|
Abstract:
|
Several methods have been proposed for defining indeterminate testing results to reduce erroneous decisions. All methods approach the problem without consideration of the likely retesting in practice when a given test result is deemed indeterminate. We argue that the analysis should better reflect the actual use of the developed diagnostic test. Suppose a multistage diagnostic testing algorithm is considered where a given test may be used up to k times sequentially for a given patient, with additional testing only taking place if the result from the previous stage is indeterminate. We propose that for stages less than k, two thresholds be defined corresponding to positive, negative, and indeterminate result. For the kth stage, where a final result is required, a single threshold is defined. A generalization of the ROC and corresponding area under the curve is constructed by considering all possible values of thresholds. We evaluate the properties of the proposed method via simulation studies and results show superior performance compared to the typical ROC-based approaches, with the improvement being more substantial as k increases. An application is considered for the case of k=2.
|
