Extracting Rules from Composite Neural Networks for MedicalDiagnostic Problems

• Authors: Mu-Chun Su;Hsiao-Te Chang
• Venue: Neural Processing Letters
• Year: 1998

Recently, neural networks have been applied to many medical diagnosticproblems because of their appealing properties, robustness, capability ofgeneralization and fault tolerance. Although the predictive accuracy ofneural networks may be higher than that of traditional methods (e.g.,statistical methods) or human experts, the lack of explanation from atrained neural network leads to the difficulty that users would hesitate totake the advise of a black box on faith alone. This paper presents a classof composite neural networks which are trained in such a way that thevalues of the network parameters can be utilized to generate If-Then ruleson the basis of preselected meaningful coordinates. The concepts andmethods presented in the paper are illustrated through one practicalexample from medical diagnosis.