Three methods for generating psychiatric diagnoses by computer are compared on measures of agreement between computer and clinical diagnoses made on actual samples of cases. Rules for two of the methods, Bayes and discriminant function classification, were derived from characteristics of subjects in a developmental sample. Rules for the third method, DIAGNO II—an example of the logical decision tree approach to differential diagnosis—had been derived a priori.
The three methods performed equally well on a cross-validation sample drawn from the same population as the developmental sample. DIAGNO II performed most accurately on a cross-validation sample from a new population. Reasons are given why, at the present time, a logical decision tree method such as DIAGNO II is preferable to the Bayes and discriminant function methods for computer diagnosis.