A system of selective sequential questioning is presented by which meaningful symptom-complex subsets may be constructed. Generalized, the procedure results in a flow sheet applicable to differential diagnosis. An ideal subset is homogeneous with regard to clinical diagnosis; distance from the ideal is measured by tabulating the frequency with which a subset is associated with a particular diagnosis. Given sufficient patients, this tabulation expresses the "conditional probability" that a patient with a given symptom-complex subset has a particular disease. Applied to the diagnosis of arthritis, the system proved 96% accurate and was in some respects more consistent than clinical diagnosis. This system avoids the problem of the interdependence of variables and increases usefulness by presenting the tabulation of its previous experience with the subset for evaluation by the user.