Application of Radial Basis Network Model for HIV/AIDs Regimen Specifications

JOURNAL OF COMPU TING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151- 9617 HTTPS://SITES.GOOGLE.COM/SITE/JOU RNALOFCOMPUTING/ 136 Application of Radial Basis Network Model for HIV/AIDs Regimen Specifications Dr. P.Balasubramanie,Kongu Engineering College, Perundurai. Prof. M.Lill y Florence, Adhiyamaan Colle ge of Engineering, Hosur. Abstr act: HIV/AIDs Regimen specification one of many problems for whi ch bioinformaticians h ave implemented and trained machi ne learning methods su ch as neural networks. Predicting HIV resistan ce would be much ea sier, but unfortunately we rarely ha ve enough structural information avail able to train a neural network. To network model designed to p redict how long the HIV patient can p rolong his/her life time with certain regimen specifi cation. To learn th is model 300 patient’s details have taken as a training set to train the network and 100 patie nts medical history has t aken to test this model. This netwo rk model is trained using MAT lab implem entation. Resources: RBN alg orithm implemented in MAT Lab, data from a ART centre in Tamil Nadu. Key words: RBN, Regimen specif ication and ANN, etc. 1. Introduction The Human Immunodeficiency Virus i s one of the main causes of death in the worl d. The HIV is a human pathogen that infe cts certain types of lymphocytes called T -helper cells, which are important to the immune system. Without a sufficient number of T-helper cells, the immune system is unable to defend the body against infections. Clinic al trail sys tem for HIV/AIDs is a complex one. This is the ca se because every patient is unique with his/her own health, set of genetic traits, predisposition to side effects and prognosis. Additionally, many symptoms a nd diagnoses are inherently imprecise in thei r definition and difficult to measure. Although clinical trial data provide excellent information regarding excepted tre atment outcomes for large groups of patient s, the prediction o f actual treatment outcomes and clinical cou rses for a particular individual patient may be subj ect to a considerable degree of uncertainty. In this paper, we will focus on the function of RBN for our problem and we have discu ssed about some of the outcomes of this trained network. 2. Radial Basis Function Network (RBFN) Model The RBFN model consists of three layers, the input , hidden and output la yer. The nodes within each layer are fully connected to the previous layer as shown in Fig 1. The input variables are each assigned to a node i n the input layer and pass directly to the hidden layer wit hout weights. The hidden nodes or u nits contain the radial basis functions also called transfer functions and are analo gous to the JOURNAL OF COMPU TING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151- 9617 HTTPS://SITES.GOOGLE.COM/SITE/JOU RNALOFCOMPUTING/ 137 sigmoid functions comm only used in the back propagation netwo rk models. They are represented by the bell shape d curve in the hidden model shown in figure1. Figure 1 FBN Model for this network 1. Input layer – There is one neuron i n the input layer for each pre dictor variable. In the case of categori cal variables, N-1 neurons are used where N is the number of categorie s. The input neurons (or processi ng before the input layer) standardize s the range of the values by subtracting the median a nd dividing by the inter quartile range. The input neurons then feed the values to each of the neurons in the hidden layer. 2. Hidden layer – This layer has a variable number of neurons (the optimal number is determined by the training process). Each neuron con sists of a radial basis function centered on a point with as many dimensions as there are predi ctor variables. The spread (radius) of the RBF function may be different for each dimension. The cente rs and spreads are determined by the training process. When presented with the x vector of input values from the input layer, a hidden neuron computes the Euclide an distance of the test case from the neuron’s center point and then appli es the RBF kernel function to this di stance using the spread value s. The resulting value is passed to the summation layer. 3. Summation layer – The value coming out of a neuron in the hidden layer is multiplied by a weight associated with the neuron (W1, W2, ...,Wn in this figure) and passed to the summatio n which adds up the weighted values an d presents this sum as the output of the network. Not sho wn in this figure is a bias value of 1.0 that is multiplied by a weight W0 and fed into the summatio n layer. For classification problem s, there is one output (and a separate set o f weights and summation unit) for ea ch target category. The value output for a category is the probability that the case being evaluated has that category. 3. Training RBF Networks The following parameters are d etermined by the training process: 1. The number of neuro ns in the hidden layer. 2. The coordinates of the center of ea ch hidden-layer RBF function. 3. The radius (spread) of each RBF function in each dimension. 4. The weights applied to the RBF function outputs as they are passed to the summation layer. Various methods have b een used to train RBF networks. One ap proach first uses K- means cluste ring to find cluster cente rs which are then used as the cente rs for the RBF functions. However, K-means cl ustering is a computationally intensive procedure, and it often does not generate the optimal number of centers. Another ap proach is to use a random subset of the training points as the cent ers. The RBF network has a feed forward structure consisting of a single hidd en layer of J locally tuned units, which are fully interconnected to an output layer of L linear units. All hidden units simu ltaneously receive the n -dimensional real valu ed input vector X (Figure 3). The main difference from that of MLP is the absence of hidden-layer weights. The hidden- unit outputs are not calculated using the weighted-sum mechanism/sigmoi d activation; rather each hidden -unit output Zj is obtained by closeness of the input X to an n -dimensi onal parameter vector mj associated with the j th hidden unit10,11. The response ch aracteristics of the j th hidden unit ( j = 1, 2, ¼, J ) is assumed as, Z j =K ( || X - µ j ||) / σ 2 j JOURNAL OF COMPU TING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151- 9617 HTTPS://SITES.GOOGLE.COM/SITE/JOU RNALOFCOMPUTING/ 138 Figure 2. Feed forwa rd neural network Figure 3 Ra dial basis funct ion neural network where K is a strictly positive radically symmetric function (kernel) with a unique maximum at its ‘centre’ mj and which drops off rapidly to zero away from the centre. The parameter sj is the width of the receptive field in the input space from unit j . This implies that Zj has an appreciable value only when the distan ce || || j X - m is smaller than the width sj . Given an input vector X , the output of the RBF network is the L - dimensional activity vector Y , whose l th component ( l = 1, 2 ¼ L ) is given by, l Y l (X) = ∑ w lj z j (X) j=1 For l = 1, mapping of eq. (1) is similar to a polynomial threshold gate. Ho wever, in the RBF network, a choi ce is made to use radi cally symmetric kernels as ‘hidd en units’. 4. Experimental Setup and Result of RBN network Model Consider an observation used to trai n the model to have r input s variables such as patient’s age, weight, CD count, HB rate, CD8 count and so on. T he weight for calculating the sum is regimen specifi cation. The regimen specification and other factors are not common for all the patients. According to the output received the weight is adjusted and again it executed until the required output i s generated. Now the adjus ted weight is the r egimen specification for the particular patient. In this way that the network is trained. In this research, we ha ve taken 500 patients medical history . Among this, 300 cases used as training set and 20 0 cases used as testing set. Among these hom ogeneous set, the output of this model defines the regimen specification for two sets. One set of cases, they can prolong their life more than 10 years only if they should follow the specified regimens, another set, they are very difficult to prolong their life even if they follow the restricted specification of regimens. In this study the maximum age of the patient s taken is 45 and only used homogenou s data. Table 1 shows the details of patie nts (sample) . Table 2 shows the suggested regimen spe cification. The definition of regimen specification i s consulted with a Physician in the famous ART centre in Tamil Nadu. Table 1 Patient ID Age Weight CD4 CD8 HB TLC First Identified Date A 23 45 204 721 10 945 23.11.07 B 23 43 187 1498 10 1769 23.10.07 C 34 42 38 812 12.5 1076 28.07.06 D 35 35 28 940 8 1202 19.12.07 JOURNAL OF COMPU TING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151- 9617 HTTPS://SITES.GOOGLE.COM/SITE/JOU RNALOFCOMPUTING/ 139 E 32 41 238 408 9 1100 Year 2000 F 38 92 33 294 9.2 571 10.12.07 G 37 43 123 1262 8.5 1605 04.09.07 H 40 35 38 745 8 1169 01.12.07 I 42 43 34 512 12.5 811 23.09.98 J 39 40 112 643 10 900 869 In Table 1 there 10 patients data h as been taken, the important factors to define the regimen are given. The underlined data are conserved as set 2, sin ce this case cannot prolong their life for the expected pe riods. The remaining data consider as set 1, they can prolong their life provided they shoul d strictly follow the defined regimen specification. Table 2 Patient ID Regimen R_Specification Prolong Period A ZLN 2 Per Day >75% B ZLN 2 Per Day >75% C ZLE 2 Per Day <50% D SLN 30 2 Per Day <50% E ZLN 2 Per Day >75% F ZLE 2 Per Day <50% G SLN 30 2 Per Day >75% H ZLN 2 Per Day <50% I ZLN 2 Per Day <50% J ZLN 2 Per Day >75% In Table 2 the outcome of this RBN has discussed one set of cases they can prolong their life time more than 75% i.e in our research we have taken maximum 10 years, so more th an 7.5 yrs they can alive if they give the co ntinuous response. The second set of ca ses they can prolong maximum of 50% of the period. There are three combination of regimens woul d be prescribed based on their TLC count for them. The specification is depen ds on their weight, HB and CD4 count. The g rams and mill gra ms of this regimens depends o n their age and weight. 5. Conclusion The sensitivity and specificity of both neu ral network models had a b etter predictive power compared to logistic reg ression. Even when compared on an external dataset, the neural network models performed better than the logistic regressi on. This study indicates the good predictive capabilities of RBF neural network. Also the time taken by RBF i s less than that of MLP in our application. Though application of RBF network is limited in biomedicine, many comparative studies of MLP and statistical methods are used. The limitation of the RBF neural network is that it is more sensitive to dimensionality and has greater difficulties if the number of units is large. Generally, neural network results presented are mostly based only on the same dataset. Here an independent valuation is done u sing external validation data and both the neural netwo rk models performe d well, with the RBF model having better prediction. The p redicting capabilities of RBF neural network had showed good results and more applications would bring out the efficiency of this model over other models. ANN may be particularly useful wh en the primary goal is classification and is important when interactions or complex n onlinearities exists in the dataset. Logistic re gression remains the clear choice when the primary goal of model development is to look for possible causal relationships between independ ent and dependent variables, a nd one wishes to easily understand the effect of predictor variables on the outcome. There have been ingeniou s modifications and restrictions to the neu ral network model to broaden it s range of applications. The bottleneck networks for JOURNAL OF COMPU TING, VOLUME 1, ISSUE 1, DECEMBER 2009, ISSN: 2151- 9617 HTTPS://SITES.GOOGLE.COM/SITE/JOU RNALOFCOMPUTING/ 140 nonlinear principle co mponents and netwo rks with duplicated weights to mimic autoregressive models are rece nt examples. When classification is the goal , the neural network model will often deliver close to the best fit. The present work was a part of our research. 6. References 1. Rumelhart, D. E., Hinton, G. E. and Williams, R. J., Learning representation by back-propagat ing error s. Nature , 1986, 323, 533–536. 2. Hecht-Nielsen, R., Neurocomputing , Addison-Wesley, Reading, MA, 1 990. 3. White, H., rtificial Neural Net works. Approxim ation and Learning The ory , Blackwell, Cam bridge, MA, 1992. 3. White, H. and Gallant, A . R., On learning the derivatives of an unkn own mappin g with multilayer feedforward networks. Neural Networks ., 1992, 5, 129 –138. 4. GROSS, G., and GALIANA, F.D.: ’Short- term load forecasting’, 2 GUPTA, P.C.: ‘A stochastic ap proach to pak demand forecasting in electric utility systems’, IEEE Trans., 1971, PkS-W, (2), pp. 824- 832 exponential sm oothing ’, IEEE Trans., 1971, PAS%, (2). pp. 900- 911 5. HUBELE, N., and CHENG, C.: ‘Identilication of aasonal shortterm load forecasting models usi ng stati stical decision functions’, hocIE. EE, 1 98 7,p p. i ss8-1~73 6. CHRISTAANSE, W.R.: ‘Shor t-tm load fomasbhg Usi ng general I EEE Trans., 1990, PWRSS, (I)p, p40- 45 7. Carpenter G .A. and Gro ssberg S. (1995). “Adaptive Re sonance Theo ry (ART)”. I n: Handbook of Brain Theor y and Neur al Networks Ed.: Arbib M. A. , MIT P ress, 1995 8. Carpente r G.A., Grossb erg S. and Rosen , D.B. (1991). “Fuzzy ar t: Fast stable learning and categorization of a nalog patterns by an adaptive resona nce system”. Neural Networks , Vol. 4, pp. 759-77 1, 1991. 9. Grossberg, S. (1976). “Ad aptive pattern classification and uni versal recordin g : I. Parallel devel opment and c oding of ne ural feature detectors”, Biologi cal Cyber netics , Vol 23, pp 121- 134. 10. Zhang, Y.-J., Liu, Z.-Q. (2002). “Self- Splitting Competitive Lea rning: a New Online Cl ustering Pa radigm ”, IEEE Transactions On Ne ural Networks , Vol. 13, No. 2, March 2002.