Novel Estimation Methods in Three-category Medical Diagnostic Study

In clinical practice, disease progression is often complex and may involve three ordinal diagnostic stages: non-diseased (healthy), early diseased, and fully diseased. For example, mild cognitive impairment (MCI) serves as a transitional phase between normal aging and advanced Alzheimer’s Disease (AD). To this end, several summary measures for three-category medical diagnostic test have been developed, including the three-class Youden index, the volume under the ROC surface (VUS), and sensitivity to the early diseased stage. However, in practice, confirmation of disease status using a gold standard (GS) test is often limited due to ethical concerns, cost, or invasiveness. As a result, only a subset of patients undergo disease verification, introducing verification bias and leading to inaccurate estimates of diagnostic accuracy if unadjusted. This dissertation addresses this challenge in three parts. First, we propose point estimators and confidence intervals for the three-class Youden index in the presence of verification bias. Second, we extend our methodology to the VUS, developing bias-corrected estimators that account for incomplete disease verification. Third, we propose estimation methods for sensitivity to the early diseased stage, conditioned on fixed specificity and sensitivity to the fully diseased stage, also under verification bias. Simulation studies and real-world applications demonstrate that the proposed estimators are accurate and robust across a range of scenarios. These bias-corrected summary indices provide a robust framework for evaluating three-class diagnostic tests when full verification is not feasible. They not only improve the validity of diagnostic test assessments but also facilitate more informed decision-making in the early detection and management of progressive diseases such as AD.