Adaptive Model for Diseases Number Prediction Based on Neuro - Fuzzy Technique ALEXANDER ROTSHTEIN, MORTON POSNER HANNA RAKYTYANSKA Dept. of Industrial Engineering and Management Jerusalem College of Technology – Machon Lev Dept. of Applied Mathematics Vinnitsa State Technical University 21 Havaad Haleumi, 91160, Jerusalem 95 Khmelnitske sh., 21021 Vinnitsa ISRAEL UKRAINE Abstract: - A neuro-fuzzy method is proposed for diseases number prediction on the basis of expert regularities that can be revealed in available experimental data. The neuro-fuzzy technology allows to organize the training process in real time and to use all of new experimental data for on-line improvement of the prediction model. Key-Words: - Fuzzy logic, linguistic approximation, neuro-fuzzy network, training of neuro-fuzzy network 1 Introduction The prediction of the number of diseases of some type or other is a necessary element of organization of medical-preventive measures. From a formal viewpoint, this problem is related to a wide class of problems of predicting discrete sequences [1] origin- ating not only in medicine but also in engineering, economics, sociology, etc. The nontrivial nature of the prediction of discrete sequences is due to the fact that, in contrast to well-algorithmisized interpolation procedures, the prediction requires the extrapolation of data on the past to data on the future. In this case, it is necessary to take into account an unknown low governing a process generating discrete sequences. A great number of papers deal with the development of mathematical models of prediction. The methods based on probabilistic-statistical means are most widely used; however, their use requires a considerable amount of experimental data, which are not always available under the conditions of even recent events. Interest has recently been revived [2] on the use of artificial neural networks for the solution of prediction problems. The networks are considered as universal models akin to the human brain, which are trained to recognize unknown regularities. However, a large sample of data is required in the case of training neural networks, as well as in the case of using probabilistic-statistical methods. Moreover, a trained neural network does not permit to explicitly interpret the weights of arcs. An approach of diseases number prediction based on expert fuzzy IF-THEN rules has been proposed in [3]. The problem of IF-THEN rules training has been solved by off-line optimization technique (gradient or genetic algorithm). This paper proposes neuro-fuzzy model for diseases number prediction which gives the possibility of training fuzzy knowledge bases in real time, i.e. on-line training. 2 Linguistic Model of Prediction We consider information on the incidence of appendicular peritonitis disease according to the data of the Vinnitsa clinic of children’s surgery in 1982-2000 that are presented in Table 1. Table 1. Distribution of the diseases number Year 1982 1983 1984 1985 1986 1987 1988 1989 Number of diseases 109 143 161 136 161 163 213 220 Year 1990 1991 1992 1993 1994 1995 1996 1997 Number of diseases 162 194 164 196 245 252 240 225 Year 1998 1999 2000 Number of diseases 160 185 174 Analyzing the disease dynamics in Fig.1, it is easy to observe the presence of four-year cycles whose third position occupies a leap year. Fig. 1. Disease dynamics These cycles will be denoted as follows: 260 232 210 180 140 100 High Above average Low Average Below average Sickness rate: 1982 1986 1990 1994 1998 2000 } { } { ... 1 1 4 3 2 1 1 4 + - i i i i i i x x x x x x ... leap year