Estimating confidence interval of difference in optimal accuracy achieved by overlapping (or nested) groups of clinical biomarkers under multivariate normality

In biomedical studies, it is a common practice to combine multiple clinical markers/tests available to achieve higher overall accuracy. In this paper, we consider some important problems in binary classification problem arising from biomarker combination. This paper focuses on the settings with multivariate normality under which optimal combination coefficients have a closed-form. Specifically, the aim is to study confidence interval estimation of the difference in optimal accuracy measured by area under ROC curve (i.e. AUC) achieved by two overlapping groups of biomarkers as well as two nested groups of biomarkers. The simulation studies indicate that proposed procedures have satisfactory coverage at finite sample sizes. A subset from Alzheimer's Disease Neuroimaging Initiative (ADNI) study is analyzed to illustrate the usage of proposed approaches in practice.