The FDG/PET methodology for early detection of disease onset: a statistical model

The development of appropriate statistical methodologies for neuroimaging studies is dependent upon the research question of interest. Often studies are analyzed with techniques that may not be appropriate for the research question but are accepted owing to convention, familiarity, or apparent statistical sophistication. Neuroimaging data are particularly complex owing to (a) the high number of potential dependent variables (i.e., regions of interest) coupled with the practical limitations on sample size; (b) the known physical properties of scanners (e.g., resolution) interacting with the intricate and variable structure of the human brain; and (c) mathematical properties introduced into the data by the physiological model for quantification. In this article, a statistical model will be discussed for addressing a particular problem in clinical studies. Given that there is a characteristic abnormality in regional glucose metabolism in a specific disease, can a probabilistic statement be made with confidence regarding the likelihood of an individual scan being similar to those from the disease group or normal subjects? The model capitalizes on known statistical aspects of normal regional glucose metabolism. To illustrate the model, data will be presented on normal subjects, patients with confirmed Huntington's disease, and subjects at risk for the disease. Reliability and clinical validity of the model will be discussed.