ADNI participants who underwent concurrent Aβ PET, T1-weighted and FLAIR MRI, and cerebrospinal fluid (CSF) examinations were considered in this study. See Supplementary Methods 1 for a description of the ADNI study. Participants were further stratified by Aβ status (A+/-) using standardized uptake value ratio (SUVR) values computed by the ADNI PET core and reported in the “SUMMARYSUVR_WHOLECEREBNORM_1.11CUTOFF” field of the “UCBERKELEYAV45_05_12_20” CSV file [30, 31], and in cognitively normal (CN), mild cognitive impairment (MCI), and AD dementia. Detailed information about the eligibility criteria for the different diagnostic cohorts can be found at http://adni.loni.usc.edu/methods/documents/. Since we are interested in the role of demyelination across the AD continuum, our primary analysis was focused on individuals in the Alzheimer’s continuum, i.e., those with cognitive impairment (MCI and AD) and a positive Aβ status (A+) as well as those cognitively unimpaired (both A- and A+). This resulted in a primary study cohort of 588 subjects. A subset of these participants had also DTI scans available at baseline (n=151). At least one longitudinal FLAIR scan was available in 493 study participants (mean follow-up time 2.1 years). Demographics and biomarker information of the study cohort are summarized in Table 1. In a secondary analysis, we also investigated non-AD amnestic participants, which included A- MCI (n=192) and A- AD dementia (n=15). Forty-four of these participants had available baseline DTI scans, and at least one longitudinal FLAIR scan was available in 203 participants (mean follow-up time 2.5 years). Demographics and biomarker information of this secondary study cohort is reported in Supplementary Table 1. All participants provided written informed consent approved by the institutional review board of each ADNI participating institution.
2.2 Magnetic Resonance Imaging
MRI acquisition and pre-processing protocols in ADNI are described in detail elsewhere . All included subjects were scanned in 3T MRI devices. Structural T1 images were normalized to Montreal Neuroimaging (MNI) space and segmented using Statistical Parametric Mapping 12 (SPM12, Wellcome Department of Clinical Neurology, London, UK). Baseline and longitudinal FLAIR scans were processed using the Lesion Segmentation Toolbox  in SPM12 as previously described [9, 33] to derive binary WM hyperintensity (WMH) masks.
2.3 18F-florbetapir PET
FBP PET scans were acquired and pre-processed following the ADNI pipeline  (http://adni.loni.usc.edu/methods/pet-analysis-method/pet-analysis/). For quantification of FBP retention in the WM, we followed the same approach used in . Briefly, FBP images were first coregistered to the corresponding T1 scan using SPM12. A WM mask containing both NAWM and WMH was then generated by merging the binarized WM segment from SPM12 (generated after thresholding the WM probability map with 0.5) and the binary WMH mask previously generated from FLAIR scans. Note that by combining both segmentations we avoid incorrect classification of WM hypointense areas in the T1 scan as CSF or grey matter . The inverse of the deformation field from spatial normalization was then used to propagate a predefined cerebral mask from MNI space into each individual’s native space, and this mask was intersected with the previously derived WM mask to obtain a cerebral WM mask (i.e., excluding the brainstem and cerebellar WM). To minimize partial volume effects, we 1) excluded small WMH clusters with fewer than 27 voxels (~ 27 mm3, corresponding to a 3-mm isotropic voxel) from the analysis to minimize contamination due to spill-in counts from surrounding NAWM and 2) we eroded the resulting WM mask so that voxels within 2 mm distance from any non-WM voxel were excluded . Finally, SUVR was calculated in NAWM and WMH regions by using the cerebellar grey matter as the reference region. While recent works point that the whole cerebellum  or the WM  might be superior reference regions, we choose the traditional cerebellar grey matter  to avoid any potential circular analysis due to the inclusion of cerebellar WM.
For voxel-wise analyses, coregistered FBP PET scans were spatially normalized to MNI space using the deformation fields obtained from the T1 spatial normalization, masked with a binary WM mask defined in MNI space, and smoothed using an 8-mm isotropic filter.
To statistically control analyses for continuous levels of Aβ burden within A+ and A- subjects, we also measured continuous SUVR (using cerebellar grey matter as reference region) within the ADNI ROI that was previously used for establishing Aβ status .
2.4 Fluid biomarkers
CSF levels of Aβ1-42 (Aβ-42) and phosphorylated tau181 (p-tau181) were measured using a fully automated Roche Elecsys electrochemiluminescence immunoassay batch . Plasma levels of neurofilament light chain (NfL), a marker of subcortical large-caliber axonal degeneration [40, 41], were measured from blood samples as described previously . A detailed description of CSF and blood sampling procedures can be found at http://adni.loni.usc.edu/methods/.
2.5 Diffusion tensor imaging
Diffusion-weighted images were acquired in a subset of participants (those scanned with GE scanners). Acquisition protocols, as well as pre-processing and post-processing steps, were described in detail in previous studies . Here, we used fractional anisotropy (FA) images provided by ADNI to quantify WM integrity  in NAWM and WMH. For this, FA images were coregistered to the corresponding T1 scan using SPM12. The results were visually inspected to ascertain a correct registration. Mean FA in NAWM and WMH was measured within the eroded WM mask defined in Section 2.3, after excluding voxels with FA < 0.25 to exclude spurious fibre tracts .
2.6 Clinical assessments
Global cognitive performance was assessed with the Alzheimer’s Disease Assessment Scale—Cognitive 13-Item (ADAS-Cog 13) at baseline and in subsequent annual follow-up visits. Changes in clinical diagnosis at follow-up were determined by a consensus Committee. Further details can be found at http://adni.loni.usc.edu/methods/.
2.7 Statistical Analysis
We first investigated how FBP retention in the WM correlates with cross-sectional MRI markers of WM degeneration. For this, we 1) compared FBP SUVR in NAWM and WMH using paired t-tests and 2) fitted linear models to assess cross-sectional associations between average FBP SUVR and FA in these regions. Linear models were adjusted for age, sex, and clinical diagnosis. Furthermore, since Aβ PET tracer retention in the WM has been found to be positively correlated with global tracer uptake in the cortex [28, 46, 47], which likely reflects partial volume effects but also binding to diffuse plaques and cerebrovascular amyloid angiopathy [46, 48, 49], we also regressed-out this confounding effect in order to isolate myelin binding contributions to FBP SUVRs. For this, we used the entire CN cohort to fit a linear model describing the dependence of NAWM and WMH SUVR with cortical FBP SUVR (see Supplementary Methods 2). After regressing-out cortical FBP dependence, we transformed the resulting variables to z-scores using CN levels as reference. We referred to this adjusted measure as “adjusted NAWM (or adjusted WMH) SUVR” in the present and subsequent analyses.
We also investigated whether low FBP uptake in NAWM is an early marker of WM damage in longitudinal analyses. Linear mixed models with subject-specific intercepts, adjusted for age, sex, and clinical diagnosis, were fitted to assess whether adjusted NAWM SUVR was a predictor of faster WMH accumulation.
Next, we investigated how FBP retention in the WM changes across the AD continuum. Voxel-wise analyses adjusted for age, sex, and global cortical FBP SUVR were conducted in the WM to assess group-level differences in FBP SUVR across the preclinical, prodromal, and dementia stages of AD. The A- CN sample was used as the reference group, and statistical maps were thresholded using pFDR<0.001. We also conducted ROI-level analyses comparing adjusted SUVR in NAWM and WMH across AD stages. In addition, we investigated the diagnostic added value of these two measures for identifying advanced disease stages in AD. For this, we analysed the incremental discriminative accuracy provided by adjusted NAWM and WMH SUVR to classify A+ CN vs A+ MCI and A+ CN vs A+ AD. The areas under the ROC curve (AUC) of two logistic regression models, one including cortical SUVR as a predictor and the second including both cortical SUVR and adjusted NAWM or WMH SUVR, were compared using a bootstrap procedure. Both models included age and sex terms as covariates. In addition, in the smaller subset with available DTI scans, we also performed preliminary a head-to-head comparison of the discriminative accuracy provided by the latter models in comparison to a model based on FA, cortical SUVR, and covariates.
We then investigated the associations of FBP retention in NAWM and WMH with fluid biomarker levels. Linear regressions adjusted for age, sex, clinical diagnosis, and cortical FBP SUVR were separately fitted in A- and A+ subjects to assess the relationship between fluid biomarkers and adjusted SUVR in NAWM and WMH. Fluid biomarker levels were log-transformed to reduce the skewness of model residuals.
Finally, we used linear mixed effects models with subject-specific random intercepts to investigate whether demyelination in NAWM, as reflected by low adjusted SUVR in this region, is associated with accelerated rates of longitudinal cognitive decline. The models were adjusted for age, sex, and years of education, as well as for global cortical FBP SUVR and WMH volume given their known associations with longitudinal cognitive deterioration [50, 51]. Cox regression, adjusted for the same covariates, was used to test whether adjusted SUVR in NAWM was associated with increased risk of progression to MCI or dementia. These associations were also tested using dichotomized versions of adjusted NAWM SUVR; the cut-point was defined as the adjusted NAWM SUVR that results in 90% sensitivity for the identification of A+ AD dementia participants [52, 53], yielding an adjusted NAWM SUVR cut-point of -0.57.
Following guidelines from the statistical literature that do not recommend the use of multiple comparisons correction for hypothesis-driven studies with a limited number of planned comparisons , we did not perform multiple comparisons correction except for voxelwise analyses.