Gray Matter Covariance Networks as Classifiers and Predictors of Cognitive Function in Alzheimer's Disease

The study of shared variation in gray matter morphology may define neurodegenerative diseases beyond what can be detected from the isolated assessment of regional brain volumes. We, therefore, aimed to (1) identify SCNs (structural covariance networks) that discriminate between Alzheimer's disease (AD) patients and healthy controls (HC), (2) investigate their diagnostic accuracy in comparison and above established markers, and (3) determine if they are associated with cognitive abilities. We applied a random forest algorithm to identify discriminating networks from a set of 20 SCNs. The algorithm was trained on a main sample of 104 AD patients and 104 age-matched HC and was then validated in an independent sample of 28 AD patients and 28 controls from another center. Only two of the 20 SCNs contributed significantly to the discrimination between AD and controls. These were a temporal and a secondary somatosensory SCN. Their diagnostic accuracy was 74% in the original cohort and 80% in the independent samples. The diagnostic accuracy of SCNs was comparable with that of conventional volumetric MRI markers including whole brain volume and hippocampal volume. SCN did not significantly increase diagnostic accuracy beyond that of conventional MRI markers. We found the temporal SCN to be associated with verbal memory at baseline. No other associations with cognitive functions were seen. SCNs failed to predict the course of cognitive decline over an average of 18 months. We conclude that SCNs have diagnostic potential, but the diagnostic information gain beyond conventional MRI markers is limited.