Successive iteration method for reconstruction of missing data


Published on

Published in: Technology, Education
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Successive iteration method for reconstruction of missing data

  1. 1. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME437SUCCESSIVE ITERATION METHOD FOR RECONSTRUCTION OFMISSING DATAT.JayalakshmiCMS College of Science and Commerce, Bharathiar University, IndiaABSTRACTThe information age has made a large amount of data available for medicalinformation processing. Occasional failures lead to missing data. The missing data may makeit difficult to apply analytical models. Data imputation techniques help us fill the missing datawith a reasonable prediction of what the missing values would have been. The implementedsystem creates successive iteration method and various other missing value techniques. Thesystem achieves better performance than other methods.Keywords: Artificial Neural Networks; Back Propagation Method; Diabetes Mellitus;Missing Value Analysis, Pre Processing Methods; Successive Iteration Method.INTRODUCTIONANN is a type of massively parallel computing architecture based on a brain likebehaviour. They work in a similar way as human brain is functioning. The neuron or nervecell is the basic building block of the brain. Neurons are connected in various ways to formbiological neural networks. They process information in the brain, and communicateinformation from different parts of the body to the brain. In total, the human brain has 1014to1015neurons [11]. Neurons receives inputs process them by a simple connections andthreshold operations and outputs a result. Neural networks are sophisticated modelingtechniques capable of modeling extremely complex functions. They are inspired by biologicalmodels of neurological systems and is an established machine learning model with robustlearning properties and simple deployment. Nowadays, they are being successfully appliedacross a wide range of problem domains, in areas like finance, medicine, engineering, patternrecognition and image processing. Anywhere, that there are problem of prediction orclassification neural networks are being introduced.The principle advantage of neural network is to generalize, adapting to signaldistortion and noise without loss of robustness. The back propagation algorithm is a veryINTERNATIONAL JOURNAL OF COMPUTER ENGINEERING& TECHNOLOGY (IJCET)ISSN 0976 – 6367(Print)ISSN 0976 – 6375(Online)Volume 4, Issue 2, March – April (2013), pp. 437-447© IAEME: Impact Factor (2013): 6.1302 (Calculated by GISI)www.jifactor.comIJCET© I A E M E
  2. 2. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME438popular training algorithm in neural network research. Previous studies show that thearchitecture of network can have significant effect on its performance. The performance ofthe network also depends on the data given to the system. Real time processing applicationsthat are highly dependent on data often suffer from the problem of missing input variables.Data incompleteness and estimation are important problems that have received relativelylittle attention in the medical informatics research community. Most of the real world data areseldom complete [20]. The problem of missing data poses difficulty in the analysis anddecision-making process [8]. Decision-making is highly depending on these data, requiringmethods of estimation that are accurate and efficient. Neural networks cannot interpretmissing values; it requires complete set of data for improving the performance [4]. There aremany situations where input feature vectors are incomplete and methods to tackle theproblem have been studied for a long time. A commonly used procedure is to replace eachmissing variable value with an estimated value or imputation obtained from the non-missingvalues of other variables in the same unit. Some intelligent techniques are needed to replacethe missing data.Diabetes is a disease that is characterized by an elevated blood glucose level. This canbe caused by a reduction of the production of insulin by the pancreas (Type I diabetes) or bythe insulin being less effective at moving glucose out of the blood stream and into cells thatneed it (Type II diabetes). The blood glucose level that is elevated for a long period can resultin metabolic complications such as kidney failure, blindness, and an increased chance ofheart attacks. To prevent or postpone such complications strict control over the diabetic bloodglucose level is needed [7]. The aim of this study is to classify the Type II diabetes of PimaIndian Diabetes data set. It is the most challenging problem in machine learning because ofthe high noise level [15][16].To remove the noise and achieve the efficient classification twoimportant techniques were introduced: missing value analysis techniques and preprocessingtechniques.This paper presents an alternative to the usual iterative method for determining theapproximate value. The method developed is based on the standard numerical analysistechnique of successive approximations or iterations. Reconstruction of missing values mayshift input patterns, and the network may have to settle on a complicated solution to satisfyall reconstructed patterns. Successive approximations give correct results comparing to othermethods. This paper organized as follows: Section II briefs about the background study,Section III describes about Artificial neural networks, Section IV deals about diabetesmellitus, Section V gives the methodology of the proposed system, and Section VI concludesthe paper.BACKGROUND STUDYSiti Farhanah Bt Jaafar and Darmawati Mohd Ali [26] proposes a network with eight inputsand four inputs and the results obtained are compared in terms of error. The highestperformance is obtained when the network consists of eight inputs with three hidden layerswith 15, 14, 14, 1 neurons respectively. Amit Gupta and Monica Lam [2] investigate thegeneralization power of a modified back propagation training algorithm such as weightdecay. The effect of the weight decay method with missing values can be applied for differentdata sets such as EPS data set, ECHO data set, IRIS data set etc. The missing values can bereconstructed using standard back propagation, iterative multiple regression, replace byaverage and replace by zero. The weight decay method achieves significant improvement
  3. 3. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME439over the standard back propagation method. Fulufhelo V.Nelwamondo, Shakir Mohamed andTshilidzi Marwala [8] discuss the expectation maximization algorithm and the autoassociative neural network and genetic algorithm combination. The results show that forsome variables EM algorithm is able to produce better accuracy while for the other variablesthe neural network and GA system is better. The findings shows that the methods used arehighly problem dependent. Colleen M.Ennett, Monique Frize, C.Robin Walker [4]investigates the impact of ANN performance when predicting neonatal mortality ofincreasing the number of cases with missing values in the data sets. They proposes threeapproaches such as delete all the values, Replace with mean, and Replace with normal topredict Canadian neonatal Intensive care unit network’s database. The experimental resultswere very promising. Junita Mohamad-saleh and Brain [ 12] proposes the PrincipleComponent Analysis method for elimination of correlated information in data. It has beenapplied to the Electrical Capacitance Tomography data, which contains highly correlated dueto overlapping sensing areas. It can boost the generalization capability of a MLP, the PCAtechnique also reduces the network training time due to the reduction in the input spacedimensionality. The findings suggest that PCA data processing method useful for improvingthe performance of MLP systems, particularly in solving complex problems. Pasi Luukka[21] presented a new approach using a similarity based Yu’s norms for the detection oferythemato squamous diseases, diabetes, breast cancer, lung cancer and lymphography. Thedomain contains records of patients with known diagnosis. The results are very promisingwhen using Yu’s norms for the diagnosis of patients taking into consideration the error rate.The use of this preprocessing method enhanced the result over 30%. Stavros J.Perantonis andVassilis Virvilis [27] proposed a method for constructing salient features from a set offeatures that are given as input to a feed forward neural network used for supervised learning.The method exhibits some similarity to principle component analysis, but also takes intoaccount supervised character of the learning task. It provides a significant increase ingeneralization ability with considerable reduction in the number of required input features.ARTIFICIAL NEURAL NETWORKSA Neural Network is a massively parallel distributed processor made up of simpleprocessing units, which has a natural propensity for storing experiential knowledge andmaking it available for use. It is very sophisticated modeling technique capable of modelingextremely complex functions. ANNs attempt to create machines that work in a similar way tothe human brain by building them using components that behave like biological neurons.However the operation of artificial neural networks and artificial neuron is far moresimplified that the operation of the human brain. The brain consists of millions of theseneurons, which may be specialized in some task or not. The behaviour of the brain inspired todevise an artificial neuron called perceptron, which is the basis of all neural network models.It resembles the brain in two respects a) Knowledge is acquired by the networks from theenvironment through a learning process. b) Interneuron connection strengths, known assynaptic weights, are used to store the acquired knowledge. Neural networks have advantagesover classical statistical approaches especially when the training set size is small comparedwith the dimensionality of the problem to be solved and the underlying data distribution isunknown. Learning is essential to most of the neural network models. Learning can besupervised, when the network is provided with the correct answer for the output duringtraining, or unsupervised, when no external teacher is present. E is the error calculated from
  4. 4. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME440the actual output (d) to the calculated output (o). The goal of the neural network learning is toiteratively adjust weights, in order to globally minimize a measure of the difference betweenthe actual output of the network and the desired output as specified by the teacher.DIABETES MELLITUSDiabetes mellitus is the most common endocrine disease. The disease is characterizedby metabolic abnormalities and by long-term complications involving the eyes, kidneys,nerves, and blood vessels. The diagnosis of symptomatic diabetes is not difficult. When apatient presents with signs and symptoms attributable to an osmotic diuresis and is found tohave hyperglycemia essentially all physicians agree that diabetes is present. The two majortypes of diabetes are Type I diabetes and Type II diabetes. Type I diabetes usually diagnosedin children and young adults, and was previously known as juvenile diabetes [22] [26]. TypeI diabetes mellitus (IDDM) patients do not produce insulin due to the destruction of the betacells of the pancreas. Therefore, therapy consists of the exogenous administration of insulin.Type II diabetes is the most common form of diabetes. Type II diabetes mellitus (NIDDM)patients do produce insulin endogenously but the effect and secretion of this insulin arereduced compared to healthy subjects [6]. Currently cure does not exist for the diabetes, thenonly option is to take care of the health of people affected, maintained their glucose levels inthe blood to the nearest possible normal values [9].METHODOLOGYThe information age has made a large amount of data available for medicalinformation processing. Occasional failures lead to missing data. The missing data may makeit difficult to apply analytical models. Data imputation techniques help us fill the missing datawith a reasonable prediction of what the missing values would have been. The implementedsystem compares various missing value techniques.5.1 CLASSIFICATION DATAThe problem that has been chosen for this research is to classify the Type II diabeticdata using Levenberg Marquardt back propagation algorithm. To investigate the performanceof the proposed neural-based method, the classifier was applied to PIMA INDIANDIABETES dataset and has been collected from UCI machine learning repository [3]. Thedata set is a two-class problem either positive or negative for diabetes disease. The data set isdifferent to classify because of the high noise level. It contains 768 data samples. Eachsample consists of personal data and the results of medical examination. The individualattributes are• Number of times pregnant• Plasma glucose concentration• Diastolic blood pressure(mmHg)• Triceps skin fold thickness(mm)• 2-hour serum insulin(mu U/ms)• Body mass index(weight in kg/(height in m)) 2• Diabetes pedigrees function• Age(years)
  5. 5. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME441The above dataset 268 patients are having diabetes, which can be interpreted as “1” and theremaining patients are not having diabetes and can be interpreted as “0”. The networktopology used for this study is 8-8-8-1. i.e. one input layer, two hidden layers and one outputlayer with eight input nodes, eight hidden nodes and one output node.5.2 NETWORK ARCHITECTUREBack propagation algorithm is the most commonly used algorithm and it is the simplefeed forward artificial neural networks. The algorithm adjusts network weights by errorpropagation from the output to the input. During the training the network minimizes the errorby estimating the weights. The minimization procedure can be performed using the gradient-descent approaches, in which the set of weight vectors consisting of weights is adjusted bythe learning parameters. The network parameters such as learning rate, momentum constant,training error and number of epochs can be considered as 0.9, 0.9, 1e-008 and 100respectively. Before training the weights are initialized to random values. The reason toinitialize weights with small values is to prevent saturation. To evaluate the performance ofthe network the entire sample was randomly divided into training and test sample. The modelis tested using the standard rule of 80/20, where 80% of the samples are used for training and20% is used for testing. In this classification method, training process is considered to besuccessful when the MSE reaches the value 1e-008. On the other hand the training processfails to converge when it reaches the maximum training time before reaching the desiredMSE. The training time of an algorithm is defined as the number of epochs required to meetthe stopping criterion.5.3 MISSING DATA ANALYSISThe problem of missing data poses difficulty in the analysis and decision makingprocesses. Decision making is highly depending on these data, requiring methods ofestimation that are accurate and efficient. Back propagation neural networks have beenapplied for classification problems in real world situations. A drawback of this type of neuralnetwork is that it requires a complete set of input data, and real world data is seldomcomplete. The problem of databases containing missing values is a common one in themedical environment. ANNs cannot interpret missing values, and when a database is highlyskewed, ANNs have difficulty in identifying the factors leading to a rare outcome. Imputationis the substitution of some value for a missing data point or a missing component of a datapoint. Once all missing values have been imputed, the dataset can then be analyzed usingstandard techniques for complete data. The analysis should ideally take into account thatthere is a greater degree of uncertainty than if the imputed values had actually been observed,however, and this generally requires some modification of the standard complete-dataanalysis methods.5.4 METHODSMissing information in data sets is more than a common scenario. There are differentmethods to perform these imputations depending on the type of variable with missing dataand on the type of auxiliary variables. Here the imputation of categorical variables from othernumerical and categorical variables is studied. This paper presents an alternative to the usualmethod for determining the approximate value. It is an alternative to the usual iterativemethod for determining the approximate value. Iterative algorithms usually used whenexplicit formulae are unavailable. The idea is that a repetition of simple calculations will
  6. 6. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME442result in a sequence of approximate values for the quantity of interest. The method developedis based on the standard numerical analysis technique of successive approximations oriterations. Successive approximations give correct results comparing to other methods. Themean of the same attribute is repeatedly calculated until it approximates the value. Theapproximated value is replaced with the missing column.a) Omit the Values: The easiest way to deal with missing values is simply delete all thecases with missing values for the variable under consideration. This technique however maylead to the loss of potentially valuable information about patients whose values are missing.b) Replace with Mean: The second approach is to replace all missing values with the mean.The method of replacing by average is to replace all missing values of an attribute by theaverage of all available values of the same attribute in the training set. Replacing the missingvalues with the means might bias the databases towards the sicker ones.c) Replace with Zero: The third technique is to replace all the missing values with zeros.The method of replacing by zero is simply to replace all missing values by zero. Replacingmissing values by zero in this study has the same effect of replacing missing values by thesmallest value of an attribute since data have been converted to be between zero and one. Ifthe values are important for clinical management the assessment of missing values leads topoor classification.d) Replace with K-nearest neighbor: The fourth technique K-nearest neighbor methodreplaces missing values in data with the corresponding value from the nearest-neighborcolumn. The nearest-neighbor column is the closest column in Euclidean distance. If thecorresponding value from the nearest-neighbor column is also contains missing value the nextnearest column is used.e) Replace with Successive Iteration Method: The technique presented in this paper is toreplace the missing data with successive iteration method.5.4.1 ALGORITHMStep 1: Find the missing element in the data set.Step 2: Calculate the mean of the attributeStep 3: Substitute the mean with missing attributesStep4: Find the new mean of the attributeStep 5: Compare the new mean with the existing meanStep 6: If both are same replace the mean with missing column attributeStep 7: Otherwise repeat the steps 3-6 until the mean converges.5.5 PREPROCESSING OF INPUT DATANeural network training could be made more efficient by performing certainpreprocessing steps on the network inputs and targets. Network input processing functionstransforms inputs into better form for the network use. The normalization process for the rawinputs has great effect on preparing the data to be suitable for the training. Without thisnormalization, training the neural networks would have been very slow. There are manytypes of data normalization. It can be used to scale the data in the same range of values foreach input feature in order to minimize bias within the neural network for one feature toanother. Data normalization can also speed up training time by starting the training processfor each feature within the same scale. It is especially useful for modeling application where
  7. 7. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME443the inputs are generally on widely different scales. Principle Component Analysis is a verypopular normalization method. Principal Component’s normalization is based on the premisethat the salient information in a given set of features lies in those features that have the largestvariance. This means that for a given set of data, the features that exhibit the most varianceare the most descriptive for determining differences between sets of data. This isaccomplished by using eigenvector analysis on either the covariance matrix or correlationmatrix for a set of data.5.6 EXPERIMENTAL RESULTSThe system has been implemented to study the impact of pre-processing and missingvalue techniques. The simulations have been carried out using MATLAB. Various networkswere developed and tested with random initial weights. The network is trained ten times andthe performance goal is achieved at different epochs.Figure. 1. SIT Method without pre-processingFigure. 2. SIT Method with pre-processingFigure.3. Accuracy (without pre-processing)
  8. 8. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME444Figure.4. Accuracy (With pre-processing)TABLE I. IMPACT OF MISSING VALUES WITHOUT PRE-PROCESSINGAccuracy6264666870727476Omit theValuesReplace withZeroReplace withMeanReplace withKNNReplace withSITMethodAccuracyPerformance020406080100120Omit theValuesReplace withZeroReplace withMeanReplace withKNNReplace withSITMethodEpochsAccuracyEpochsPerformance 1e-009Method Accuracy EpochsPerformance1e-009Omit the Values 97.1 24 5.76Replace with Zero 99 15 5.11Replace with Mean 99.2 14 6.48Replace with KNN 99 35 8.33Replace with SIT 99.5 14 5.88
  9. 9. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME445TABLE II. IMPACT OF MISSING VALUES WITH PRE-PROCESSINGThe impact of the missing values can be assessed by taking the average of ten runs andmeasured in terms of classification accuracy (Table.1) and training time (Table.2). It showsthat the accuracy was improved when replace with successive iteration method incombination of PCA pre-processing method.6 CONCLUSIONThis paper demonstrates the impact of missing value technique and pre-processingtechnique. It proves that some combination of missing values and pre-processing theaccuracy was tremendously improved. This shows that pre-processing and missing valuesplay a major role in classification. In future this can be applied for different training methodsand networksREFERENCES[1] Ambrosiadou, V., Gogou, G., Pappas, C., and Maglaveras, N., “Decision support forinsulin regime prescription based on a neural network approach”, MedicalInformatics, 1996, pp.23–34.[2] Amit Gupts and Monica Lam (1998), The Weight Decay backpropagation forgeneralizations with missing values, Annals of Operations Research, Sciencepublishers, pp.165-187.[3] G. Arulampalam and A Bouzerdoum, “Application of shunting Inhibitory ArtificialNeural Networks to Medical Diagnosis”, Seventh Australian and New ZealandIntelligent Information Systems Conference,18-21 November 2001, pp.89 – 94.[4] Colleen M Ennett, Monique Frize, C.Robin Walker, “Influence of Missing Values onAritificial Neural Network Performance”, Proceedings of Medinfo, 2001, pp.449-453.[5] DeClaris, N., and Su, M. C., “A neural network based approach to knowledgeacquisition and expert systems”, IEEE Systems Man and Cybernetics Proc. Pp.645–650, 1991.[6] Edgar Teufel1, Marco Kletting1, Werner G.Teich2, Hans-Jorg Pfleiderer1, andCristina Tarin-Sauer3, “Modelling the Glucose Metabolism with BackpropagationThrough Time Trained Elman Nets”, IEEE 13th Workshop on Neural Networks forSignal Processing, NNSP03, 17-19 Sept. 2003, pp.789 - 798[7] Eng Khaled Eskaf, Prof.Dr.Osama ,Badawi , Prof.Dr.Tim Ritchings, “Predictingblood Glucose Levels in Diabetics using feature Extraction and Artificial NeuralNetworks”.Method AccuracyOmit the Values 72.41Replace with Zero 67.84Replace with Mean 66.53Replace with KNN 70.99Replace with SIT 75.29
  10. 10. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME446[8] Fuluf helo V Nelwamondo, Shakir Mohammed and Tshilidzi Mawala, “Missing Data:A comparison of neural network and expectation maximization techniques”, CurrentScience, Vol 93, No 11, 2007.[9] Humberto M.Fonseca; Victor H.Ortiz, Agustin LCabrera., “Stochastic NeuralNetworks Applied to Dynamic Glucose Model for Diabetic Patients”, 1st ICEEE,2004, pp.522 - 525[10] Igor Aizenberg, Claudio Moraga, “Multilayer Feedforward Neural Network Based onMulti-Valued neurons (MLMVN) and a back propagation learning algorithm”, SoftComputing, 20 April 2006, pp.169-183.[11] Jain.A.K, J.Mao and K.M.Mohuddin (1996), Artificial Neural Networks: A Tutorial,IEEE Computer, Vol.29,No.3,PP.31-44.[12] Junita Mohammad-saleh, Brain S.Hoyle, “Improved Neural Network performanceusing Principle Component Analysis on Matlab”, Int. journal of the computer, theinternet and management, Vol.16, No 2, 2008, pp.1-8.[13] Lakatos,G., Carson, E. R., and Benyo, Z., “Artificial neural network approach todiabetic management”, Proceedings of the Annual International Conference of theIEEE, EMBS, pp.1010–1011, 1992.[14] Maglaveras, N., Stamkopoulos, T., Pappas, C., and Strintzis M., “An adaptive back-propagation neural network for real-time ischemia episodes detection. Developmentand performance analysis using the European ST-T database”, IEEE Trans. Biomed.Engng. pp.805–813, 1998.[15] Md.Monirul Islam, Md.Shahjahan, and K.Murase, “Exploring ConstructiveAlgorithms with Stopping Criteria to Produce Accurate and Diverse IndividualNeural Networks in an Ensemble”, IEEE International Conference on Systems, Man,and Cybernetics, Volume 3, 7-10 Oct. 2001 pp.1526 – 1531.[16] Md.Shahjahan1,M.A.H.Akhand2, and K.Murase1, “A Pruning Algorithm for TrainingNeural Network Ensembles”, SICE 2003 Annual Conference ,Volume 1, 2003,pp.628 – 633.[17] Miller,A. S., Blott,B. H., and Hames, T. K., “Review of neural network applicationsin medical imaging and signal processing” Med. Biol. Eng. Comput. 1992, pp.449–464.[18] M.Nawi1, R.S. Ransing and M.R. Ransing, “An Improved Conjugate Gradient Basedlearning Algorithm for Back Propagation Neural Networks” International journal ofComputational Intelligence, 2008.[19] Ping Zuo,Yingchun Li, Jie Ma SiLiang Ma, “Analysis of Noninvasive Measurementof Human Blood Glucose with ANN-NIR Spectroscopy”, International Conferenceon Neural Networks and Brain, ICNN&B 05. Volume 3, 13-15 Oct. 2005, pp. 1350 –1353.[20] P.K.Sharpe and R.J.Solly, “Dealing with Missing Values in Neural Network basedDiagnostic Systems”, Neural Computing and Applications, Springer-Verlag LondonLtd, 1995, pp.73-77.[21] Pasi Luuka (2007), Similarity classifier using similarity measure derived form Yu’snorms in classification of medical data sets, Elsevier, Computer in Biology andMedicine, pp.1133-1140.[22] Rajeeb Dey and Vaibhav Bajpai, Gagan Gandhi and Barnali Dey, “Application ofArtificial Neural Network (ANN) technique for Diagnosing Diabetes Mellitus”, IEEERegion 10 Colloquium and the 3rd ICIIS, Dec 2008, PID 155.
  11. 11. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 2, March – April (2013), © IAEME447[23] Richard B.North, J.Paul Mcnamee, Lee wu, and Steven Paintadosi, “Artificial neuralnetworks: Application to electrical stimulation of the human nervous system”, IEEE1996.[24] Roelof K Brouwer PEng, PhD, “An Integer Recurrent Artificial Neural Network forClassifying Feature Vectors”. ESSANN, Proceedings – Eur. Symp on ArtificialNeural Networks, April 1999, D-Facto public, pp. 307-312.[25] S.G.Mougiakakou,K.Prountzou,K.S.Nikita , “A Real Time Simulation Model ofGlucose-Insulin Metabolism for Type 1 Diabetes Patients” , 27th Annual Int. Conf.of the Engineering in Medicine and Biology Society, IEEE-EMBS 17-18 Jan. 2006pp.298 - 301.[26] Siti Farhanah, Bt Jafan and Darmawaty Mohd Ali, “Diabetes Mellitus Forecast usingArtificial Neural Networks (ANN)”, Asian Conference on sensors and theinternational conference on new techniques in pharamaceutical and medical researchproceedings (IEEE), Sep 2005, pp. 135-138.[27] Stavros J.Perantonis and Vassilis Virvilis (1999), Input Feature Extraction forMultilayered Perceptrons Using Supervised Principle Component Analysis, Neuralprocessing letters, pp.243-252.[28] T.Waschulzik, W.Brauer, T.Castedello, B.Henery, “Quality Assured EfficientEngineering of Feedforward Neural Networks with Supervised Learning (QUEEN)Evaluated with the pima Indian Diabetes Database”, Proceedings of the IEEE-INNS-ENNS International Joint Conference on Neural Networks, IJCNN, Volume 4, 24-27July 2000 pp.97 – 102.[29] Chaitrali S. Dangare and Dr. Sulabha S. Apte, “A Data Mining Approach forPrediction of Heart Disease using Neural Networks” International Journal ofComputer Engineering & Technology (IJCET), Volume 3, Issue 3, 2012, pp. 30-40,ISSN Print : 0976 – 6367, ISSN Online : 0976 – 6375.[30] D. Kanakaraja, P. Hema And K. Ravindranath, “Comparative Study on Different PinGeometries of Tool Profile in Friction Stir Welding using Artificial NeuralNetworks”, International Journal of Mechanical Engineering & Technology (IJMET),Volume 4, Issue 2, 2012, pp. 245-253, ISSN Print: 0976 – 6340, ISSN Online: 0976 –6359.[31] Y. Angeline Christobel and P. Sivaprakasam, “Improving the Performance of K-Nearest Neighbor Algorithm for the Classification of Diabetes Dataset with MissingValues”, International Journal of Computer Engineering & Technology (IJCET),Volume 3, Issue 3, 2012, pp. 155-167, ISSN Print : 0976 – 6367, ISSN Online : 0976– 6375.