Abstract
Some 700 icteric patients have now been investigated by the Copenhagen study group for computer-aided differential diagnosis of jaundice (the COMIK study). The paper presents experience based on preliminary test runs using selected subsets of the data. In particular, the effect of the number of intervals into which quantitative laboratory variables are grouped has been studied. Computer performance depends little on such grouping and, by inference, on analytical imprecision. As the grouping becomes finer, performance is even seen to deteriorate. These findings arc in accord with theory and experience from other areas of pattern recognition, the problem being one of sample size. It is concluded that we are in a paradoxical situation of knowing from studies of single analytes that precision matters; yet when it comes to assembling information about a patient's condition we cannot make effective use of the precision of today's laboratories. This would seem to hold good whether data are processed by a computer, by physician, or by the two in collaboration.