Amerca Joural of Sgal Processg, (5): 45-49 DOI:.593/.asp.5.8 Geder Classfcato from ECG Sgal Aalyss usg Least Square Support Vector Mache Raesh Ku. rpathy,*, Ashutosh Acharya, Sumt Kumar Choudhary Departmet of Bo Medcal Egeerg, I, Rourkela, Ida Departmet of Electrocs ad Comm. Egeerg, SI, Bhubaeswar, Ida Abstract I ths preset paper t deals wth the Geder Classfcato from ECG sgal usg Least Square Support Vector Mache (LS-SVM) ad Support Vector Mache (SVM) echques. he dfferet features extracted from ECG sgal usg Heart Rate Varablty (HRV) aalyss are the put to the LS-SVM ad SVM classfer ad at the output the classfer, classfes whether the patet correspodg to requred ECG s male or female. he least square formulato of support vector mache (SVM) has bee derved from statstcal learg theory. SVM has already bee marked as a ovel developmet by learg from examples based o polyomal fucto, eural etworks, radal bass fucto, sples ad other fuctos. he performace of each classfer s decded by classfcato rate (CR). Our result cofrms the classfcato ablty of LS-SVM techque s much better to classfy geder from ECG sgal aalyss terms of classfcato rate tha SVM. Keywords ECG, HRV, SVM, LS-SVM, CR. Itroducto Geder s almost ts most salet feature, ad geder classfcato accordg to ECG s oe of the most Challegg problems perso detfcato Bo metrcs[]. Compared wth other research topcs Bo metrcs, the academc researches o geder classfcato s less. I realty, successful geder classfcato wll boost the performace of Patet recogto large Medcal database. Fro m last two decades It was observed that a varety of predcto models have bee proposed the mache learg that clude tme seres models, regresso models, adaptve euro-fuzzy ferece systems (AFIS), artfcal eural etwork (A) models ad SVM models[]. Due to the effectveess ad smoothess of A model, t s wdely used varous felds lke patter recogto, regresso ad classfcato. For classfcato ad o-lear fucto estmato, the recetly proposed SVM techque s a ovatve kd of mache learg method troduced by Vapk ad co-workers[3, 4, 5]. hs method s further ehaced by varous vestgators for dfferet applcatos lke classfcato, feature extracto, clusterg, data reducto ad regresso dfferet dscples. SVM have remarkable geeralzato performace ad may more advatages over other methods, ad hece SVM has * Correspodg author: raesh.tr@gmal.com (Raesh Ku. rpathy) Publshed ole at http://oural.sapub.org/asp Copyrght Scetfc & Academc Publshg. All Rghts Reserved attracted atteto ad gaed extesve applcato. Suykes ad hs group[6] have proposed the use of LS-SVM for smplfcato of tradtoal of SVM. Apart from ts use classfcato varous areas of patter recogto, t has bee extesvely used hadlg regresso problems successfully[7, 8]. I LS-SVM, a set of oly lear equato (Lear programmg) s solved whch s much easer ad computatoally more smple whch made t advatageous tha SVM. I the preset study, both SVM ad LS-SVM classfers have bee desged, traed ad tested usg varous kerel fuctos lke lear ad (Radal Basc Fucto) RBF kerel. RBF kerel LS-SVM gves better performace terms of classfcato rate amog other classfers.. Materals ad Methods F gure. Custom made ECG devce
46 Raesh Ku. rpathy et al.: Geder Classfcato from ECG Sgal Aalyss usg Least Square Support Vector Mache usg Least Square Support Vector Mache I ths curret study, the ECG of dfferet people was recorded. A -house developed ECG data acqusto system was used for ths study. he ECG system maly composed of two parts, vz. ECG electrodes ad ECG-DAQ. ECG-DAQ, a USB-based devce, helped recordg ECG sgals PC. Fgure shows the ECG devce coected wth PC whch was used for data recordg our study. Bomedcal Starter Kt from atoal Istrumet software was used to extract the dfferet tme ad frequecy doma features of HRV (e.g. mea heart rate (HR), mea RR, Mea, stadard devato of RR tervals (SD), root mea square of successve dfferece (RMSSD), 5, p5 ad SD ad SD of Pocare plot). he HRV features were the aalysed by o-lear statstcal aalyss usg Classfcato ad regresso trees (CAR) ad Boosted tree (B) classfcato to determe the sgfcat features SAISICA (v7) fgure. A combato of the features was subsequetly used for Geder Classfcato usg SVM ad LS-SVM MALA B.. After predcto of sgfcat features lke(rr Mea, RR Stadard Devato(std.),HR mea, HR Std, root mea square of successve dfferece (RMSSD), 5, p5,lf peak, HF Peak, LF Power, HF Po wer, LF/HF rato) fro m CA R ad B the these feature values are the put to both SVM classfer. At the output of both classfers we wll classfy two classes.e. (boy ad grl).the bary values assgs to grl class as ad for boy class the value s.after classfcato we compare both the values terms of Classfcato rate (CR). he put data are ormalzed before beg processed the etwork as follows: I ths method of ormalzato, the maxmum values of the put vector compoet are gve by:,max = max ( ( p) ), p= p, =. Where (p) s the put features ad p s the umber of patters the trag set ad testg set,, or ( p) ( p) =, p =.p, =,,max After ormalzato, the put varables le the rage of to. F gure. Sgfcat Iput parameters for geder classfcato 3. SVM Classfcato SVM techque s a attractve kd of mache learg method troduced by Vapk ad co-workers[3, 4, 5]. hs method s further modfed by varous scetsts for dfferet applcatos lke classfcato, feature extracto, clusterg, data reducto ad regresso dfferet dscp les of egeerg. Our preset aalyss s based o the classfcato of bary class data by employg SVM techque. hs method bulds a Hyperplae for separato of data to two classes smple bary classfcato of lear separable trag data vector ( x, x, x3... x ) dmesoal Space. A class decso fucto assocated wth Hyperplae s weghted sum of trag data set ad bas are symbolzed[,, ] as yx ( ) = w φ( x) + b (3.) Where w ad b are weght vector ormal to Hyperplae ad bas value respectvely. ew test data accordg to the classfcato by SVM classfer s assged to a class accordg to sg of decso fucto as: estg data belogs to class-(male class) f w φ ( x ) + b (3.) est data belogs to class-(female class) f Ad w φ ( x ) + b (3.3) w φ ( x) + b= (3.4) Is the decso boudary correspodg to Weght vector ad bas value for optmal Hyperplae. Support vectors are obtaed by maxmzg the dstace betwee the closest trag pot ad the correspodg Hyperplae. hs ca be doe by maxmzg the marg defed as M = same w as mmzato of w (3.5) Uder the methodologes y ( w φ ( x ) + b) (3.6) Dfferet umber of mathematcal algorthms exsts for
Amerca Joural of Sgal Processg, (5): 45-49 47 determg the value of weght ad bas uder the codto (3.5) ad (3.6). Oe of the most effcet method used SVM s Quadratc Optmzato problem. Its soluto of the problem volves costructo of dual problem wth the use of Lagrage multp ler α whch s gve as follows: α αα yy xx = = = (3.7) he equato (3.7) s maxmzed uder the codtos α y = Ad x For all value of,,... =.After solvg Quadratc optmzato problem, the values of weght ad bas are obtaed as Ad Where w= α yx (3.8) b y wx = (3.9) x s support vector for each ozero value ofα. Hece, the classfcato fucto for a test data pot x s er product of support vector ad test data pot, whch s gve as follows y( x) = αyx x+ b (3.) SVM maps -dmesoal data vector to a d-dmesoal feature space (d>) wth help of a mappg fucto for bary classfcato of olear trag data pots. hs Mappg fucto or kerel fucto provdes a Hyperplae wh ch separate the classes hgh dmesoal feature space. Usg stadard Quadratc Programmg (QP) optmzato techque, the Hyperplae maxmzes the marg betwee classes. hose data pot whch are Closest to the Hyperplae are used to measure the marg ad amed as support vectors. I dual formulato of quadratc optmzato problem stead of usg dot product of trag data pots hgh dmesoal feature space, kerel trck s used. Kerel fucto defes the er product of trag data pots hgh dmesoal feature space. hus the kerel fucto s defed as kx (, x) = φ ( x) φ ( x) (3.) he varous advatages of kerel fucto are, It reduces the mathematcal as well as the computatoal complexty hgher dmesoal feature space. I ths paper, commoly used kerel fuctos are lear, polyomal, radal Gaussa ad sgmod are defed as follows: k( x, x ) = xx (Lear kerel Fucto) k( x, ) ( ) d x = xx + c d ad c > For (Polyomal Kerel Fucto) x x σ kx (, x) = e For σ > (Radal Gaussa kerel fucto) he ew classfcato fucto usg kerel fucto s defed as follows: y( x) = αyk ( x, x) + b (3.) Further, the testg data pots are classfed wth these bary class traed model ad fal decso about class of data pot s take o the bass of maorty votg of class. he performace of model s determed by classfcato rate equal to (CR) = (P+F)/(P++FP+F) (3.3) Where P=rue Postve, =rue egatve, FP=False Postve ad F=False egatve. 4. LS-SVM Classfcato he formulato of LS-SVM s troduced as follows. Gve a trag set {, } xy = wth put data x R ad y +.he correspodg bary class labels {, } followg Classfcato model ca be costructed by usg o-lear mappg fucto φ (x)[]. ( ) y w x b = φ + (4.) { w φ ( x ) + b, f y = + } { w φ ( x ) + b,f y = } Where w weght s vector ad b s the bas term[9,, ]. As LS-SVM, t s ecessary to mmze a cost fucto C cotag a pealzed regresso error for bary target, as follows: m C( we, ) = ww+ γ e (4.) = Subect to equalty costrats y = w φ x + b+ e =,,..., (4.3) ( ), he frst part of ths cost fucto s a weght decay whch s used to regularze weght szes ad pealze large weghts. Due to ths regularzato, the weghts coverge to smlar value. Large weghts deterorate the geeralzato ablty of the LS-SVM because they ca cause excessve varace. he secod part of eq. (4.) s the regresso error for all trag data. he parameter γ, whch has to be optmzed by the user, gves the relatve weght of ths part as compared to the frst part. he restrcto suppled by eq. (4.3) gves the defto of the regresso error. o solve ths optmzato problem, Lagrage fucto s costructed as L( wbe,,, α) = w + γ e = α{ w φ( x) + b+ e y} (4.4) = Where α are the Lagrae multplers. he soluto of eq. (4.4) ca be obtaed by partally dfferetatg wth respect to w, b, e ad α he αφ ( ) γ φ( ) (4.5) w= x = e x = =
48 Raesh Ku. rpathy et al.: Geder Classfcato from ECG Sgal Aalyss usg Least Square Support Vector Mache usg Least Square Support Vector Mache Where a postve defte Kerel s used as follows: (, ) φ( ) φ( ) K x x = x x (4.6) A mportat result of ths approach s that the weghts (w) ca be wrtte as lear combato of the Lagrage multplers wth the correspodg data trag (x ). Puttg the result of eq. (4.5) to eq. (4.), the follo wg result s obtaed as y = αφ ( x) φ( x) + b (4.7) = For a pot y to evaluate t s: (4.8) ( ) φ( ) y = αφ x x + b = he vector follows from solvg a set of lear equatos: α y A = b Where A s a square matrx gve by I K + A = γ (4.9) (4.) Where K deotes the kerel matrx wth th elemet eq. (4.5) ad I deotes the detty matrx, =[..]. Hece the soluto s gve by: α y = A b (4.) All Lagrage multplers (the support vectors) are o-zero, whch meas that all trag obects cotrbute to the soluto ad t ca be see from eq. (4.) to eq. (4.). I cotrast wth stadard SVM the LS-SVM soluto s usually ot sparse. However, by prug ad reducto techques a sparse soluto ca easly acheved. Depedg o the umber of trag data set, ether a teratve solver such as cougate gradets methods (for large data set) ca be used or drect solvers, both cases wth umercally relable methods. I applcato volvg olear regresso t s ot φ x, φ x eq. eough to chage the er product of ( ) ( ) (4.7) by a kerel fucto ad the th elemet of matrx K equals to eq. (4.5). hs shows the followg olear regresso fucto: y = αk( x, x) + b (4.) = For a pot x to be evaluated, t s: (, ) (4.3) y = α K x x + b For LS-SVM, there are lot of kerel fucto (Radal bass fucto, Lear polyomal), sgmod, bsple, sple, etc. However, the kerel fucto more used s a smple Gaussa fucto, polyomal fucto ad RBF. hey are defed by: x x K( x, x) = exp σ sv d (, ) ( ) (4.4) K x x = x x + t (4.5) σ Where d s the polyomal degree ad s the squared sv varace of the Gaussa fucto, to obtaed support vector t should be optmzed by user. I order to acheve a good geeralzato model t s very mportat to do a careful model selecto of the tug parameters, combato wth the regularzato costat γ. 4. Result ad Dscusso I ths study, the proposed modellg are carred out; 5 sets of put-output patters used for trag both etworks ad for testg purpose the remag sets are used. he software programs developed are used for mplemetato usg MALAB verso.. I the begg, SVM etwork was traed wth commoly used kerels lke Lear ad RBF Kerel Fucto. able(-) shows the performace of model meas of Percetage of Classfcato rates(cr) obtaed from the processg of testg data wth respect to SVM ad LS-SVM model wth RBF kerel ad Lear kerel. able. Cofuso matrx for est data sheet for SVM Modelg (usg Lear Kerel Fucto) rue Class Predcted Class (Boy) (Grl) Classfcato (Boy) 5 67% (Grl) 3 a ble. Cofuso matrx for est data sheet for SVM Modelg (usg RBF Kerel Fucto) rue Class Predcted Class (Boy) (Grl) Classfcato (Boy) 5 84% (Grl) 5 a ble 3. Cofuso matrx for est data sheet for LS-SVM Modelg (usg Lear Kerel Fucto) rue Class Predcted Class (Boy) (Grl) Classfcato (Boy) 5 75% (Grl) 4
Amerca Joural of Sgal Processg, (5): 45-49 49 a ble 4. Cofuso matrx for est data sheet for LS-SVM Modelg (usg RBF Kerel Fucto) rue Class Predcted Class (Boy) (Grl) Classfcato (Boy) 5 9% (Grl) 5. Coclusos 6 hs paper proposes a geder classfcato from ECG sgal usg LS-SVM ad SVM techque. he dfferet put features extracted from HRV aalyss are drectly feed to SVM ad LSSVM classfer. Both the classfers are desged, traed ad tested. LS-SVM classfer wth RBF kerel gves better classfcato rate of 9% amog other models. hs shows LS-SVM as promsg results for the classfcato of the Geder based o HRV ad ECG sgal aalyss. REFERECES [] Z.H Wag, Z.C Mu, Geder classfcato usg selected depedet-features based o geetc algorthm Proceedgs of the Eghth Iteratoal Coferece o Mache Learg ad Cyberetcs, Baodg, -5 July 9. [] W. Bartkewcz euro-fuzzy approaches to short-term electrcal load forecastg,ieee Explorer.vol 6, pages 9-34,. [3] V. Vapk, he ature of Statstcal Learg heory, Sprger, ew York, 995. [4] V. Vapk, Statstcal Learg heory, Wley, ew York, 998a. [5] V. Vapk, he support vector method of fucto estmato, I olear Modelg: Advaced Black-box echques, Suykes J.A.K., Vadewalle J. (Eds.), Kluwer Academc Publshers, Bosto, pp.55-85, 998. [6] J.A.K Suykes, Va Gestel, J De Brabater, B De Moor, J Vadewalle, Least Square SVM. Sgapore: World Scetfc,. [7] D. Habay, A expert system based o least square support vector maches for dagoss of the valvular heart dsease, Expert Systems wth Applcato, 36(4), part-, 8368-8374.9 [8] Z Sahl,,A.Mekhald, R.Boudssa, S.Boudrahem, Predcto parameters of dmesog of sulators uder o-uform cotamated codto by multple regresso aalyss, Electrc Power System Research, Elsever press. [9] G. Zhu, D.G Blumberg, () Classfcato usg ASER data ad SVM Algorthms: he case study of Beer Sheva, Israel, Remote Ses. of Evro., 8, 33-4. [] V. Kecma, (5) Support Vector Maches: heory ad Applcatos Lpo Wag (Ed.) Support vector mache a troducto, Sprger-Verlag Berl Hedelberg,-48. [] V. Kecma, () Learg ad soft computg: Support vector mache, eural etwork ad Fuzzy logc models, MI Press, Cambrdge, MA, 48-76.