Explaining and Generalizing Diagnostic Decisions

Abstract

We present methods for explaining, generalizing, and learning from diagnostic decisions involving tradeoffs. The methods combine decision theory with extensions of explanation-based learning and qualitative reasoning. The explanation methods search for comprehensible justifications of diagnostic decisions. “Qualitative logics of decision” are used to provide generalizations that enable users to understand individual decisions as members of classes of similar decisions. Generalization supports automated knowledge acquisition using abduction and explanation-based macro learning. We illustrate the methods using detailed examples of two prototypical diagnostic decisions involving tradeoffs, deciding whether or not to operate and whether to test or treat. We provide evaluation metrics and the results of empirical measurements. Empirical results indicate that rules learned using a coarse form of qualitative reasoning provide correct decisions on more important tradeoffs.