The method of neuron weight vector initial values selection in Kohonen network

Diagnosing of morbid conditions by means of automatic tools supported by computers is a significant and often used element in modern medicine. Some examples of these tools are automatic conclusion-making units of Parotec System for Windows (PSW). In the initial period of PSW system implementation, the units were used for recognition of orthopaedic diseases on the basis of the patient's walk and posture [15,17]. Subsequently, many additional options have been implemented, which have been used for purposes of diagnosing neurological diseases [1,2,3,9,12]. During automatic classification of diseases the additional units use elements of neural networks. The vectors based on normalised diagnostic measures [3] are inputs of the units. The measurements describe a patient's posture condition, his walk and overloads occurring on his feet. The Counter-Propagation (CP), two-layer network has been used in one of the automatic conclusion-making units. During CP network activity, we can see not only supervised but unsupervised learning processes as well. This is a characteristic feature of the CP network. The initial steps of the CP network learning process are very important, because the success of the network training process depends on them to a great extent. Therefore, a new method of weight vector initial values selection was proposed. The efficiency of the method was compared with classical methods. The results were very satisfactory. Owing to the proposed method, the time of the network training process as well as the mean-square error and the classification error was reduced. The research has been carried out using clinical cases of some neurological diseases: Parkinson's Disease, left-lateral hemiparesis and right-lateral hemiparesis after ischemic stroke. The measurements, which were made on a control group of patients without any neurological diseases, were the reference for these diagnostic classes.