Neurodegenerative diseases often progress gradually in a long period in which the earliest stage is critical to recognize for early interventions. A statistical challenge is to diagnose the early disease stage accurately, in comparison to the healthy and fully developed disease stages. The accuracy of a diagnostic marker in this setting can be summarized by the volume under the ROC surface. When the diagnostic populations are clustered by families, we propose to model the marker by a general linear mixed model that takes into account of the correlation from members of the same clusters. This model facilitates the ML estimation and inferences of the diagnostic accuracy, and allows the incorporation of covariates and missing data when clusters do not have subjects on all diagnostic groups. We assess the performance of the estimator by simulations, and apply the method to the neuroimaging and biomarker data from the Dominantly Inherited Alzheimer Network, an international family-clustered registry, by estimating the accuracy of several cerebrospinal fluid and imaging biomarkers in differentiating 3 groups: normal non-mutation carriers, asymptomatic and symptomatic mutation carriers.