen Sgnl Processng Conference (EUSIPCO),009, Glsgow, Scotlnd [3] Hyun-Seng Joo et l, A Blnd SLM PAPR reducton scheme usng Cyclc Shft n STBC MIMO-OFDM System Interntonl Conference on Informton nd Communcton Technology Converge (ICTC), Jeju, 010 [4] Y-Sheng Su, A Computtonlly effcent PAPR reducton schemes for STBC-OFDM Systems, Proceedngs of TENCON-Sprng (013) [5] Sunl Kumr P et l, Performnce Anlyss of PAPR reducton n STBC MIMO-OFDM System, Proceedngs of the Ffth Interntonl Conference on Advnced Computng (ICOAC ) held t Mdrs Insttute of Technology Cmpus, Ann Unversty-Chenn, December 18-0 (013), pp: [6] Krongold B nd Jones D, PAR Reducton n OFDM v Actve Constellton Extenson, IEEE Tr. Brodcstng (003), pp edtor@eme.com

15 Hrumr Rjguru nd Sunl Kumr Prbhr [7] Zhfe Tng et l, Projecton onto convex sets super-resoluton mge reconstructon bsed on wvelet b-cubc nterpolton, Interntonl Conference n Electronc nd Mechncl Engneerng nd Informton Technology, 1-14 August, 351(011)-354, Chn [8] Slehn et l, POCS method for photo coustc tomogrphy wth non-negtve constrnt, 35 th Annul Interntonl Conference of the EMBC, Os, 3-7 July (013). [9] Armstrong J, Pe-to-Averge Power Reducton for OFDM by Repeted Clppng nd Frequency Domn Flterng, Electr. Letters, pp , Feb (00). [30] Sunl Kumr Prbhr, Hrumr Rjguru, ICA, LGE nd FMI s Dmensonlty Reducton Technques followed by GMM s Post Clssfer for the Clssfcton of Eplepsy Rs Levels from EEG Sgnls, 9 th IEEE Europen Modellng Symposum,October 6-8, Mdrd, Spn (015) [31] Hrumr R, Sunl Kumr P, Performnce Comprson of EM, MEM, CTM, PCA, ICA, Entropy nd MI for Photoplethysmogrphy Sgnls, Bomedcl nd Phrmcology Journl, Vol.8, No.1(015), pg no: [3] Indr, V., Vsnthumr, R. nd Sugumrn, V. Smple Sze Determnton for Clssfcton of EEG Sgnls usng Power Anlyss n Mchne Lernng Approch. Interntonl Journl of Advnced Reserch n Engneerng nd Technology, 3(1), 01, pp [33] V.V.Rmlngm, S.Gnesh umr, V. Sugumrn, Anlyss of EEG Sgnls Usng Dt Mnng Approch, Interntonl Journl of Computer Engneerng & Technology (IJCET), Volume 3, Issue 1, Jnury- June (01), pp edtor@eme.com