A.N. Yusupov - Ph. D. (Tech.), Leading Engineer-Programmer, Senior Lecturer of Wellness Foundation \"MEDINEF\" BGTU \"Voenmech\" RTC (St. Petersburg). E-mail: a.n.yusupov@gmail.com

At the work was shown an approaches to restoration of the frequency\'s conditional function using math apparat of the artificial neural network. The statistical evaluation can be used in the different analytic systems of the experimental data including medical diagnostic methods. The frequency\'s conditional function allows to determine simultaneously the most probable patient\'s health condition (sick-healthy) and to evaluate an activity of the biological processes which are caused the investigated disease. Also it is possible to develop multiple different individual and popular indicators based on this: probability of the disease developing during defined period of time (prognosis), biological age, evaluation of the treatment efficiency, influence of harmful factors on health. An evaluation of the frequency\'s conditional function on the multidimensional data we can obtain by the «kn nearest neighbors» method also by the linear discriminant analysis (Fisher\'s model). Both of the methods have known disadvantages. For instance the «kn nearest neighbors» method is extremly sensitive to fluctuated data, and the linear discriminant analysis has a requirement to the data is not feasible in biomedical statistic. As effective alternative method for said ones artificial neuron network of the direct spread with sigma-function of neurons activation is turned out. The last slice of the network - the one neuron with activation of the sigma-function, which is completely coincide with the Fischer\'s model. Other neuron networks slices deform feature space, provide data to the form of fitness for a linear discrimination which is implemented by the last neuron. In numerous computational experiments the neural network (with the due regularization training) stably shows its obvious advantages in given frequency\'s conditional function with different forms and in indicators distribution with different characters. At the same time neural network method does not have any requirements to the nature of initial indicators, to their type of probable distribution and to the character of indicators\' interrelation that qualitatively distinguishes it from linear discriminant analysis and from the «kn nearest neighbors» method what about the demonstrative example of recovery the frequency\'s conditional function for the modeling data is evidenced.

 

1. Bikel P.Doksam K. Matematicheskaja statistika: Per. s angl. JU.A. Danilova. M.: Finansy i statistika. 1983. Vyp. 1.
2. Ossovskijj S. Nejjronnye seti dlja obrabotki informacii // Nejjronnye seti dlja obrabotki informacii: Per. s polsk. I.D. Rudinskogo. M.: Finansyistatistika. 2002. 344 s.
3. Fisher R.A. The Use of Multiple Measurements in Taxonomic Problems // Ann. Eugen. 1936. № 7. 178 p.
4. Hecht-Nilsen R. Neurocomputing. Amsterdam: Addison Wesley 1991. P. 433.