Shusaku Tsumoto and Hiroshi Tanaka
Automated knowledge acquisition is an important research area to solve the bottleneck problem in developing ezpert systems. For this purpose, several methods of inductive learning, such as induction of decision trees, AQ method, and neural networks, have been introduced. However, most of the approaches focus on inducing rules which classifT/ cases correctly. On the contrary, medical experts also learn other information which is important for medical diagnostic procedures from databases. In this paper, a rule-induction system, called PRIMEROSE3 (Probabilistic Rule Induction Method based on Rough Sets version 3.0), is introduced. This program first analyzes the statistical characteristies of attribute value pairs from training samples, then determines what kind of diagnosing model can be applied to the training samples. Then, it extracts not only classification rules for differential di-agnosis, but also other medical knowledge needed for other diagnostic procedures in a selected diagnosing model. PRIMEROSE3 is evaluated on three kinds of clinical databases and the induced results arc compared with domain knowledge acquired from medical experts, including classification rules. The experimental results show that our proposed method correctly not only selects a diagnosing model, but also extracts domain knowledge.