ESTIMATING ALZHEIMER’S DISEASE ANATOMICAL RISK ACROSS IMAGING DATABASES USING HIGH-DIMENSIONAL MACHINE LEARNING METHODS

Abstract

Significant efforts have been dedicated to derive Alzheimer's disease (AD) biomarkers using structural brain MRI. Previously we developed the AD Pattern Similarity (AD-PS) scores using elastic net regularized logistic regression (EN-RLR) operating directly in the whole-brain voxel space based in the Alzheimer's Disease Neuroimaging Initiative (ADNI). Here we extend this work by estimating the AD-PS scores for the Women Health Initiative Memory MRI (WHIMS-MRI) study cohort. In WHIMS-MRI, brain MRI scans (n=1365) were performed at baseline in 2005–2006 and during the follow-up (n = 717) in 2009–2010 across 14 US clinical sites. MRI data from 359 ADNI-1 participants and WHIMS-MRI participants were jointly processed using the Advanced Normalization Tools software. After segmenting and aligning MRI data from both studies into a common space, EN-RLR classifiers were estimated using grey matter (GM) images from 188 cognitively normal (CN) participants and 171 AD patients from ADNI-1. Baseline and follow-up GM images from the WHIMS-MRI cohort were provided as input to the EN-RLR classifiers to generate class-conditional probabilities (called AD-PS scores) of having AD-like spatial patterns of atrophy. To assess the construct validity of both baseline and follow-up AD-PS scores in WHIMS-MRI, we examined their associations with age, global cognitive function (assessed by the modified mini-mental state examination or 3MS), white matter lesion volume (WMLV) and overall mortality. The scores estimated using WHIMS-MRI baseline scans discriminated AD patients from CN individuals producing sensitivity of AD detection of 90.0% while specificity was 82.8% using a 0.5 probability threshold. The AD-PS score at baseline were correlated with age (rho = 0.42, p <0.001), WMLV (rho = 0.24, p <0.001), mortality (rho = 0.28, p<0.001), and associated with lower 3MS (rho=-0.19, p < 0.001). A similar pattern of associations was found with AD-PS scores at the follow-up (age:rho = 0.37, p <0.001; 3MS:rho = -0.17, p <0.001; WMLV:rho = 0.20, p <0.001; and mortality:rho = 0.20, p<0.001). Neuroanatomic measures of AD risk estimated for a nationwide cohort of older women in the US were found to be significantly correlated with age, cognitive function, WM lesion volumes, and mortality.