We included data from 179 cognitively normal (CN) participants (58 Aβ-positive), 177 Aβ-positive patients with a clinical diagnosis of AD dementia, and 217 Aβ-positive patients with MCI (i.e. prodromal AD) from the ADNI-1, ADNI-GO/2 and ADNI-3 cohorts (adni.loni.usc.edu). Detailed inclusion criteria for the different diagnostic categories have been described in detail before (16) and are available on the ADNI website (http://adni.loni.usc.edu/methods/documents/). Evidence of Aβ pathology was based on AV45-PET or, in case this measure was not available, on CSF Aβ levels (see below for details). The ADNI is a longitudinal multicentre study aimed at investigating whether neuroimaging methods such as MRI and PET, together with genetic, clinical and neuropsychological measures can be used to characterize progression of MCI and AD. The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD.
Neuropsychological Test Scores
Cognitive performance in the ADNI is assessed using neuropsychological test batteries covering various cognitive domains. We used previously established composite cognitive scores for memory (ADNI-MEM) and for executive function (ADNI-EF) (17, 18). Additionally, we used the score on the Clock Drawing test as a measure of visuospatial function, and the Boston Naming Test score as a measure of language function(4). Mini Mental State Examination (MMSE) scores were used for characterizing global cognitive impairment. We also calculated the difference between the ADNI-MEM and the ADNI-EF composite scores (ADNI-DIFF) to characterize differential decline in these two domains. Positive values in this variable thus represent a more pronounced executive impairment compared to the memory deficit, and vice versa for negative values.
We analysed longitudinal changes in cognitive functions of the prodromal AD subtypes for participants with available follow-up data (n = 200). We used longitudinal measures of ADNI-MEM, ADNI-EF, Clock Drawing test and the Boston Naming Test. Additionally, longitudinal Clinical Dementia Rating (CDR) scores were used as a criterion indicating progression from prodromal (CDR = 0.5) to clinically manifest AD dementia (CDR ≥ 1). Mean follow-up period was 44 months (range 12-72 months), and 72.5% of participants had at least 36 months of follow-up available.
Measures of cortex-to-whole cerebellum AV45 standard uptake value ratios (SUVR) have been calculated by the ADNI PET core (Jagust Lab, UC Berkeley) and were downloaded from the ADNI server. We selected Aβ positive patients with AD or MCI if their AV45 SUVR values were greater than or equal to the recommended threshold of 1.11 (19).
ADNI CSF values in the current study were derived from electrochemiluminescence immunoassays for Aβ (1-42), phospho-Tau (181P), and total-Tau on an automated Elecsys cobas e 601 instrument. We included participants who had CSF Aβ values lower than the threshold of 880 pg/ml proposed by Hansson et al. (20).
APOE genotype was determined using DNA extracted from a 3-mL aliquot of EDTA blood samples by Cogenics (21). Genotype information was coded in a binary APOE ε4 variable indicating the presence of at least one APOE ε4 allele.
Structural MRI images were used to derive hippocampal and cortical volume measures. MRI images in ADNI are acquired at multiple sites using scanner-specific T1-weighted sagittal 3D MPRAGE sequences and undergo standardized image pre-processing steps to improve uniformity across the scanners (http://adni.loni.usc.edu/methods/documents/). We extracted regional grey matter volumes from these scans using a previously described automated volumetry approach implemented in statistical parametric mapping software (SPM8, Wellcome Trust Center for Neuroimaging) and the VBM8-toolbox (22, 23). Briefly, this involves automated tissue class segmentation and high-dimensional spatial normalization to an aging/AD-specific reference template. Spatially normalized grey matter (GM) maps were visually inspected for segmentation and normalization accuracy, and voxel values were modulated for volumetric changes introduced by the high-dimensional normalization, so that the total GM volume present before warping was preserved. Hippocampal (HV) and regional cortical grey matter volumes were extracted from these scans using regions-of-interest defined in the Harvard-Oxford anatomical atlas (24). Individual volumes were divided by the total intracranial volume (TIV), calculated as the sum of total volumes of all tissue segments. In analogy to previous MRI-based subtyping studies (2, 25), cortical grey matter volumes were extracted from selected frontal, temporal, and parietal association areas (see Supplementary table 1), summed into a measure of bilateral cortical total volume (CTV), and further used to calculate the hippocampal to cortical volume ratio (HV:CTV).
Finally, we included white matter hyperintensity (WMH) volume as a measure of small vessel vascular disease burden. These values have been calculated by the ADNI MRI core and were downloaded from the ADNI server. In ADNI-1 data WMH values were obtained via analysis of the proton density (PD), T1 and T2 MRIs (26). In ADNI-GO/2, a fluid-attenuated inversion recovery (FLAIR) MRI sequence was used to calculate WMH volumes (27). In the present study we pooled available WMH measures (4) and controlled statistical analyses of this variable for different segmentation methods using a dummy-coded confound variable.
FDG-PET data acquisition and preprocessing
FDG-PET data were retrieved in a pre-processed form from the ADNI server. FDG-PET images were obtained on multiple scanners with protocols specific to platforms. Dynamic 3D scans of six 5-minute frames were acquired 30-60 minutes after injections of 185 MBq of 18F-FDG. All original ADNI FDG-PET scans underwent standardized image pre-processing steps to improve uniformity across the scanners. Detailed information on FDG-PET acquisition and pre-processing is available on the ADNI website (http://adni.loni.usc.edu/methods/documents/). For the present study, FDG-PET images were further spatially normalized to a customized FDG-PET standard space template and smoothed with a Gaussian smoothing kernel of 8 mm full-width at half maximum (FWHM) using SPM8 (28).
The patients with AD dementia were classified into hypometabolic subtypes using agglomerative hierarchical clustering of voxel-wise FDG-PET data with Ward’s linkage as implemented in MATLAB software (29, 30). Individual FDG-PET profiles were scaled to their global mean prior to clustering analysis so that clustering relies on differences in regional FDG-PET patterns rather than on global signal differences across patients. In the clustering procedure, the algorithm progressively combines closest voxel-wise FDG-PET profiles of the participants into larger clusters, as well as most similar clusters with each other. Output of the algorithm is a hierarchical dendrogram in which the level of branching indicates the degree of dissimilarity between the clusters. The optimal number of separable clusters in the data was evaluated using standard performance measures for clustering solutions including the Davies-Bouldin criterion (31) and the silhouette criterion (32).
To visualize patterns of hypometabolism in the identified subtypes, we conducted voxel-wise two-sample t-tests between FDG-PET images from each of the subtypes and the CN group, using age, gender, and years of education as covariates. Images were scaled to the average signal in a pons reference region prior to analysis. An explicit grey matter mask was applied to the images and obtained t values were converted into Cohen’s d effect size values.
Classification of patients with prodromal AD
We classified FDG-PET scans of patients with prodromal AD according to the identified AD subtypes using a fully automated classification procedure. For that, we first screened patients for evidence of regional hypometabolism by assessing whether at least one of the 48 bilateral cortical areas defined in the Harvard-Oxford atlas had an FDG-PET signal (scaled to pons) of at least one standard deviation below the mean of the control group. Participants with no such regions were classified into a “no hypometabolism” subtype (10, 11). The remaining patients with prodromal AD were classified into one of the subtypes identified in the AD dementia group based on the smallest Euclidean distance between the individual patient’s voxel-wise FDG-PET profile (scaled to global values) and the mean FDG-PET profile of each of the AD dementia subtypes (4).
Statistical analyses were conducted using RStudio and R version 3.5.2 with a statistical significance threshold of P < 0.05 (two-tailed). Chi-squared tests with post-hoc pairwise proportion tests were used to compare gender compositions of subtypes and frequencies of the APOE ε4 genotype. For this and other post-hoc tests comparing subtypes we used the false discovery rate (FDR) correction (33) as implemented in R. Age and years of education were compared across subtypes using ANOVA. Differences in cognitive measures and biomarkers across the subtypes were tested with ANCOVA using age, gender, and education as covariates (34, 35).
For patients with prodromal AD with available clinical follow-up data, we also conducted Cox proportional hazards regression analyses for analysing differential risks of progression to dementia across FDG-PET defined subtypes. Progression to dementia was operationalized as a change in CDR score from 0.5 to ≥ 1. Models included age, gender, and education as covariates. Participants were censored if they did not progress to dementia before the last available follow-up CDR score.
In addition, linear mixed effects regression models were used to assess differences in domain-specific longitudinal changes in memory, executive function, visuospatial function or language function. Models included time of follow-up measured in months from baseline, a factor variable indicating subtype, and an interaction term for time by subtype as independent variables. The estimates for interactions between subtype and time indicated whether subtypes had differential cognitive trajectories over time. Age, gender, and education were included as covariates. Regression models included random intercepts and random slopes for participants; t-tests used Satterthwaite approximations for degrees of freedom.
 The Elecsys β-Amyloid(1-42) CSF immunoassay in use is not a commercially available IVD assay. It is an assay that is currently under development and for investigational use only. The measuring range of assay is 200 (lower technical limit) – 1700 pg/mL (upper technical limit). The performance of the assay beyond the upper technical limit has not been formally established. Therefore, use of values above the upper technical limit, which are provided based on an extrapolation of the calibration curve, is restricted to exploratory research purposes and is excluded for clinical decision making or for the derivation of medical decision points.