High con dence association rules for medical diagnosis

This paper elaborates a simple and general decision model based on the so-called con rmation rules. Con rmation rules are generated separately for each diagnostic class so that selected rules cover (and should hence be able to reliably predict) a signi cant number of cases of the target class. At the same time, a con rmation rule should not cover the cases of non-target diagnostic classes, and when used for prediction it should exclude the possibility of classifying any of the nontarget cases into the target class. In this work we have used and tested the association approach for rule generation, accepting only extremely high con dence rules with reasonable support level as potentially good con rmation rules. Experimental results in the problem of coronary artery disease diagnosis illustrate the approach.