Capitalizing on complementary FDG PET and florbetapir PET data sets to distinguish between early and late mild cognitive impairment using multi-modal partial least squares

We previously introduced a voxel-based image analysis algorithm known as multi-modal partial least squares (MMPLS) to characterize the linkage between covarying patterns in two or more complementary complex data sets. Here we used the MMPLS to characterize and compare covarying patterns in FDG PET and florbetapir PET images from 150 patients with early mild cognitive impairment (eMCI) and 78 patients with late MCI (lMCI) in the Alzheimer's Disease Neuroimaging Initiative (ADNI). MMPLS was used in conjunction with SPM8 to characterize and compare MMPLS subject scores irrespective of the subject's clinical status and free from the Type I error inflation associated with multiple regional comparisons. Brain maps of the between-group differences in covarying FDG PET and florbetapir PET latent variables were compared to maps of between-group differences in FDG PET and florbetapir PET measurements using convention univariate statistics. Differences between the eMCI and lMCI MMPLS subject scores were highly significant (Hotelling 2 sample T 2: P = 2.2e-06, with no need to correct for multiple regional comparisons). By comparison, voxels associated with maximal between group differences in the florbetapir PET and FDG PET brain maps generated using conventional statistics were associated with p-values of P = 3.1e-06 and 9.0e-06, respectively, prior to any correction for multiple regional comparisons and p-values of P = 0.015 and P = 0.082 after correction. Between-group florbetapir PET and FDG PET brain maps simultaneously generated using the MMPLS were similar to those generated in independent comparisons with conventional univariate statistics. When applied to FDG and florbetapir images from the same persons, the MMPLS appears to distinguish between eMCI and lMCI patients with improved statistical power and freedom from multiple regional comparisons. Additional studies are needed to clarify the value of voxel-based image analysis algorithms that capitalize on two or more complementary data sets from the same persons in the early detection, tracking, and differential diagnosis of AD and the evaluation of AD-modifying treatments.