Belief rule mining using the evidential reasoning rule for medical diagnosis

Abstract

A belief rule mining approach is proposed to generate belief rules with a customized set of criteria by mining from multiple belief rules that are trained using data with varied sets of criteria. As the theoretical basis of the belief rule mining approach, the key concepts are defined, including the weights and reliabilities of cases, criteria, models, and belief rules. Based on the key concepts, multiple sub-models composed of belief rules with varied sets of criteria are initialized and optimized. Then, the optimized sub-models are integrated using the evidential reasoning rule into belief rules with a customized set of criteria. In the belief rule mining process, the weights and reliabilities of the sub-models are considered according to the weight and reliability calculation procedures of models proposed in this study. The proposed approach is used to help diagnose thyroid nodules with 527 medical cases, in which its applicability is demonstrated. By comparative experiments, the diagnostic correctness of the proposed approach is verified to be higher than those of the directly-optimized model and the approach without the consideration of reliability.