Evaluation of medical technologies: a generalized ROC analysis.

The author describes a new methodology to solve medical decision problems involving a choice between alternatives under conditions of risk and uncertainty when imperfect empirical information from diagnostic technologies is available. The main new concept is the generalized ROC (GROC) curve, which extends considerably the scope of analysis of the traditional ROC curve as well as the threshold approach to medical decision making. Both techniques become special cases of the new general approach. The author shows how to apply the technique and derive comprehensive clinical guidelines to solve the medical decision problem involving one patient in the most general situation. The larger problem of evaluating the performances of diagnostic technologies for a population of patients with varying prior probabilities of disease is then addressed. A new measure of performance is proposed that goes beyond the well-accepted area under the ROC curve index. This measure is used to compare different technologies for a population of patients, and a new methodology is proposed to carry out such comparisons.