||Neural networks have been widely used in general classification tasks. They have the advantages of being flexible and tolerant to noisy and incomplete data. However, when applied to diagnostic problems, some traditional neural network approaches such as nearest neighbour rule, Hebb rule or perceptron learning rule do not often perform satisfactorily. We propose that a diagnostic problem differs from a typical pattern recognition problem in having a non-uniform distribution of information in the data. Some diagnostic information are more informative and some appear in high correlation, thus reducing the applicability of conventional classification techniques, where data information are usually treated on the same basis. Recently, a number of rule-based expert systems such as classification trees are established to tackle problems in this nature. They work relatively well when training examples are inadequate but informative, or when the data contain much redundant information. However, due to their hierarchical structure, their performances are easily curtailed by data noise. Neural networks and classification trees are found to have complementary advantages, therefore combining them is a natural way to make the system more robust and applicable. In this thesis, we compare the classification performance of general approaches and the combined rule-based neural network approach, and present simulations for real diagnostic problems and the artificial ones. We also propose a model for diagnostic data and study the effects on generalization by varying the abundance of informative data and the strength of information contents. Results imply that different problems may require specialized techniques for better performance, and the use of classification trees as data pre-processing for perceptron classifier is found applicable to diagnostic problems.