Abstract

In this paper I shall try to analyse appropriate logic and actual methods as a preliminary to developing expert systems that will simulate clinical diagnosis. It is doubtful that all diagnoses address the same kind of problem and hence no one logic will suffice. Sometimes the signs and symptoms manifest an underlying disorder that cannot be observed directly (the substantialist model); sometimes there seems to be no underlying disorder and the diagnosis is a rearrangement of the data (the nominalist model). In deciding their merits in the particular case, we must identify the sources of variation; what is meant by saying that the data are independent; and what can be meant by wrong diagnosis and misclassification. Problems arise in human diagnosis because the data may be false, even lies; because the known list of diseases is not exhaustive; and because the probability of multiple unrelated diseases increases with age. Special difficulties arise from homeostasis (which is fundamental in defining and understanding disease) because the primary disturbances may be hard to distinguish from the homeostatic responses to them.