Refinement of medical knowledge bases: a neural network approach

Abstract

One important issue in designing medical knowledge-based systems is the management of uncertainty. Among the schemes that have been developed for this purpose, probability and CF (certainty factor) are the most widely used. If rules are organized according to a connectionist model, then neural network learning suggests a promising solution to this problem. When most rules are correct, semantically incorrect rules can be recognized if their associated certainty factors are weakened or change signs after training with correct samples. The techniques for rule base refinement are examined under this approach. The concept has been implemented and tested in an actual medical expert system.<>