Longitudinal validity of PET‐based staging of regional amyloid deposition

Abstract Positron emission tomography (PET)‐based staging of regional amyloid deposition has recently emerged as a promising tool for sensitive detection and stratification of pathology progression in Alzheimer's Disease (AD). Here we present an updated methodological framework for PET‐based amyloid staging using region–specific amyloid‐positivity thresholds and assess its longitudinal validity using serial PET acquisitions. We defined region‐specific thresholds of amyloid‐positivity based on Florbetapir‐PET data of 13 young healthy individuals (age ≤ 45y), applied these thresholds to Florbetapir‐PET data of 179 cognitively normal older individuals to estimate a regional amyloid staging model, and tested this model in a larger sample of patients with mild cognitive impairment (N = 403) and AD dementia (N = 85). 2‐year follow‐up Florbetapir‐PET scans from a subset of this sample (N = 436) were used to assess the longitudinal validity of the cross‐sectional model based on individual stage transitions and data‐driven longitudinal trajectory modeling. Results show a remarkable congruence between cross‐sectionally estimated and longitudinally modeled trajectories of amyloid accumulation, beginning in anterior temporal areas, followed by frontal and medial parietal areas, the remaining associative neocortex, and finally primary sensory‐motor areas and subcortical regions. Over 98% of individual amyloid deposition profiles and longitudinal stage transitions adhered to this staging scheme of regional pathology progression, which was further supported by corresponding changes in cerebrospinal fluid biomarkers. In conclusion, we provide a methodological refinement and longitudinal validation of PET‐based staging of regional amyloid accumulation, which may help improving early detection and in‐vivo stratification of pathologic disease progression in AD.

Numerous studies have shown that early phases of amyloid accumulation can be observed in cognitively unimpaired older individuals decades before the onset of dementia (Braak et al., 2011;Jack et al., 2013;Sperling, Mormino, & Johnson, 2014;Thal et al., 2002).
Amyloid imaging with positron emission tomography (PET) has become an important diagnostic tool for AD (Villemagne et al., 2011), displaying high sensitivity and specificity by comparison with neuropathological findings Sabri, Seibyl, Rowe, & Barthel, 2015;Villeneuve et al., 2015). Recent amyloid PET studies have attempted to characterize and stage the regional amyloid pathology spread in-vivo, either using an a priori distinction between early neocortical and later subcortical amyloid deposition (Cho et al., 2018;Hanseeuw et al., 2018;Thal et al., 2018) or more comprehensive data-driven models including regional deposition differences within both cortical and subcortical areas (Cho et al., 2016;Grothe et al., 2017;Sakr et al., 2019). In our previous work  we have developed an in-vivo staging model that adopts a commonly used analytic approach for determining regional staging schemes in neuropathological studies (Braak & Braak, 1991;Josephs et al., 2016;Thal et al., 2002) and is based on the frequency of regionally measured uptake positivity in cognitively normal older participants. The hierarchical stage model showed a progression pattern evolving from temporobasal and frontomedial areas (I), over the remaining associative neocortex (II), to primary sensory-motor cortex and the medial temporal lobe (III), and finally the striatum (IV). Across individuals the amyloid deposition patterns were highly consistent with the predicted hierarchy across these brain regions, allowing to classify almost all individual profiles (>98%) into one of four progressive amyloid stages. Furthermore, in a follow-up study we were able to show the validity of this model in an independent cohort consisting of people with subjective memory complaints enrolled in the French INSIGHT-pre-AD study (Sakr et al., 2019). The first stages of amyloid accumulation identified by the in-vivo staging scheme were already paralleled by concomitant decreases in CSF-Aβ values, but in both cohorts were largely classified as amyloid-negative by commonly used mean global cortical radiotracer uptake measurement. Moreover, while AD dementia patients unanimously exhibited advanced in-vivo amyloid stages (III/IV), amyloid burden among nondemented individuals varied across the full spectrum of amyloid stages and correlated with decreased memory scores in these individuals.
Together, the proposed method of in-vivo amyloid staging shows promising results for early detection of amyloid accumulation and a stratification of the predementia phase of AD based on an individual's extent of amyloid pathology. However, in analogy to neuropathologic staging schemes, the model describes cross-sectional data, and thus it is still unclear whether this frequency-based approach reflects actual longitudinal progression patterns. Moreover, it is important to keep in mind that the staging approach also depends on a range of methodological settings that may affect the final staging outcome. In particular, our staging approach as originally devised employed a constant threshold for defining regional amyloid-positivity, which may not accurately account for differing noise levels across different brain regions. Thus, amyloid PET studies have shown considerable regional signal variations even in young adults (Joshi et al., 2012) who are highly unlikely to exhibit any cerebral amyloid deposition (Braak et al., 2011;Kok et al., 2009), indicating region-specific signal confounds.
In the present study we updated our original in-vivo amyloid staging method using a technically more appropriate approach for detecting regional amyloid-positivity based on region-specific signal thresholds defined in a young control population, and then further assessed the longitudinal validity of the cross-sectionally estimated staging model by studying individual amyloid progression patterns in serial PET assessments.

| Participants
As in our previous staging study  In addition, we included Florbetapir-PET scans of 13 cognitively normal young controls (YC) to estimate region-specific amyloidpositivity thresholds. All YC subjects were younger than 45 years of age (mean ± SD: 27 ± 4.3 y), the sample had equal gender distribution (M/F: 6/7), and low APOE4 positivity prevalence (9 negative, 1 positive, 3 unknown). The data was originally acquired to serve as a YC reference data set for scaling of global Florbetapir-PET signal to the standardized Centiloid scale (Navitsky et al., 2018a)  As reported previously Navitsky et al., 2018a), written informed consent was obtained for all study participants and/or authorized representatives, and data collection and sharing were performed in accordance with the ethical standards of the institutional and/or national research committees (http://adni. loni.usc.edu/methods/documents/) (Navitsky et al., 2018a). Only deidentified imaging and clinical data was used for the analysis according to data sharing and publication policies of the providing institutions (http://adni.loni.usc.edu/data-samples/access-data/).

| Imaging data
Detailed acquisition and standardized pre-processing steps of the Pre-processing and analysis of both baseline and follow-up imaging data followed previously reported procedures Sakr et al., 2019). Briefly, MRI images were realigned, segmented into tissue types, and spatially normalized to an aging/AD-specific reference template space using standard procedures (Ashburner, 2007) implemented in Statistical Parametric Mapping software (SPM8, the Wellcome Trust Centre for Neuroimaging). Florbetapir-PET images were rigidly aligned to their corresponding structural MRI scan in native space, corrected for partial volume effects (PVE) using the "Müller-Gärtner" method (Gonzalez-Escamilla et al., 2017;Müller-Gärtner et al., 1992), and spatially normalized to the reference template space using transformation parameters from the corresponding MRI. Fifty-two regional volumes of interest (rVOI) were defined within the reference template using the Harvard-Oxford atlas, including 48 different cortical regions, as well as the hippocampus, amygdala, striatum, and thalamus, all masked for >50% gray matter probability. Regional standard uptake value ratios (SUVR) for rVOI were calculated by scaling mean tracer uptake values to the mean uptake value of the whole cerebellum. The employed 3-compartment PVE correction method (Müller-Gärtner et al., 1992) has been shown to provide improved estimates of regional amyloid tracer binding in the brain's gray matter (Gonzalez-Escamilla et al., 2017;Rullmann et al., 2016), but it does not provide values for the white matter compartment. Thus, the signal of the whole cerebellar reference region (including cerebellar gray and white matter) that is most commonly used for scaling Florbetapir-PET data Navitsky et al., 2018a) was extracted from the PET data before PVE correction .
Adopting a commonly used approach for defining global signal cut-offs (Jack Jr et al., 2017;Joshi et al., 2012), regional amyloidpositivity within each rVOI was defined as 1.65 SDs above the mean value in the YC sample. Acquisition parameters for the YC imaging data followed similar protocols as for the ADNI data and are detailed in the original publication of this dataset (Navitsky et al., 2018a). Preprocessing and analysis of Florbetapir-PET images of the YC group followed the identical MRI-based processing pipeline as described above. Given the relatively small sample size, mean and SD values for all rVOIs were calculated on 10,000 bias-corrected and accelerated bootstrap resamples. In sensitivity analyses we used alternative cutoff definitions based on 2 and 3 SDs above the mean value, as well as using the maximum value observed in the YC sample.
For comparison, we also studied standard global cortical SUVRs (with whole cerebellum reference) for all ADNI PET scans using centrally calculated values that are made available on the ADNI server (https://adni.loni.usc.edu/methods/pet-analysis). Global amyloidpositivity was defined using a recommended cut-off of SUVR >1.1 (Joshi et al., 2012;Landau et al., 2014).

| Cerebrospinal fluid (CSF) AD biomarkers
Amyloid-β 1-42 peptide (Aβ 1-42 ), total tau (t-tau), and phosphorylated tau (p-tau) in the CSF were measured using the multiplex xMAP Luminex platform with Innogenetics immunoassay kit-based reagents. Immunoassay reagents, analytical platform, and other details of biomarker quantification are described elsewhere (Shaw et al., 2009). 90% of the included ADNI sample in this study (603 subjects) had complete baseline CSF biomarker data measured at the same study point as the neuroimaging data.

| Cognitive assessment
Cognition was assessed using the mini-mental state examination (MMSE) (Folstein, Folstein, & McHugh, 1975) as a measure of global cognitive performance, as well as 30-min delayed recall (DR) of the Rey Auditory Verbal Learning Test (Schoenberg et al., 2006) and the Trail Making Test B (TMT-B) (Reitan, 1958) as domain-specific measures of episodic memory and executive function, respectively.
2.5 | Cross-sectional model of regional amyloid progression and staging scheme Estimation of a regional amyloid progression model and derived staging scheme followed the same general approach as in our previous staging study , including calculation of regional frequencies of amyloid-positivity in the CN cohort (percentage of individuals showing suprathreshold SUVR in the different rVOI) as an indicator of regional pathology progression, and grouping single regions into four larger anatomical divisions based on their estimated temporal involvement (frequency count). Methodological improvements over the original approach included the calculation of regional amyloid-positivity based on region-specific cut-offs as determined in the YC data, as well as defining the four larger anatomical divisions using the optimal k-means clustering algorithm (Wang & Song, 2011) instead of equal partitions of the observed amyloid-positivity frequency range. Moreover, in sensitivity analyses we assessed potential effects of confounding factors on the estimated regional progression pattern by repeating the analysis in separate subsamples of males and females, APOE4 carriers and non-carriers, older and younger individuals (using a median split at 73y), as well as higher and lower educated individuals (median split at 16 years of education). The similarity of the progression sequences derived from the different subgroups was assessed by calculating pair-wise Spearman rank correlations between the respective regional frequency counts . were also used to derive the regional progression model, an unbiased estimate of non-stageable CN subjects was calculated using 10,000 resamples of randomly selected test and train subsamples. In sensitivity analyses, we assessed staging outcomes when systematically altering both the rVOI positivity requirement for an anatomical division to be considered positive (initially set to 50%) as well as the overall number of considered anatomical divisions (initially set to 4).

| In-vivo stages and longitudinal amyloid progression
To estimate the validity of the cross-sectionally estimated amyloid progression model, we studied regional amyloid progression using 2-year follow-up imaging data. First, in order to validate the crosssectional in-vivo staging approach, we determined the in-vivo amyloid stages for the follow-up dataset and assessed how many subjects progressed, remained stable, or displayed unexpected progression patterns (reverse stages or progression along deviating patterns).
In addition to the longitudinal staging analyses, we also explored the temporal progression of regional amyloid-positivity independently of the cross-sectionally estimated staging scheme. For this we used a probabilistic model that starts with the longitudinal detection of earliest amyloid accumulation in cases that are amyloid-negative at baseline, and then iteratively searches for individuals at increasing amyloid progression phases in the baseline data to record their regional progression over follow-up. Thus, based on the longitudinal rVOI changes in the subset of subjects that displayed no suprathreshold signal in any of the rVOI at baseline, we defined a regional profile corresponding to the first signs of amyloid pathology deposition. This profile was then used to select another subset of subjects who already showed the corresponding rVOI positivity profile at baseline. Assuming that these subjects were at the second year of the progression from an allnegative rVOI state, we analyzed their follow-up data to define a profile for an estimated fourth year of amyloid pathology progression. In total,

| Statistical analyses
Cross-sectional associations of in-vivo amyloid stages with baseline CSF AD biomarkers and cognitive test scores were assessed using age-corrected Spearman rank correlation coefficients (ρ). In addition, for each amyloid stage differences in baseline CSF biomarker values were compared to its preceding stage using Mann-Whitney U tests.
Similarly, in the longitudinal staging analysis Mann-Whitney U tests were used to compare baseline CSF biomarker values between groups that remained stable in their respective stage and those who progressed to a different stage at follow-up. Differences in the distribution of baseline in-vivo amyloid stages between clinical groups were evaluated using Pearson's Chi-squared test (χ 2 ).

| RESULTS
3.1 | Updated cross-sectional staging model of regional amyloid progression As expected, regional means and SDs of SUVR values in the YC data differed considerably between rVOIs, resulting in variable region-specific amyloid-positivity cut-offs (Supplemental Figure 1). Applying these cut-offs to the regional SUVRs in the CN cohort resulted in a frequency-based progression model (Figure 1 and Supplemental Figure 2) where most frequent involvement was observed for the temporal pole (67%), anterior parts of the middle temporal (62%) and parahippocampal gyri (56%), and the planum polare (51%) (stage I).
The anterior temporal lobe regions were followed by widespread frontal, temporal, and parietal association regions (40-20%), including for example, the anterior inferior temporal gyrus (39%), anterior (32%) and posterior (23%) cingulate cortex, and precuneus (28%) (stage II). A third stage involved posterior temporobasal and medial temporal areas, as well as most parts of the occipital cortex and the postcentral gyrus (20%-10%), whereas the final stage was characterized by amyloid deposition in the precentral cortex, hippocampus, and striatum (10-0%). Regional progression models estimated in subsamples stratified by demographic (age, gender, education) or genetic (APOE4) variables were all highly similar to the original model calculated in the whole CN group (all pair-wise Spearman rank correlation coefficients ρ > .96, p < .001), indicating little effect of these variables on the regional distribution of amyloid-positivity (Supplemental Figure 3). Moreover, regional progression models calculated for varying cut-off definitions all showed a highly similar ordering of the 52 brain regions (all pair-wise Spearman rank correlation coefficients ρ > .96, p < .001; Supplemental Figure 4).
Amyloid-positivity in the four anatomical divisions defining the cross-sectional staging scheme showed a highly consistent regional hierarchy across individual amyloid deposition profiles (Figure 2). Only 1% (2 individuals) of the amyloid-positive subjects in the analyzed sample and on average only 2 ± 4% of the CN subjects in the sensitivity analyses displayed non-stageable patterns. Sensitivity analyses altering the total number of considered anatomical divisions and the methodology to define positivity for each anatomic division indicated that the a priori settings (4 divisions, 50% positivity requirement) offer a good trade-off for sensitive detection of regional amyloid-positivity in a relatively detailed staging model while keeping the number of unstageable subjects as low as possible (Supplemental Figure 5).
F I G U R E 1 Four-stage model of regional amyloid progression. Brain renderings on the left illustrate the frequency of regional amyloidpositivity across individuals on a color scale from black/blue (lowest) to yellow/red (highest). In the resulting four-stage model of regional amyloid progression (I-IV), incremental stages are defined by involvement of higher numbered anatomic divisions (in red), in addition to the affected areas of the previous stage (blue) F I G U R E 2 Staging of individual amyloid burden. Six hundred and sixty-seven study participants are plotted as rows of the matrix, anatomical divisions are plotted as columns. Absence or presence of amyloid is displayed in gray and red, and two subjects that deviated from the predicted pattern are displayed in yellow (1 CN and 1 MCI) Stage I subjects were mostly classified as amyloid-negative according to standard global cortical SUVRs but displayed significantly lower CSF-Aβ 1-42 values as compared to stage 0 (−6.6%, p < .001) (Table 2), and overall, higher amyloid stages were strongly associated with lower CSF-Aβ 1-42 levels (ρ = −.72, p < .001). Additionally, higher amyloid stages were associated with higher t-tau (ρ = .52, p < .001) and p-tau (ρ = .56, p < .001) levels. In pairwise comparisons, both t-tau and p-tau levels were significantly higher in stage II compared to stage I and stage 0, but did not differ between stage 0 and stage I (Table 2). No significant CSF-based amyloid or tau biomarker differences were found between stage III and stage IV, probably reflecting saturation effects in these measures (Jack Jr. et al., 2012;Mattsson et al., 2015).
Distribution of in-vivo amyloid stages differed significantly between clinical groups (χ2 = 117.66, p < .001; Figure 3). Advanced amyloid stages III and IV accounted for 86% of amyloid deposition profiles in AD, 45% in MCI, and 20% in the CN group, with only a single CN subject classified as stage IV. Within separate clinical groups, higher amyloid stages (I-IV) were associated with lower DR scores in the CN, MCI, and AD groups, as well as with higher TMT-B scores and lower MMSE scores in the MCI group (Table 3). Similar associations were observed for standard global cortical SUVR values (> 1.1), although associations with DR in the CN group did not reach statistical significance.

| Longitudinal staging of individual amyloid progression
Out of 436 subjects that had 2-year follow-up imaging data available, 73.4% of subjects remained at the same stage after two years, 18.3% showed a model-conform transition to a higher stage, 8.3% decreased to a lower stage, and no subjects followed a regional progression pat-

| DISCUSSION
In the present study we have reassessed our previously developed invivo amyloid staging model  by applying new methodology that accounts for regionally varying noise levels in the Florbetapir-PET imaging data. Additionally, we evaluated the validity of the cross-sectionally estimated amyloid progression model using longitudinal PET data. Similarly to the previous model  and to the established neuropathologic amyloid staging schemes (Braak et al., 2011;Braak & Braak, 1991;Thal et al., 2002), longitudinal modeling approach lied on the regional order of amyloid progression as indicated by the occurrence of suprathreshold SUVR signal (i.e., regional amyloid-positivity), which is the metric used by the regional staging approach. We note that these two metrics, occurrence of suprathreshold SUVR signal and SUVR accumulation rates, provide complementary information on the regional onset of pathology and its further temporal dynamics, respectively. Interestingly, a comparison of the evolution of mean regional SUVR values with the probabilistic information on regional amyloid-positivity in our longitudinal probabilistic model suggests that the observed regional onset of pathology does not coincide with fastest progression in amyloid accumulation ( Figure 5). Thus, SUVR values of the temporal regions identified to exhibit the first suprathreshold signal displayed only a rather moderate increase over the modeled time intervals, and were surpassed at later time points by the higher SUVR increases in AD-typical frontoparietal regions that were also consistently identified as fast accumulating regions in previous longitudinal amyloid-PET studies (Bilgel et al., 2016;Marinescu et al., 2019;Mattsson et al., 2019;Villain et al., 2012). Gonneaud et al. (2017) suggested that the early signal increase in the temporal neocortex may be attributable to "physiologic" agerelated amyloid accumulation that occurs independently from ADrelated changes in the anterior and posterior midline regions. However, these early signal elevations could also reflect some form of (age-related) off-target binding, or a confound by blood flow and radiotracer delivery effects as previously described for settings with low global amyloid levels (Sojkova et al., 2015). Here we find that the regionally restricted temporal lobe signal defining stage I in our study corresponds to lower CSF-Aβ 1-42 values (but not different CSF tau values) as compared to stage 0, which suggests that this signal at least partly represents amyloid accumulation. The initial signal increase in anterior temporal lobe regions is followed by amyloid-positivity in several neocortical association areas that have previously been linked to early AD-related amyloid accumulation (together defining stage II of our staging scheme) (Gonneaud et al., 2017;Mattsson et al., 2019;Palmqvist et al., 2017;Villeneuve et al., 2015). Interestingly, we also found that, in addition to a further decrease in CSF-Aβ values, involvement of these stage II regions was associated with a significant increase in tau biomarker values, thus corroborating their closer link to AD-related pathology. Nonetheless, the exact nature of early signal increases in the temporal neocortex and their significance for AD-related pathologic changes remains to be studied in more detail. To date all amyloid-PET studies reporting early signal increases in the temporal neocortex employed 18 F-based amyloid radiotracers (Cho et al., 2016;Gonneaud et al., 2017;Rodrigue et al., 2012), and thus studying these regional signal elevations using 11 C-based PIB-PET may provide additional information as to the tracer-specificity of this finding. For instance, tracers may differ in their binding characteristics to specific types of amyloid-β aggregates, potentially also resulting in differential binding affinities to the early "diffuse" and later "neuritic" forms of amyloid plaques Ni, Gillberg, Bergfors, Marutle, & Nordberg, 2013;Seo et al., 2017).
In the longitudinal staging analysis, none of the progressed individuals deviated from the regional progression pattern predicted by  values . Thus, mean global SUVR changes in our study are well in the range of this previous study, and indicate increased global cortical amyloid accumulation from stage I onwards.
The data-driven probabilistic longitudinal trajectory model revealed a pattern of regional amyloid spread that was highly consistent with the pattern estimated by regional frequency counts in the crosssectional staging scheme, thus supporting the validity of this staging approach as a method for estimating an individual's pathology progression on the basis of a single cross-sectional PET scan. This method may be particularly useful for risk stratification in preclinical cohorts, where the utility of clinical data for staging disease progression is very limited. Importantly, increasing amyloid stages were also associated with lower memory scores among cognitively normal individuals, thus indicating the potential of this method for estimating risk of disease progression towards clinically relevant stages during the preclinical phase of AD. Recent studies of the topography of amyloid progression suggest that amyloid spread to subcortical regions is associated with higher risk of cognitive decline among nondemented individuals (Cho et al., 2018;Hanseeuw et al., 2018). Hanseeuw et al. (2018) suggested to stratify subjects into three stages according to their cortical and striatal radiotracer uptake values (low cortical, high cortical, and high cortical with high striatal signal), and showed that high neocortical and striatal signal predicted faster cognitive decline than elevated cortical signal alone. A similar staging approach was employed by Cho et al. (2018), who observed cognitive changes over a decade-long period and reported that involvement of subcortical regions (amygdala and striatal regions) was associated with the steepest longitudinal changes in cognitive measures. Our data-driven staging system very similarly places striatum and other subcortical structures to the latest stage of regional amyloid accumulation but additionally offers a more detailed subdivision of the preceding cortical deposition stage. Farrell et al. (2018) suggested that earliest amyloid-related memory decline among cognitively normal individuals most closely corresponds to regional amyloid-PET signal increases in the medial and lateral parietal neocortex (corresponding to stage II of our regional staging scheme).
Consistently, using long-term clinical follow-up information of up to 6 years we could most recently demonstrate that an individual's invivo amyloid stage as defined by our four-stage model translates into a stage-proportional risk for cognitive decline, where cognitively normal individuals from stage II onwards were at a significantly increased risk for conversion to MCI compared to stage 0 (Teipel et al., 2020).

| Methodological considerations
Our regional analysis of Florbetapir-PET data from young healthy adults indicates that even in this assumedly amyloid-negative group Florbetapir-PET signal varies significantly across brain regions, thus strongly arguing for the use of region-specific amyloid-positivity cutoffs. However, the derived cut-offs in the present study are limited to a relatively small sample of young controls with available Florbetapir-PET data (Navitsky et al., 2018a), and larger samples of truly amyloidnegative individuals would be desirable for a more robust estimate of regional amyloid-positivity thresholds. In our regional amyloid staging approach, we apply partial volume effect correction to the Florbetapir-PET images, which has previously been shown to improve amyloid-PET quantification (Brendel et al., 2015;Gonzalez-Escamilla et al., 2017;Rullmann et al., 2016;Su et al., 2016), particularly by accounting for confounding effects of high nonspecific white matter binding (Matsubara et al., 2016). Nonetheless, quantitative amyloid imaging results are also affected by the choice of several other experimental variables, including type of acquisition, preprocessing techniques, and reference tissue selection (Landau et al., 2014;Schmidt et al., 2015). Thus, future methodological work should also assess potential effects of other commonly applied reference regions Su et al., 2016) or of different approaches to PET signal quantification (e.g., full kinetic modeling instead of the simplified SUVR approach [Ottoy et al., 2017]). Finally, our longitudinal modeling of amyloid progression was limited to two time points with a relatively short interval of 2 years, which previously was associated with less stable amyloid trajectories than models based on three or more time points (Landau et al., 2018). Observing baseline amyloidnegative subjects over longer intervals would result in more robust temporal amyloid progression estimates with less reliance on modeling assumptions and could also provide better insights into the longterm stability of the observed regressions in early stage I subjects.

| CONCLUSION
We have updated our previously published in-vivo amyloid staging scheme to account for regionally varying noise levels in Florbetapir-PET signal and provided evidence for the longitudinal validity of the cross-sectionally estimated amyloid progression pattern. This technique may offer a valuable tool for stratification of pathologic disease progression based on a single cross-sectional PET scan, with potentially important implications for early detection and risk assessment in the preclinical phase of AD. Further evaluation of the effects of regional amyloid progression on longitudinal dynamics of cognitive decline in at-risk individuals is warranted.

ACKNOWLEDGMENTS
The present study was funded in part by a grant from the Alzheimer