Medical Diagnosis using Incremental Evolution of Neural Network

Transcription

1 Medcal Dagnoss usng Incremental Evoluton of Neural Network Rahul Kala 1, Harsh Vazran 2, Anupam Shukla 3 and Rtu Twar 4 1, 2, 3, 4 Soft Computng and Expert System Laboratory, Indan Insttute of Informaton Technology and Management Gwalor Morena Lnk Road, Gwalor, Madhya Pradesh , INDIA 1 rahulkalatm@yahoo.co.n, 2 harshtmg@gmal.com, 3 dranupamshukla@gmal.com, 4 rt_twr@yahoo.co.n Ctaton: R. Kala, H. Vazran, A. Shukla, R. Twar (2010) Medcal Dagnoss usng Incremental Evoluton of Neural Network, Journal of Hybrd Computng Research 3(1), ABSTRACT: There s an ncreasng concern for the rapd growth of the dseases, whch necesstates the need of automated systems for dsease dagnoss. The automatc dagnoss of dsease not only helps n early detecton of the dsease and takng of necessary preventve measures, t also helps n precse dagnoss that may be of a lot of help to the medcal practtoner. The use of neural networks n dsease detecton s common. The neural networks try to learn from the hstorcal data n the tranng phase. Ths learnng s generalzed n the testng phase to new data. The major problem n neural networks s ther fxed archtecture whch poses a problem for the network desgner to fnd out the optmal archtecture. Ths leads to the use of automatc generaton of neural networks. In ths paper we propose the use of an ncremental evoluton of mult-layer perceptron for the dagnoss of PIMA Indan Dabetes. Here the network starts wth a small number of hdden layer neurons, whch ncrease wth tme. Evolutonary algorthm s used for the optmzaton. The evoluton uses Smulated Annealng as a local search strategy. The resultng dagnoss usng ths strategy showed a better performance as compared to a number of classcal neural approaches. Ths hence makes an excellent dagnostc system. KEYWORDS: Medcal Expert System, PIMA Indan Dabetes, Bo-medcal engneerng, Evolutonary Neural Networks, Smulated Annealng. I. INTRODUCTION Automatc medcal dagnoss s an mportant problem that deals wth the use of technology for the dagnoss of varous dseases. The automated dagnostc systems fnd a lot of applcaton n assstng the doctors for carryng out quck and effectve dagnoss [1]. The dagnostc systems make use of the past data of the patents to carry out the dagnoss. These systems analyze and montor the trends n the hstorcal occurrences or non-occurrences of these dseases. These trends are later generalzed to the new data as well [2]. Ths result n the engneerng of ntellgent expert systems that carry out the task of dagnoss as doman experts based on the nputs provded [3]. The medcal expert systems make use of a hstorcal database that s a collecton of past trends, knowledge base based on whch the decson of the presence of dsease s made, nference engne etc. Dabetes s an emergng problem n the medcal world. The dsease may be classfed nto three types, namely type 1, type 2, and gestatonal dabetes that occur durng pregnancy. Each of these types has ts own symptoms and s found s dfferent populaton dstrbutons [4]. Type 2 dabetes s ncreasng throughout border area, along wth rsk factors for the dsease. Some 1.1 mllon border resdents 18 and older suffer from type 2 dabetes, and 836,000 are pre-dabetc as per 2007 PAHO survey [5]. It was estmated that more than 2.5 mllon people worldwde experence vson loss due to dabetc retnopathy [6]. Accordng to the survey by Amercan Dabetes Assocaton (ADA), 49 percent of the U.S. adults polled sad they most feared cancer as a potental health problem whle just 3 percent sad they worred about dabetes. The survey results suggest that people need to assess ther dabetes rsk and take t more serously [7]. The feld of bomedcal-engneerng reles upon the use of machne learnng for the decson makng. Machne learnng deals wth the analyss of the hstorcal database havng nstances of occurrences and non-occurrences of the dsease. Based on ths database we try to analyze the trends n the hstorcal database. These trends consttute the knowledge base of the system. A good learnng technque results n the extracton of good rules that may easly be generalzed to the testng data as well. The system hence gves the correct output to any appled nput, provded t obeys the trends of

2 the hstorcal database. The problem of medcal dagnoss s essentally classfcatory n nature. The two classes of the system denote the presence and absence of the dsease [8, 9]. Any system s supposed to classfy the nputs nto ether of the two classes. The classfcatory problems behave best when the nter-class separaton s hgh and the ntra-class separaton s low. In such a case t s reasonably easy for the system to construct decson boundares that separate the classes. However for many problems the attrbutes are not deal enough to showcase an easy separaton of the classes by decson boundares. Ths emphaszes on the need of good machne learnng technque to construct flexble decson boundares. The effect of nose especally adds a problem. The nosy data may many tmes result n befoolng the system and tryng to drve t to the wrong drecton. It further may result n poor generalzaton of the resultng system. Neural networks fnd ample of applcatons n medcal dagnoss. They are easly able to learn from the hstorcal data and generalze ther learnng to the testng data. A number of neural network models have been appled to the dagnoss of numerous knds of dseases. Ths ncludes the curve fttng neural network models lke Mult-Layer Percepton, Radal bass Functon Networks, etc.; the classfcatory models of neural networks lke Self-Organzng Maps, Learnng Vector Quantzaton, etc.; or the recurrent neural network models [10-13]. The common dseases for whch expert systems are engneered nclude skn dseases, heart dseases, thyrod dsorders, etc. All the neural models are able to solve the varous dseases by varyng magntudes. One of the bggest dsadvantages of the neural approaches s that the network archtecture needs to be fxed by a human expert. The varous tranng parameters also need to be specfed. The performance of the network largely depends upon these parameters. The human fxng of the parameters may many tmes result n sub-optmal parameter fxng. Further the tranng algorthm may result n the network convergng to local mnma. These problems are solved by the evolutonary neural networks that use the evolutonary algorthms to evolve the most optmal neural network. The evoluton of neural network s a very tme consumng actvty. Ths may hence requre the assstance of a heurstc technque to act as a local search strategy n the evoluton process. Ths results n every ndvdual readly gettng to the most optmal pont n the vcnty. In ths paper we propose an ncremental evoluton of the neural network. The algorthm uses Mult-Layer Perceptron as the neural network model. An evolutonary algorthm facltates the neural network tranng. The maxmum number of permssble neurons n the neural network ncrease along wth evolutonary algorthm generatons. The algorthm uses smulated annealng as the local search strategy. Ths paper s organzed as follows. Secton 2 presents some of the related works. The algorthm desgn and workng s presented n secton 3. Secton 4 presents the smulaton results and the comparson of the algorthm to the other conventonally used algorthms. The concluson remarks are presented n secton 5. II. RELATED WORK In ths secton we brefly overvew the advances n the doman of bomedcal expert systems as well as classfcaton. Numerous models on medcal dagnoss have been developed and tested by Shukla et al. [10-13]. These models use a varety of methods namely Mult-Layer Perceptron wth Back Propagaton Algorthm, Radal Bass Functon Networks, Self Organzng Maps, Learnng Vector Quantzaton, Adaptve Neuro Fuzzy Inference Systems, etc. for the dagnoss of dseases. The major dseases nclude dabetes, heart dseases, eplepsy, breast cancer, thyrod, etc. In all combnatons of model and dsease, an effectve dagnoss could be made. Ths emphaszes on a hgh degree of accuraces of the ndvdual systems. There s however always a scope to remove the ndvdual lmtatons of the models and further enhance the recognton score. A number of models from the hybrd soft computng have been appled on the problem of PIMA Indan Dabetes by Kala et al. [14]. Ths ncludes the ensemble approach, neurofuzzy system, and evolutonary neural networks. All the hybrd methods gave a good accuracy for dagnoss. Based on the comparsons n the same work t was clear that the evolutonary neural networks and ensemble technques remove the lmtatons exstng n the ndvdual neural network models. An extended verson of these works may be found n [15, 16]. The problem of classfcaton especally requres a good system modelng n order to enable the system separate the varous classes n the system. The major problem s especally the classfcaton of the nputs that le close to the decson boundares. Kala et al made an mplementaton of a Modular Neural Network for machne learnng [17]. Ths model clustered the entre nput space nto clusters. Each cluster was solved usng ts own neural network. The approach was appled for learnng of a self made database of face recognton. Results proved that the approach could better dentfy the faces. Further the system was scalable to handle much more data. In another approach [18] an ensemble approach s used for problem solvng. Here a varety of models were used for solvng the same set of nputs and outputs. The ntegrator used a votng mechansm for decdng the fnal output. Ths approach was on a speech database. The combnaton of face and speech was appled along wth a better ntegraton technque n [19]. Here each module returned the probabltes of the occurrence of the varous classes. These were summed up for all the modules to get the fnal probablty vector. The ntegrator declared the class correspondng to the maxmum sum as the fnal output class. One good modular neural network model s presented for the bometrc recognton n [20, 21]. Here the authors make three dfferent modules for a mult-modal bometrc recognton system. One module s dedcated to each bometrc modalty.e. face, speech and fngerprnt. Fuzzy ntegraton s the ntegraton technque of use. Each module n turn uses a herarchcal Modular Neural network wth an evolutonary base and a fuzzy ntegraton technque. In another work [22] co-evoluton s used as a mechansm of evoluton of a modular neural network. In ths model the

3 varous modules of the modular neural network evolve n a co-evolutonary approach. The varous modules help each other to evolve wth good recognton rate and develop dstnct characterstcs for an optmal overall system performance. III. ALGORITHM FRAMEWORK The general algorthm s an evolutonary neural network. The neural network model used s Mult-Layer Perceptron. Ths model conssts of a number of neurons arranged n a layered archtecture. The frst layer s a passve layer called as the nput layer. The last layer s the output layer, where the outputs are collected. There may be a number of hdden layers n between. The use of multple hdden layers complcates the problem wth the constructon of rapdly changng decson boundares. The resultng network gves a reasonable performance n the tranng database, but the performance s very poor n the testng database due to low generalzaton [2]. Most of the bo-medcal dagnoss problems can be effectvely solved by a sngle hdden layer. We therefore assume that the neural network conssts of a sngle hdden layer. Further the actvaton functons of the hdden layer and output layer are assumed to be constant that do not change n the evolutonary process. We assume the hdden layer to have an actvaton functon of tansg and the output layer to have an actvaton functon of pureln. Ths s shown n fgure 1. Inputs Weghts Fgure 1 General Archtecture of Mult-Layer Perceptron Neural Network Each neuron here does the weghted sum of the nputs wth addton of bas as gven n equaton (1) and then passes the resultant sum for applcaton of the actvaton functon as gven n equaton (2). Ths becomes the output of the neuron that may be gven to the other neurons as per the network archtecture. y w x b (1) Input Layer o f ( y) f ( w x b) (2) Hdden Layer Output Bas Output Layer The evolutonary approach here does the task of fxng of the number of neurons n the hdden layer as well as the varous weghts and bases. The evolutonary algorthm used here follows an ncremental evoluton approach. An evolutonary approach always creates neural networks of varyng szes. Ths depends upon the evolutonary operators as well as the system specfcatons. In ths algorthm we restrct the maxmum number of neurons n the hdden layer to n max (t). Any network cannot have more than ths number of neurons. Ths crteron s explctly checked nto all the evolutonary operators. Ths number s kept varable and changes as the algorthm proceeds. Intally the maxmum permssble neurons n max (t) s kept to a very low value. As the generatons ncrease, ths number gradually ncreases n a lnear manner. Hence the ntal few generatons wtness very small networks. As the generatons ncrease, the networks start gettng complex n nature. The ncrease n n max (t) s gven by equaton (3). n max nmax ( G) nmax (0) ( t) nmax (0) t (3) G Here G s the maxmum possble generatons. The complete evolutonary algorthm s shown n fgure 2. The algorthm frst generates random neural networks. It s ensured that each of the ntal neural networks have neurons less than n max (0). Ths becomes the ntal populaton for the evoluton. The evoluton takes place untl the maxmum generatons are not reached. At each generaton the frst task s to generate the populaton of the next generaton. Ths s done by the applcaton of the evolutonary operators on the prevous generaton populaton. The varous evolutonary operators are dscussed n the next sub-sectons. The factor n max (t) ncreases at every generaton and needs to be calculated. Smulated Annealng serves as a local search strategy for the ndvdual to hunt for the best place n the vcnty. Ths greatly relves the complexty of the evolutonary process. Each ndvdual of the evolutonary process s a mult-layer perceptron. The conventonal smulaton s done for all the tems of the tranng data. Ths determnes the performance of the neural network or the ftness of the ndvdual. The larger networks are gven a penalty. The varous steps are dscussed one by one. A. Evolutonary Operators The task of generaton of next generaton populaton from the prevous generaton s done by the applcaton of the evolutonary operators. The varous operators used n ths algorthm nclude () selecton, () mutaton, () crossover, (v) elte, (v) jump and (v) new. Selecton operator does the task of selecton of the ndvdual for the evolutonary process. The selecton s more lkely to select the ftter ndvduals than the weaker ndvduals. The algorthm uses stochastc unform selecton mechansm. The selecton uses a rank based scalng mechansm. Ths scalng saves the algorthm from beng over

4 domnated by ftter ndvduals. Mutaton operator tres to add new characterstcs nto the ndvdual. Ths operator adds new characterstcs to the system by slght modfcaton of the exstng ndvduals. The algorthm uses a Gaussan mutaton technque. All the weghts and bases are modfed randomly by an amount that depends upon Gaussan random number. No Intal Neural Networks Stoppng Crteron Met Compute n max (t) Evolutonary Operators Ftness Evaluaton Yes Testng Neural Network Performance Computaton Smulated Annealng Fgure 2 General Structure of the Algorthm The crossover operaton mxes two ndvduals or neural networks to produce the chldren neural networks. The problem n ths problem wth the crossover s of varable szes of the two networks. Let the two parents have szes n 1 and n 2. The crossover operator creates chldren that have szes cel((n 1 +n 2 )/2) and floor((n 1 +n 2 )/2). The dstrbuton of the weghts and bases from the parents to the chldren takes place n a random manner where any of the weghts or bas to any chld may randomly come from ether of the parents. Ths s the scattered crossover technque. The other evolutonary operator s elte. Ths operator smply passes the best ndvdual of one generaton drectly nto the hgher generaton. The next evolutonary operator used s jump. Ths operator modfes the structure of the neural network ndvdual. It adds a new hdden layer neuron to the ndvdual. The correspondng weghts and bas are gven random values. Smlar operator s the new evolutonary operator. Ths operator adds completely random ndvduals wth the maxmum allowable sze n max (t) to the populaton pool. The phlosophy of ncremental evoluton s used here where the ndvduals are expected to grow nto complex shapes wth tme. These operators hence try to advance neural networks to attan larger shapes along wth tme. It may be noted that the smaller neural networks are preferable as they lead to better generalzaton. Smaller networks hence have greater chance of survval n the evolutonary process by beng selected n the elte, selecton, and crossover operatons. Ths makes the evolutonary process bas towards smaller networks and the smaller networks start domnatng the populaton pool. It s equally mportant to same larger networks from extncton as well as to add an exploratory nature to the algorthm. The addton of neurons may have a large mpact on the performance. Addton of a few neurons may result n better learnng and hence better recognton. Ths exploraton for hgher performance by larger networks s caused by these two operators. B. Smulated Annealng The neural networks may have a varety of archtecture. Each of these archtectures s a collecton of a large number of weghts and bases. Ths makes the search space for the evolutonary algorthm very large n sze. Besdes the search space has a very complex archtecture. Evolutonary algorthms face problems n such complex ftness landscapes. The evolutonary process hence needs to be asssted by a local search strategy. The local search strategy helps every ndvdual to reach the most optmal pont n ts vcnty. Ths ads the evolutonary algorthm whose job now s prmarly to place ndvduals near the mnma. In ths algorthm smulated annealng s used as a local search strategy. We run a few teratons of ths evolutonary process. The smulated annealng modfes all the parameters by some small amount. If the resultant ndvdual has a better ftness, t s accepted. If t has a lower ftness t s accepted wth a probablty gven n equaton (4). f p e T (4) Here T s the temperature constant. δf s the change n ftness. C. Neural Network Any ndvdual of the evolutonary approach represents a neural network. The ftness of the ndvdual s the performance of the neural network n solvng the dagnoss problem. The tranng database s used for the measurement of the performance. The tranng targets consst of an entry of 1 denotng the presence of the dsease and 0 denotng ts absence. The total error between the outputs and the targets s measured and summed for all the tems n the tranng database. Ths s gven n equaton (5). o t N E (5) Here N s the total number of elements n the tranng database. The algorthm puts a penalty on the networks wth more number of neurons. Ths saves the algorthm from producng very large networks that may have a poor generalzaton. The penalty s proportonal to the total number of neurons. The

5 net ftness of the ndvdual that needs to be mnmzed by the evolutonary algorthm may hence be gven by equaton (6) Ft = E + α H (6) Here α s the penalty constant and H s the total number of neurons n the neural network. IV. RESULTS The algorthm was mplemented n JAVA. The data module of the entre program was used to feed data nto the system. Smlarly there were modules for the neural network as well as the evolutonary algorthm. The smulated annealng module was kept as a part of the neural network module. The ftness functon of the evolutonary algorthm executed the smulated annealng module whch n turn used neural network module for the evaluaton. The am of the system s to solve the problem of detecton of PIMA Indan Dabetes. For ths we make use of the database of UCI Machne Learnng Repostory [23]. The PIMA Indan Dabetes data set conssts of a total of 8 attrbutes. These decde the presence of dabetes n a person. Ths database places several constrants on the selecton of these nstances from a larger database. In partcular, all patents here are females at least 21 years old of Pma Indan hertage. The frst attrbute s the number of tmes the women was pregnant. The next attrbute s Plasma glucose concentraton a 2 hours n an oral glucose tolerance test. We further have the attrbutes Dastolc blood pressure (mm Hg), Trceps skn fold thckness (mm), 2-Hour serum nsuln (mu U/ml), Body mass ndex (weght n kg/(heght n m)^2), Dabetes pedgree functon and Age (years). The entre data was dvded nto tranng and testng data sets. The tranng dataset conssted of about 70% of the data randomly chosen from the dataset. The remanng 30% data was used as the testng dataset. The tranng data set was used for the evoluton of the neural network usng the dscussed approach. The best ndvdual after the entre evoluton was the most optmal neural network. Ths network was then executed aganst the tranng and the testng data sets and the correspondng performance was noted. The methodology s shown n fgure 3. The smulaton of the algorthm was done for 50 generatons wth 50 ndvduals. The number of neurons could vary from 1 to 25. At ant generaton 40% of the ndvduals were contrbuted by the applcaton of crossover operator, 20% by the applcaton of mutaton operator, 25% by the operaton of new operator and 15% by the applcaton of the jump operator. The elte count was kept to 1. The penalty constant was kept as 0.1 and the smulated annealng temperature constant had a value of 2. The smulaton under these parameter values took approxmately 2 mnutes. At the end of the evolutonary process, the neural network so obtaned had an accuracy of dagnoss of on the tranng data and 82.38% on the testng data. The plot of the best ndvdual n varous generatons s shown n fgure 4. Tranng Data Set System Intalzaton System Tranng Testng on Tranng Data Set Entre Data Set Testng Data Set Testng on Tranng Data Set Fgure 3 The workng methodology Fgure 4 Plot of the ftness of best ndvdual v/s generatons We further appled a few commonly used methods to the same data set to compare the effcency of the proposed algorthm aganst these methods. The frst method appled was of Artfcal neural Network (ANN) traned wth Back Propagaton Algorthm. Here we used a sngle hdden layer whch conssted of 12 neurons. The actvaton functons for the hdden layer was tansg and pureln. The tranng functon used was trangd. The other parameters were a learnng rate of 0.05 and a goal of Tranng was done tll 2000 epochs. After the network was traned and tested, the performance of the system was found out to be % for the tranng data set and % for the testng data set. The second method appled was on ensembles. Here we had used 4 modules or ANNs. Each one of them was traned separately usng the same tranng data set. The 4 ANNs were more or less smlar to each other wth small changes. These had 12, 14, 10 and 12 neurons respectvely. The numbers of epochs were 2500, 200 and The four ANNs were traned separately. Here we had used a probablstc pollng n place of the normal polng. The resultng system had a total performance of % for the tranng data and % for the testng data. The thrd method appled was of ANFIS. Here we used the same tranng as well as testng data sets. The FIS was generated usng a grd parttonng method. Each of the

6 attrbutes had 2 MFs wth t. The system was allowed to be traned for a total of 100 epochs. The fnal system so obtaned had a performance of % for the tranng data and % for the testng data. The fourth system was connectonst archtecture of evolutonary ANN. Here the chromosome stored the presence or absence of connectons n between neurons (along wth weghts) to elmnate a fully connected archtecture. The parameters of the GA were a maxmum number of 25 neurons, 25 as the populaton sze wth an elte count of 2. The creaton functon was unform and double vector representaton was chosen. Rank based ftness selecton was used. Stochastc Unform selecton method was used. Crossover rato was 0.8. The algorthm was run for 75 generatons. The fnal system had a performance of 77.38% for the tranng data set and % n the testng data set. The last system appled was the conventonal RBFN. Here the neurons had a spread of 55. The system so generated had a performance of 79.25% on the tranng data and 78.41% on the testng data. The performance of the varous models s analyzed n table 1. We can easly see that the proposed system gave the best performance than all the commonly known methods. Ths system hence presents an effectve system for dagnoss of PIMA Indan dabetes. Table 1. Comparson of results from varous algorthms. Method Tranng Accuracy Testng Accuracy Proposed Algorthm 82.96% 82.38% ANN wth BPA 77.33% 77.73% Ensemble (wth BPA) 78.72% 76.98% ANFIS 88.97% 66.52% Evolutonary ANN 77.38% 73.81% RBFN 79.25% 78.41% V. CONCLUSIONS In ths paper we studed the problem of dagnoss of PIMA Indan dabetes. The problem was solved usng an ncremental evoluton of neural network. Here the maxmum permssble number of neurons n the hdden layer ncreased along wth generatons. As the algorthm proceeded, the neural network started attanng optmzed archtecture as well as weghts and bases. The entre evoluton was done usng a varety of evolutonary operators. The evolutonary process was asssted by smulated annealng technque. Ths served as a local search strategy for the search of the global mnma. The algorthm was able to acheve a hgh degree of accuracy for both tranng as well as testng data sets. The performance of the algorthm was compared wth a varety of technques. Ths ncluded the use of conventonal Mult-Layer Perceptron wth Back Propagaton algorthm for tranng, modular neural networks, connectonst evoluton of neural network, radal bass functon network and adaptve neuro fuzzy nference system. The proposed algorthm performed better as compared to all these networks. Ths shows that the algorthm may be effectvely used for the medcal dagnoss. The proposed algorthm was tested agest a sngle data set of dabetes. The experment may be repeated for more data sets to ensure the generalzaton of the observatons. The algorthm may further be asssted by more ntellgent heurstc or soft computng technques to serve as the local search strategy. The complete search space of the neural network evoluton s hghly complex. Better technques to adaptvely explore ths search space as well as to strke a good balance between exploraton and explotaton may be devsed. All ths should result n escapng from the local mnma and tmely convergence to the global mnma. All ths may be carred n future. REFERENCES [1] J. D. Bronzno, Bomedcal Engneerng Fundamentals, CRC Press, 2006 [2] A. Shukla, R. Twar, R. Kala, Real Lfe Applcatons of Soft Computng, CRC Press, 2010 [3] N. K. Kasabov, Foundaton s of Neural Networks, Fuzzy Systems, and Knowledge Engneerng, MIT Press, 1998 [4] Western Sydney Endocrne Centre. Avalable va content&task=vew&d=6&itemd=29, 2004 [5] Pan Amercan Health Organzaton/World Health Organzaton, Dabetes Increasng along U.S.-Mexco Border. Avalable va betes/surveyresults/tabd/318/language/en- US/Default.aspx, 2006 [6] Brussels, Internatonal Dabetes Federaton s World Dabetes Day 2002 Focuses On Eye Complcatons. WDD Avalable va [7] Kathleen Doheny, Too Few Understand Dabetes' Dangers. HealthDay Reporter.. Avalable va 10/28/too-few-understand-dabetes-dangers.html, 2008 [8] R. Kala, A. Shukla, R. Twar, A Novel Approach to Classfcatory problem usng Neuro-Fuzzy Archtecture, Internatonal Journal of Systems, Control and Communcatons (IJSCC), Inderscence Publshers, 2010 [9] R. Kala, A. Shukla, R. Twar, A Novel Approach to Classfcatory Problem usng Grammatcal Evoluton based Hybrd Algorthm, Internatonal Journal on Futurstc Computer Applcatons, 2010 [10] R. R. Janghel, A. Shukla, R. Twar, Decson Support system for fetal delvery usng Soft Computng Technques, Proceedngs of the Fourth Internatonal Conference on Computer Scences and Convergence Informaton Technology, eeexplore, pp , 2010, Seoul, Korea. [11] R. R. Janghel, A. Shukla, R. Twar, P. Twar, Clncal Decson support system for fetal Delvery usng Artfcal Neural Network, Proceedngs of the 2009 Internatonal Conference on New Trends n Informaton and Servce Scence, NISS 2009, pp , 2009, Gyeongju, Korea [12] A. Shukla, R. Twar, P. Kaur, Intellgent System for the Dagnoss of Eplepsy, Proceedngs of the IEEE World Congress on Computer Scence and Informaton

7 Engneerng (CSIE), eeexplore, pp , 2009, Los Angeles/Anahem, USA [13] A. Shukla, R. Twar, P. Kaur, R. R. Janghel, Dagnoss of Thyrod Dsorders usng Artfcal Neural Networks, Proceedngs of the IEEE Internatonal Advanced Computng Conference, eeexplore, pp , Patala, Inda [14] R. Kala, A. Shukla, R. Twar, Comparatve analyss of ntellgent hybrd systems for detecton of PIMA ndan dabetes, Proceedngs of the IEEE 2009 World Congress on Nature & Bologcally Inspred Computng, NABIC '09, pp , 2009, Combatote, Inda [15] A. Shukla, R. Twar, Intellgent Medcal technologes and Bomedcal Engneerng: Tools and Applcatons, IGI Global Publshers, 2010 [16] A. Shukla, R. Twar, Bomedcal Engneerng and Informaton Systems: Technologes, Tools and Applcatons, IGI Global Publshers, 2010 [17] R. Kala, A. Shukla, R. Twar, Fuzzy Neuro Systems for Machne Learnng for Large Data Sets, Proceedngs of the IEEE Internatonal Advance Computng Conference, eeexplore, pp , 2009, Patala, Inda [18] A. Shukla, R. Twar, H. K. Meena, R. Kala, Speaker Identfcaton usng Wavelet Analyss and Modular Neural Networks, Journal of Acoustc Socety of Inda (JASI), Vol 36, No. 1, pp 14-19, 2009 [19] R. Kala, H. Vazran, A. Shukla, R. Twar, Fuson of Speech and Face by Enhanced Modular Neural Network, Proceedngs of the Internatonal Conference on Informaton Systems, Technology and Management, ICISTM 2010, CCIS 54, pp , 2010, Bankok, Thaland [20] P. Meln, O. Castllo, Hybrd Intellgent Systems for Pattern Recognton Usng Soft Computng, Sprnger, 2005 [21] P. Meln et al, Modular Neural Networks and Fuzzy Sugeno Integral for Face and Fngerprnt Recognton, In Appled Soft Computng Technologes: The Challenge of Complexty, Sprnger, 2006 [22] N. G. Pedrajas, C. H. Martnez, J. M. Perez, Multobjectve cooperatve coevoluton of artfcal neural networks (mult-objectve cooperatve networks), Neural Networks, Vol. 15, 2002, pp [23] Vncent Sgllto, UCI Machne Learnng Repostory [ The Johns Hopkns Unversty. 1990, Avalable At: es