Modeling diagnosis at multiple levels of abstraction

This dissertation develops a method for computer-based diagnostic problem-solving occurring at multiple levels of abstraction. The basic motivation for this research is to model human experts' inference behavior in initially forming a set of diagnostic hypotheses for given abnormal findings at a high level (i.e., using general diagnostic classes). Not only is this an efficient strategy, but it also serves to focus diagnostic problem-solving attention during subsequent consideration of more specific disorders as further evidence becomes available. This research is circumscribed to automating the use of causal knowledge between disorders and observed abnormal findings. The model of diagnostic problem-solving developed in this research consists of two major parts. First, a knowledge representation framework is developed for the encoding of causal knowledge in terms of tangled classification hierarchies (is-a hierarchies). Such knowledge is directly acquired from domain experts and contains the information needed to efficiently and appropriately focus diagnostic problem-solving attention at multiple levels of abstraction. A systematic analysis of the semantics of high-level causal knowledge in that context is performed. High-level causal relationships are defined in terms of a virtual underlying knowledge base consisting of specific causal associations. Desirable conditions (completeness, consistency, and adequacy) for a multi-level causal knowledge base are identified together with set of guidelines helping the acquisition of such knowledge. Second, a diagnostic inference mechanism is developed that specifies how to use the encoded knowledge to guide the formation of plausible diagnostic hypotheses for observed abnormal findings at multiple levels of abstraction based on the parsimonious covering approach to diagnostic problem-solving. A prototype implementation of the ideas developed in this research is constructed to demonstrate how they can be applied in the actual construction of knowledge-based diagnostic systems. This research has advanced the understanding of how to automate diagnostic problem-solving at multiple levels of abstraction by systematically studying some of the important issues involved in such a problem-solving process. The knowledge representation and inference model developed provide a precise framework for encoding knowledge and performing diagnosis at multiple levels of abstraction suitable for future evaluation with real-world applications. The prototype implementation has shown encouraging results demonstrating that the presented framework can lead to desirable inference behavior.