In vivo staging of regional amyloid progression in healthy middle-aged to older people at risk of Alzheimer’s disease

Background We investigated regional amyloid staging characteristics in 11C-PiB-PET data from middle-aged to older participants at elevated risk for AD enrolled in the Wisconsin Registry for Alzheimer’s Prevention. Methods We analyzed partial volume effect-corrected 11C-PiB-PET distribution volume ratio maps from 220 participants (mean age = 61.4 years, range 46.9–76.8 years). Regional amyloid positivity was established using region-specific thresholds. We used four stages from the frequency-based staging of amyloid positivity to characterize individual amyloid deposition. Longitudinal PET data was used to assess the temporal progression of stages and to evaluate the emergence of regional amyloid positivity in participants who were amyloid-negative at baseline. We also assessed the effect of amyloid stage on longitudinal cognitive trajectories. Results The staging model suggested progressive accumulation of amyloid from associative to primary neocortex and gradually involving subcortical regions. Longitudinal PET measurements supported the cross-sectionally estimated amyloid progression. In mixed-effects longitudinal analysis of cognitive follow-up data obtained over an average period of 6.5 years following the baseline PET measurement, amyloid stage II showed a faster decline in executive function, and advanced amyloid stages (III and IV) showed a faster decline across multiple cognitive domains compared to stage 0. Conclusions Overall, the 11C-PiB-PET-based staging model was generally consistent with previously derived models from 18F-labeled amyloid PET scans and a longitudinal course of amyloid accumulation. Differences in longitudinal cognitive decline support the potential clinical utility of in vivo amyloid staging for risk stratification of the preclinical phase of AD even in middle-aged to older individuals at risk for AD. Supplementary Information The online version contains supplementary material available at 10.1186/s13195-021-00918-0.


Supplementary table 1.
Correlations between amyloid-positivity frequencies based on different amyloid-positivity thresholds.
1D-GMM = one-dimensional Gaussian Mixture Model approach; 2D-GMM = two-dimensional Gaussian Mixture Model approach; Young subsample = resampled frequencies using thresholds based on data from 20 youngest participants; Previous 18F-Florbetapir based model = frequencies of regional amyloid-positivity in the ADNI cohort as estimated in the previous study by Grothe et  2 Groups 0-IV represent five groupings of participants based on the five-part division of the range of global non-PVE-corrected 11 C-PiB-PET DVR signal. Global DVR groupings highly correlated with the amyloid stages (Spearman's rho = 0.782, p < 0.001).

Supplementary table 4.
Mixed-effects regression models of longitudinal change in composite cognitive scores across groups of participants stratified by global non-PVE-corrected 11 C- -0.14 -1.765 Unstandardized estimates are presented with t-statistics. * p < .05, ** p < .01, *** p < .001. Groups 0-IV represent five groupings of participants based on the five-part division of the range of global non-PVE-corrected 11 C-PiB-PET DVR signal. For interactions between the follow-up time in years and stage, Group 0 acts as a reference. Random intercepts for participants are included to account for multiple measurements.

Methods
We evaluated the impact of several PET processing methods and threshold definitions on the derived staging model. For assessing the impact of the use of region-specific amyloid-positivity thresholds on the current staging model, we estimated a comparison model in which a uniform threshold was applied to every region. The amyloid-positivity threshold was based on the previously established threshold of 1.08 (2) which was converted for use with PVE-corrected data via a regression model with the resulting value of DVR = 1.03 (Supplementary figure 2). Next, we assessed the potential effect of applying PVE-correction by deriving a staging model using non-PVE corrected PET data. Finally, we also evaluated potential differences between the main staging model derived based on 11 C-PiB-PET DVR data and a model based on SUVR data. Briefly, individual SUVR maps were obtained using dynamic PET frames from a 50 to 70 min window and using a cerebellar gray matter reference region. These SUVR maps were then processed identically to the DVR maps, including correction for partial volume effects using the 3-compartment "Müller-Gärtner" method in subject's native space (3,4), and spatial normalization to the reference template space using transformation parameters from the corresponding MRI. Analogously to the model based on DVR maps, region-specific amyloid-positivity thresholds for the SUVR data were obtained using regional and global composite amyloid signal in 2D-GMM. A frequency-based staging model was then estimated and used for staging individual 11 C-PiB-PET SUVR profiles. The correspondence between regional amyloid-positivity frequencies of these models was assessed using Spearman rank correlations.

Results
Brain renderings illustrating the regional amyloid-positivity frequencies of the different models are shown in Supplementary figure 3, and the corresponding Spearman rank correlations for pair-wise comparisons of the models are listed in Supplementary table 5. Generally, the staging models estimated using SUVR images and non-PVE-corrected data showed only relatively minor differences compared to the main model with PVE-corrected DVR data, such that the respective regional amyloid-positivity frequencies were highly correlated (rho ~ 0.86; Supplementary table 5).
By contrast, the use of a constant universal cut-off for all brain regions had a major influence on the regional staging model, yielding low rank correlations with the regional amyloid-positivity frequencies of all other models (rho ≤ 0.22). While highest frequencies among cortical regions were similarly observed in anterior and posterior midline regions, regional frequencies were also very high in subcortical structures, including the striatum and thalamus (54% and 95% respectively), when using a constant cut-off. In terms of individual staging results, all models based on alternative PET processing strategies resulted in one or two additional participants being classified as non -stageable, indicating a minor advantage of the main model (using PVE-corrected DVR maps and region-specific thresholds) in this regard (see Supplementary table 6).