Diagnostic tests: a statistical review.

Common measures of the accuracy of diagnostic tests are reviewed. It is shown that the actual performance (predictive value) of these tests depends not only on their sensitivity and specificity, but also on the prevalence of the disease in the population tested (Bayes' theorem). The effect of an inaccurate "gold standard" on the calibration of a new diagnostic test is discussed. Receiver operating characteristic (ROC) curves are introduced as a tool for selecting an optimal cutpoint for a test, and for comparing different tests. Schemes are given for combining tests to improve their accuracy. When multiple continuous measurements are available, methods of discriminant analysis (and logistic regression) are shown to provide measurement combinations with improved accuracy. Examples and key references are provided.