Joint graph embedding applied to networks estimated from correlations in volumetric data reveal Alzheimers neurodegeneration patterns

Abstract

AbstractBackgroundNeurodegeneration measured through volumetry in MRI is recognized as a potential Alzheimer’s Disease (AD) biomarker, but its utility is limited by lack of specificity. Quantifying spatial patterns of neurodegeneration, rather than a few structures of interest, may help address this. We turn to network based analyses developed for studying functional connectivity from signal correlations, measured through fMRI on the time scale of seconds. Our contribution applies these methods to studying volume‐change correlations, measured through structural MRI on the time scale of years.MethodWe study the ADNI1 3Yr1.5T longitudinal dataset, with scans roughly every 6 months for 3 years, and FreeSurfer volumes shared publicly by UCSF. We formed two groups using clinical dementia rating thresholded at <=1. We measured correlation between 108 structure volumes for 494 (322 vs 172) individuals, and computed adjacency matrices by thresholding at magnitude 0.8 (Figure a).We model this dataset with the multiple random eigengraphs framework. Edge probabilities are computed from population‐specific network modes (i), and patient‐specific loadings (ii) (Figure b). We extend the method to guarantee meaningful finite‐sample results: a constant times identity is included to model the diagonal, and a sigmoid ensures probabilities are in [0,1].The parameter sets (i) and (ii) are estimated by maximum likelihood using gradient descent, and F statistics are computed to analyze group differences. Familywise error rate is controlled at 5% using permutation testing on the maximum statistic.ResultUsing a Scree plot, we found four network modes gave accurate graph embeddings (Figure c). We found three networks gave statistically significant differences. As seen in Figure d, limbic, temporal, and ventricular structures dominate these networks.ConclusionThe results reveal networks dominated by structures associated to AD neurodegeneration, indicating the framework has promise for studying AD. It provides an opportunity to leverage graph based tools, already well accepted in fMRI, to study neurodegeneration patterns. It has the potential to uncover networks that may increase the specificity of neurodegeneration biomarkers over those revealed from conventional mass‐univariate analysis.