A hybrid system for diagnosing multiple disorders

This paper investigates the advantages of introducing feedback between the processes of automated medical diagnosis and automated diagnostic-knowledge acquisition. Experimental results show that a diagnostic system with such feedback is capable of an efficiency/accuracy trade-off when applied to the problem of diagnosing multiple disorders. A primary feature of this work is a new mechanism, called the "diagnostic-unit" representation, for remembering results of previous diagnoses. The diagnostic-unit representation is explicitly tailored to capture the most likely relationships between disorders and clusters of findings. Unlike typical bipartite "If-Then" representations, the diagnostic-unit representation uses a general graph representation to efficiently represent complex causal relationships between disorders and clusters of findings. In addition to the basic diagnostic-unit concept, this paper presents experience-based strategies for incrementally deriving and updating diagnostic units and the various relationships between them. Techniques for selecting diagnostic units relevant to a given problem and then combining them to generate solutions are also described.