Artificial neural networks are being used in diagnosis support systems to detect different kind of diseases. As the design of multilayer perceptron is an open question, the present work shows a comparison between a traditional empirical way and neuroevolution method to find the best architecture to solve the disease detection problem. Tuberculosis and appendicitis databases were employed to test both proposals. Results show that neuroevolution offers a good alternative for the tuberculosis problem but there is lacks of performance in the appendicitis one.