Categorical and Probabilistic Reasoning in Medical Diagnosis

In the mid-1970s, when Gorry left M.I. T. to go to Baylor College of Medicine, Peter Szolovits took over as head of the Clinical Decision-Making Group at Project MAC (now known as the Laboratory for Computer Science). He renewed ties with the collaborators at Tufts University with whom Gorry had previously worked (Pauker, Schwartz, and Kassirer). The following chapter is an early result of those developing ties. It was written for a special issue of Artificial Intelligence that dealt solely with applications of AI in biomedicine (Sridharan, 1978). In the article Szolovits and Pauker review the lessons of the major four AIM programs of the early 1970s. The review begins by noting that medical decision making can be viewed along a spectrum, with categorical (or deterministic) reasoning at one extreme and probabilistic (or evidential) reasoning at the other. The authors discuss classical flow charts as the prototype of categorical reasoning and decision analysis as the prototype of probabilistic reasoning. Within that context they compare MYCIN, PIP, CASNET, and INTERNIST-the four systems described in Chapters 5 through 8. They note that, although all four systems can exhibit impressive expertlike behavior, none of them is capable of truly expert reasoning. They argue that a program that can demonstrate expertise in the area of medical consultation will have to use a judicious combination of categorical and probabilistic reasoning-the former to establish a sufficiently narrow context and the latter to make comparisons among hypotheses and eventually to recommend therapy. We include the paper here because it nicely summarizes and integrates the