How to choose the training data for neural network medical diagnosis systems

This paper presents a research work on the generalization capability of Multilayers Feedforward Neural Networks with Backpropagation under the point of view of training data. The research is focused on the case of neural networks medical diagnosis systems and it is shown through 3-D plots of the neural network performance an appropriate way to select the training data distribution which should be used for the design of the neural network. A real world distribution which can be obtained by randomly sampling a medical data base is more appropriate than an equal distribution which is the normally used distribution in the bibliography. The results are presented by using a medical diagnosis data base: different diagnosis of one main symptom: vaginal discharge of non-menstruating women.