Performance Evaluation of ANFIS for Classification of PCG Signal Using Wavelet Transform

Transcription

1 Internatonal Journal of Advanced Research n Electroncs and Communcaton Engneerng (IJARECE) Performance Evaluaton of ANFIS for Classfcaton of PCG Sgnal Usng Wavelet Transform Ajay Kumar Roy, Abhshek Msal Abstract Heart s the most mportant part of any lvng beng, and hence for the humans also. In recent years, t has been seen that the deaths have been hghly ncreased due to heart dsease all over the world. The need of fast and accurate detecton of heart dsease has been ncreased and acqures the hghest mportance n the feld of research. There has been a lot of work done for the ECG sgnal for ts segmentaton, and classfcaton, but t does not fulfll the requrement of heart dsease detecton so accurately. The dscovery PCG sgnals gave a great hope, and hence a lot of work has also been done of PCG Sgnal segmentaton and feature extracton. In contnuaton, the expermental results from dfferent lteratures need to be studed and to be used for detecton of heart dsease. The PCG sgnal has been analyzed and the feature has been extracted based on the best method shown n the lterature. Further, these feature features has been used for the classfcaton of the PCG sgnal for dfferent heart dsease acqured by the patent. ANFIS s the tool used for the classfcaton and on tranng the system t proves to be the best for classfcaton of heart dsease. Index Terms ANFIS, Wavelet, knn, PCG, Membershp Functon (MFs). I. INTRDUCTIN The heart s the most mportant part of any lvng beng, and hence for the humans also. In recent years, t has been seen that the deaths have been hghly ncreased due to heart dsease all over the world [1]. The requrement of accurate detecton of heart dsease has forced researchers to develop a system whch can help to detect the dsease and cure t as soon as possble. The dscovery of PCG sgnal s a pave to drag the concentraton towards ths topc. In recent years a lot of work has been done on ECG sgnal and a lot of nformaton has been extracted and has been used for dfferent purposes. Although ECG sgnal has been analyzed to a greater level, t s not suffcent to detect the heart dsease because t deals wth the electrcal behavor of the heart, whle abnormaltes n heart are mostly due to change n shape of the chambers of the heart. The blood flow from heart to lungs and then from lungs to heart and dfferent parts of the body, ths flow of bloods wth Manuscrpt receved July, 014. Ajay Kumar Roy, Department of Electroncs and Telecommuncaton Engneerng, Chhatrapat Shvaj Insttute of Technology Durg, Durg, Inda. Abhshek Msal, Department of Electroncs and Telecommuncaton Engneerng, Chhatrapat Shvaj Insttute of Technology Durg, Durg, Inda. specfc pressure and volume, produces the heart sound. These heart sounds are called as Phonocardogram sgnal. Phonocardogram sgnal s non-statonary sgnals wth a frequency of 10 KHz and provde the experenced doctors a path to fnd the dsease acqured by a person. Hearng to PCG sgnal the experenced doctors understand the dsease and cure accordngly. There s always exsts a chance of wrong detecton of dsease because of the doctor's nablty to hear the sound properly, hs perseverance and hs experence. Hence there s a need of developng a decson support system that can support doctors ndependent of ther experence and any unfavorable physcal condtons. The system ncludes the feature extracton of sgnal, here PCG sgnal, usng dscrete wavelet transform specally dabocous wavelet as ths can provde better nformaton than other wavelets lke her, smulate, coflet etc. Phonocardogram sgnal s a nonstatonary sgnal and hence dscrete wavelet transform s the best to analyze []. Then for the purpose classfcaton of PCG sgnal, Adaptve Neuro Fuzzy Inference System (ANFIS) has been used. The tranng of the system s done by the heart sound avalable n varous webste of medcal scences on the World Wde Web. ANFIS system s powered wth the feature of Neural Network as well as Fuzzy Inference System both and hence can classfy the PCG sgnal more accurately compared to other type of classfer lke knn classfer or NN classfer. Despte the recent developments on echocardography for heart examnatons are now avalable, cardac auscultaton remans the most mportant and screenng, dagnostc method for early dagnoss of heart valve dseases. The necessty of early detecton s to gve a warnng sgn for further nvestgatons before the serous pathologcal condtons of heart dseases occur. In addton, detecton of heart sounds can be obtaned by a technque known as phonocardography. Phonocardography dsplays the graphcal representaton of the heart sounds. It s easy to use and non-nvasve. It provdes the dagnostc nformaton for detecton of the abnormal functon of the cardac valves n clncal practce. All Rghts Reserved 014 IJARECE II. PRBLEM IDENTIFICATIN AND SLUTINS Phonocardogram sgnal s the nonstatonary sgnal, hence a perfect tool s requred to analyze the sgnal. Varous methods of feature extracton are avalable lke Fourer Transform, Fast Fourer Transform, Dscrete Fourer 1034

2 Internatonal Journal of Advanced Research n Electroncs and Communcaton Engneerng (IJARECE) Transform, Wavelet Transform, Dscrete Wavelet Transform etc. the selecton of approprate feature extractor s a bg challenge for the system. Lterature suggests that DWT provdes better feature compared to other feature extractor wth no loss n data [3]. Sgnal vares much n ts length and contans many redundant nformaton, these sgnals needs to be reshaped to get t n the requred format. Decdng the length of the sgnal whch can provde better results n feature extracton plays an mportant role. Varaton n length of the sgnal msgudes the system at the tme of tranng whch leads to wrong detecton. n analyss a length of 8555 has been observed to produce better results. After decdng the feature extracton method, the database s prepared so as to use t to tran the ANFIS system. There exsts a number of classfer agan decdng the best one s beg challenge. Lterature suggests that fuzzy systems produce better results compared to other classfer and ANFIS whch ncorporates the feature of neural network and fuzzy nference system produces the best results compare to other type of classfer. Then the nput sgnal whch needs to be detected for dsease are gven as an nput to the system, and here agan the feature of the nput sgnal s extracted and passed to the traned system for matchng the sgnal wth the database so as to produce the output as the name of the dsease acqured by the patent. Durng nput we agan have to take care of the sgnal that t should not have greater length, whch may lead to msclassfcaton of the sgnal. III. METHDLGY The detecton of the heart dsease va a decson support system s a tool for the doctors and especally for the novce doctors who know about the dsease and ther remedy, but due to lack of experence they are not able to detect the dsease partcularly. The system s developed usng MATLAB 01a. Ths system takes as an nput the heart sound, processes t and then produces the output as the name of the dsease acqured by the person whose heart sound s under test. The processed output s the name of dsease acqured by the patent. The process s hghly reactve and as soon as the dsease s dentfed by the doctors they can soon take the correctve measures to prevent the further losses that can happen to the patent due to the long term acquston of the dsease by the patent. The data flow dagram shown below presents the process followed by the system n the detecton of the heart dsease. The process nvolves fve dfferent stages n the development of the system. Intally the PCG sgnal s loaded to the system. The sgnal s then converted nto the standard format,.e. n the desred form and then t s gven as n nput to a functon whch extracts the feature of the PCG sgnal usng db10 wavelet. The Mtral Regurgtaton sgnal s analyzed by Daubeches Wavelet (db10) at a level of 5 and the analyzed sgnal can be obtaned as below. Fg -: Sgnal analyss usng db10 at level5 The graph shown n red color s the nput sgnal; the graph n blue color represents the approxmate nformaton through approxmate coeffcent whle graph n green color represents the detaled nformaton through detaled coeffcent. Through ths technque the system analyzes the nput sgnal and features lke Entropy, Energy and Varance of the sgnal can be extracted. Entropy provdes better nformaton about the PCG sgnal and hence can be used for the classfcaton of sgnal. For the purpose of classfcaton ANFIS (Adaptve Neuro-Fuzzy Inference System) has been used. The archtecture of ANFIS ncludes fve layers; Layer1 (Inputs Layer), Layer (Fuzzy AND peraton), Layer3 (Normalzaton), Layer4 (Normalzaton of Each Rule Frng Strength), and Layer5 (utput Layer), whch follows the process as shown below. x y X A 1 П w1 w1 N w 1 f 1 Load the PCG sgnal Convert the recorded PCG sgnal nto standard sgnal Call a functon whch wll extract Entropy as features of PCG sgnal usng Wavelet A Ʃ F Gve the extracted feature of the nput sgnal as an nput to the classfer and detect the dsease. Fg -1: Block dagram representaton of whole process Tran the ANFIS classfer wth the database of the sgnal. Fg -3: Archtecture of ANFIS system A Two Rule Sugeno Type ANFIS has rules of the form: All Rghts Reserved 014 IJARECE Y B 1 B П N w f w w x y 1035

3 Internatonal Journal of Advanced Research n Electroncs and Communcaton Engneerng (IJARECE) If x s A If x s A 1 and y s B1 THEN f1 p1x q1 y r1 and y s B THEN f px q y r For the tranng of the network, there s a forward pass and a backward pass. The forward pass propagates the nput vector through the network, layer by layer. In the backward pass, the error s sent back through the network n a smlar manner to Backpropagaton network. At Layer1 the output of each node s: 1, A ( x) for 1, (1) 1, B ( y) for 3,4 () So, 1, (x) the s essentally the membershp grade for x and y. The membershp functons could be anythng but here a bell shaped functon s used whch s gven by: 1... (3) A ( x) b x c 1 a where a, b, and c are the premse parameters. At Layer, every node s fxed. Ths s where the t-norm s used to AND the membershp grades gven by:, w A B ( x) ( y), 1, (4) At Layer3 fxed nodes are present, whch calculates the rato of the frng strengths of the rules gven by: w 3, w... (5) w w 1 At Layer4 the nodes are adaptve and perform the consequences of the rules: w f w ( p x q y r ) (6) 4, The parameters p, q and r n ths layer are to be determned and are referred to as the consequent parameters. decomposng the sgnal usng db10 should be chosen approprately. Here hghest eght values are taken nto account to gve as an nput to the system to tran t. The ANFIS system s traned by defnng the followng parameters along wth the database. Table 5.1: ANFIS archtecture and tranng parameters Modelng Descrpton Settng Number of nputs 8 Number of outputs 1 Type of nput membershp functons Number of nput membershp functons Type of output membershp functons gaussmf Lnear Number of Rules 56 Learnng rule (ptmzaton Method) Max. Epochs (Stoppng Crteron) 50 hybrd Error tolerance (Tranng Error Goal) 0.01 Intal step sze 1 As mentoned above, eght values are chosen for nput and the number of membershp functons s two hence the number of rules formed s 56 ((8 )). However, f the number of membershp functon and the number of nput ncreases, more number of rules wll be prepared, whch s good n one way, but the problem s the computaton complexty and so the system wll take more tme to tran tself moreover the software may run out of memory. Hence we have to choose an optmum value to tradeoff between number of nputs and number of membershp functons. A good par for the same has been obtaned as eght nputs wth two membershp functon whch s used n the proposed system. The structure and rule base of the system s shown below. At Layer5 there s a sngle node that computes the overall output, whch s gven by: w f w f (7) 5, w Ths s how, typcally, the nput vector s fed through the network layer by layer. The process dscussed above shows the process how an ANFIS system works. In the proposed work the system s traned by the database. The database s prepared by collectng many heart murmurs from dfferent places and then those sgnals are processed to obtan the entropy of the sgnals. The set of entropy nformaton of all the murmurs forms a database. The number of values of entropy.e. the number of values among all the values extracted by Fg -3: Structure of ANFIS system All Rghts Reserved 014 IJARECE 1036

4 Internatonal Journal of Advanced Research n Electroncs and Communcaton Engneerng (IJARECE) The nput sgnal s taken and then t s processed to get t nto the desred form so any extra nformaton that does not dsnfect the output. The fgure below shows one of the sgnals, and then the processed sgnal, the entropy of whch s gven as an nput to the ANFIS system. Fg -3: Structure of ANFIS system As soon as the system s traned t s ready to accept the test nput sgnal. The test sgnal s also processed to attan ts entropy and then the system, accordng to the fuzzy logc developed after tranng, dentfes the dsease represented by the test sgnal. Here at the pont of nputtng the test sgnal t should be taken nto consderaton that the test sgnal should also be changed to standard format accordng to the requrement of the system; a very short sgnal may not be able to provde suffcent nformaton for classfcaton whle a very lengthy system may confuse the system n the course of detecton of the dsease. Under these consderatons, the system s developed and produces the output as the name of dsease acqured by the patent. Fg -5: Input sgnal (Mtral Regurgtaton) IV. RESULTS AND DISCUSSIN The valdty of the proposed procedure s proven by means of extensve computer smulatons. The results obtaned for Aortc Stenoss, Aortc Regurgtaton, Mtral stenoss, Mtral Regurgtaton and Normal Heart Sounds s llustrated n ths secton. The fgure below shows the ANFIS structure whch s generated n MATLB whle smulaton of the system. The database s prepared n requred format and s passed on to the system for tranng purpose. The fgure below shows there are eght nput feature for the system whch are the entropy extracted as feature by db10 at level5 wth membershp functon, 56 rules and one constant output whch s partcularly the name of the dsease conceved by the patent. Fg -6: Transformed nput sgnal The nput sgnal s fuzzfed wth two membershp functons whch are represented below. The MF (Membershp Functon) can be ncreased more than two, but wth the ncrease n MF, the rules get ncreased and n turn t s observed that wth large MFs MATLAB may run out of memory. Fg -4: System Archtecture Fg -7: Membershp grade of nputs wth MF All Rghts Reserved 014 IJARECE 1037

5 Internatonal Journal of Advanced Research n Electroncs and Communcaton Engneerng (IJARECE) These membershp functons after fuzzfcaton gves the values to the next level of normalzaton and for the applcaton of rules formed by the system whch leads to tran the ANFIS system and to provde the result as the detected dsease. Whle tranng the system, varous errors have been ncurred by the system at each epoch of tranng. The RMSE wth respect to tranng epochs has been shown below. -manager s, Journal on Communcaton Engneerng and Systems, Volume 1. No. 1, 01, pp. no [] Abhshek Msal and G.R. Snha, Separaton of Lung Sound from PCG Sgnals Usng Wavelet Transform, Journal of Basc and Appled Physcs, Vol. 1 Iss., 01, pp no [3] Anta Dev, Abhshek Msal, Feature Extracton of PCG Sgnals usng DWT, Natonal Conference on Advances n Computng ( NCAC 13 ), 013, pp [4] Harun Ug uz., Adaptve neuro-fuzzy nference system for dagnoss of the heart valve dseases usng wavelet transform wth entropy, Neural Comput. & Applc. 01, [5] Abhshek Msal & Snha G.R., Denosng of PCG sgnal by usng wavelet transforms, Advances n Computatonal Research, 01, pp [6] Sabarmala Mankandan M., & Dandapat S., Wavelet-Based ECG and PCG Sgnals Compresson Technque for Moble Telemedcne, Internatonal Conference on Advance computng and communcatons, 007, pp [7] S. R. Debbal, J. Agzaran and D. Abbott., ptmal wavelet denosng for phonocardograms, Mcroelectroncs Journal, Vol. 3, 001, [8] Francesco Bertell, and Salvatore Serrano, Bometrc Identfcaton Based on Frequency Analyss of Cardac Sounds, IEEE Transacton on Informaton forenscs and securty, 007, [9] Anta Dev Twar, Abhshek Msal, G.R. Snha. Analyss of PCG Sgnals usng Daubeches Wavelet Famly, -manager s Journal on Communcaton Engneerng and Systems, 013, 3-9. Fg -8: Error output whle tranng the system The graph above shows that the error s ncreasng and decreasng wth the number of epochs selected, and has mnmum error at the ntal pont. The advantage of usng ANFIS s that t uses the parameter wth respect to the mnmum error, no matter whether t occurs n the frst or last or at any epoch n between. The tranng epochs can be ncreased to any level, but t has been observed that the computng complexty and tme requred ncreases sgnfcantly wth ncrease n tranng epochs. Ajay Kumar Roy: Pursung M.E. at Chhatrapat Shvaj Instture of Technology, Durg. Completed B.E. n 010. Publshed paper n IJRETM, IJCA, NCRAIT-014, and AICN-014. Workng on PCG sgnal classfcaton. Member of ISTE. Abhshek Msal: M.E., B.E. Publshed papers on varous journals lke IJCA, -Manager etc. Workng n varous felds lke Image Processng, Sgnal Processng, bomedcal Instrumentaton, and watermarkng. Member of varous Commttees lke ISTE, IETE, etc. V. CNCLUSINS The smulaton results show that the developed system performs better wth the method adopted for feature extracton, and classfcaton of the sgnal. Whle analyzng the result, t has been observed that the system s workng perfectly for the sgnal whch s 30 to 40 second long, but t produces 99% results for the sgnal whch msmatches the desred sgnal length. ne percent devaton s due to the napproprate sgnals whch are ether very nosy of whose sgnal feature could not be extracted at the desred level. The developed system produces a very effectve and accurate decson support system for the novce doctors to cure the patent whch are n urgent need of auscultaton for detecton of pathologcal condton of the heart. It can also be used at the home premse for early detecton of the pathologcal condton of the patent by any person wth some knowledge of medcal background whenever t felt to be so necessary. REFERENCES [1] Abhshek Msal, G.R. Snha, R.M. Potdar, M.K. Kowar Comparson of Wavelet Transforms for Denosng and Analyss of PCG Sgnal, All Rghts Reserved 014 IJARECE 1038