An analysis of diagnostic reasoning. 3. The construction of clinical algorithms.

As "input" data are converted to "output" conclusions, diagnostic reasoning traverses a complex series of intermediate decisions, each of which is intended to identify and preferably to explain the entities cited in the preceding stages (1, 2). Because these intermediate decisions are ignored during the formulation of Bayesian and other statistical theories (3-11) about the diagnostic process, a purely statistical approach to diagnosis has two insurmountable handicaps (2). For purposes of identification, calculations of statistical probability cannot provide the precise diagnostic evidence that is desired in modem science and that can often be obtained with suitable technologic tests. For purposes of explanation, current statistical strategies do not delineate the sequence of morbid anatomic and pathophysiologic entities that act as "proximate causes" for the observed clinical manifestations. The statistical conclusions may produce the name of a "disease" as a likely candidate in diagnostic nomenclature, but they do not demonstrate the disease, or explain what has happened. The statistical strategies, however, have a powerful intellectual attraction. Because the input data are specified, and because their manipulation with Bayesian or other calculations is also specified, statistical strategy offers the scientific advantage of expressing a rational process in mathematical symbols. This advantage would be lost if clinicians, trying to preserve their customary "art" in diagnostic reasoning, were to renounce the new statistical formulations in favor of traditional methods of branching logic. The total rejection of computational tactics in diagnosis would deprive clinicians of a unique scientific opportunity to elevate their mode