Computer-assisted classification of disease has largely relied upon testing the diagnostic algorithm in the same population from which it was originally derived, as a means of validation. To evaluate the accuracy of a diagnostic program in which discriminant function analysis is used, we applied it to a separate population, selected by different criteria from those used to define the original case material on which the diagnostic program was based. We selected a group of 315 patients having abnormal values for alkaline phosphatase, bilirubin, or aspartate aminotransferase for further biochemical and immunological investigations. We used a computer program involving discriminant function analysis and classification procedures primed with the results of 10 tests obtained on each of 535 patients in a previous series to allocate those 173 new patients who had diseases of the liver or biliary tree into one of 13 disease groups. The classification was less accurate than was the case in previous cross-validation studies. We developed new discriminants with the new case material, using the same group of tests, and when cross-validation was performed, overall accuracy was greatly improved. These experiences point to the powerful influence of group selection upon computer-assisted diagnostic procedures, and the hazards of applying to one clinical population discriminant functions derived from a different population.