Abstract:
|
The preferred scenario in imaging studies where a Test Diagnostic classifies patients for presence/absence of a condition is to have a Truth Standard for validation of the results. In many imaging trials the Truth Standard is unavailable. For instance, the Test might be PET detection of amyloid in the brain; with Truth the validation of amyloid deposits during autopsy. In such scenarios a reliable Surrogate Truth is needed. The surrogate evaluated here is a diagnostic derived from the Test itself: A large number of readers read the images and the Surrogate binary truth is the Majority Read: ST = Surrogate Truth = 1 iff a majority of readers score image = 1 ; =0 otherwise. It is shown that ST converges to the Truth as the number of readers increases. Moreover, for each reader: Surrogate Sensitivity = P{Test = 1 | ST = 1} converges to true sensitivity and Surrogate Specificity = P {Test = 0 | ST = 0} converges to true specificity.Thus, criteria for success of a Test Diagnostic with respect to specified levels of Sensitivity and Specificity can be translated into criteria for success of the Test with respect to achievement of specified levels of Surrogate Sensitivity and Specificity.
|