Time between milestone events in the Alzheimer’s disease amyloid cascade

Objective Estimate the time-course of the spread of key pathological markers and the onset of cognitive dysfunction in Alzheimer’s disease. Methods In a cohort of 336 older adults, ranging in cognitive functioning, we estimated the time of initial changes of Aβ, tau, and decreases in cognition with respect to the time of Aβ-positivity. Results Small effect sizes of change in CSF Aβ42 and regional Aβ PET were estimated to occur several decades before Aβ-positivity. Increases in CSF tau occurred 11-12 years before Aβ-positivity. Temporoparietal tau PET showed increases 4-5 years before Aβ-positivity. Subtle cognitive dysfunction was observed 7-9 years before Aβ-positivity. Conclusions Increases in tau and cognitive dysfunction occur years before the presence of significant Aβ. Explicit estimates of the time for these events provide a clearer picture of the time course of the amyloid cascade and identify potential windows for specific treatments.


Introduction
Disconcerting clinical trial results for the treatment of Alzheimer's disease (AD) have led to a shift toward earlier intervention, focusing on the early clinical or presymptomatic phases, when biomarkers are needed to identify the disease. The amyloid cascade (Hardy and Selkoe, 2002) is thought to start with elevated levels of two key amyloids in the brain, -amyloid (A) and tau, and end with severe cognitive and functional impairment (Jack et al., 2010). Growing evidence suggests that an early sign that the cascade has begun is change in cerebrospinal fluid (CSF) A, potentially detectable prior to significant A deposition in the brain as measured by positron emission tomography (PET) (Palmqvist et al., 2016). This accumulation of A has been suggested to be followed by increases in CSF tau and the spread of tau pathology beyond the temporal lobe (Braak and Braak, 1991;Schöll et al., 2016). The build-up and spread of these two brain pathologies is paralleled by gradual cognitive and functional decline (Zetterberg and Mattsson, 2014).
Previous neuropathological and biomarker data suggest that the overall time course of AD is several decades (Li et al., 2017;Villemagne et al., 2013). In autosomal dominant AD, the estimated years to clinical onset has been used to estimate the timecourse of different biomarkers in AD (Bateman et al., 2012). However, the time-course of the spread of A and tau and the onset of clinical symptoms in sporadic AD is unknown.
With repeated measures of A over time, the level and rate of change with respect to the key initiating AD pathology may offer a measure of disease progression in sporadic AD.
With level and change information, the time from the threshold for significant A pathology can be estimated within individuals, providing the temporal disease progression information important for evaluating biomarker trajectories. Without longitudinal information, cross-sectional studies frequently categorize subjects into two groupsthose below a threshold for significant pathology and those above, where subjects just below the threshold who will cross over within months are considered pathologically equivalent to subjects who will not cross over for decades. By incorporating longitudinal information, disease progression with respect to A pathology can be represented to reflect its continuous nature, resulting in a more powerful way to model the relationship between A and downstream processes.
The aim of this study was to evaluate time-from-A-positivity (TFA+) in sporadic AD. Using serial 18F-florbetapir (A) PET measurements, rates of change of A were estimated and used to calculate the time-from-threshold for each subject. These subject-specific estimates of the proximity to the threshold for A-positivity (A+) were then used to model the trajectories and temporal ordering of other key markers in AD including CSF A42, regional A PET, several measures of tau including CSF phosphorylated (P-tau) and total tau (T-tau), regional 18F-flortaucipir (AV-1451) tau PET, and cognition. Estimates of the time and ordering of these pathophysiological changes may facilitate the design of future prevention trials and identify a window for early treatment.

Standard protocol approvals, registrations, and patient consents.
This study was approved by the Institutional Review Boards of all of the participating institutions. Informed written consent was obtained from all participants at each site.

Participants
Data were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (http://adni.loni.usc.edu/, www.adni-info.org) on 1/21/2020. An initial analysis was done on all ADNI participants with available A PET data (in N=963 CU, A+ MCI and A+ AD), to facilitate the estimation of TFA+. The population in the primary analysis only included ADNI participants with measurements of both A and tau PET. Of these, all cognitively unimpaired (CU), prodromal AD (A+ MCI) and A+ AD dementia participants were included in the analysis, where A-positivity was defined using a previously established threshold (Standardized Uptake Value Ratio, SUVR = 1.10) . A-MCI (N=224, including A-CU to MCI progressors) and A-"AD dementia" subjects (N=51, including A-MCI to AD dementia progressors; we consider these to be misdiagnosed, because we assume AD requires A+) were not included in the main analysis given our aim to model disease progression over the AD continuum and not other diseases, but visualizations of their biomarker data are included

Statistical Analysis
The aims of these analyses were to evaluate the relationship between the estimated TFA+ and CSF, PET, and cognitive responses. Because TFA+ was not directly observed, in a first step, linear mixed-effects models were fit to all available longitudinal global A PET SUVR data to estimate subject-specific intercepts and slopes of A pathology. Because A slopes are unlikely to remain constant over long periods of time as subjects move toward and away from the A threshold, natural splines (Hastie and Tibshirani, 1990) were used to estimate the nonlinear shape of the slopes with respect to baseline A, using quantile regression. Rather than modeling the mean A slope with respect to baseline A, quantile regression provides a separate curve for each quantile, allowing the relationship between slope and intercept to differ depending on the location in the distribution of A slope. For each subject, TFA+ was estimated by integrating over each subject's quantile curve between the subject's intercept and the threshold for A-positivity (PET SUVR = 1.1). For example, for a subject with a baseline SUVR of 1.2 and a slope in the 0.6 quantile, TFA+ was taken to be the time it would take to go from SUVR = 1.1 to 1.2, using the slope estimates from the quantile curve. For incremental changes on the x-axis (baseline SUVR), the time required to travel the incremental distance is equal to distance/rate. Using the trapezoid rule (Atkinson, 1989), TFA+ is the sum of these incremental times spanning SUVR = 1.1 to 1.2. An example of calculating TFA+ is given in the top left panel of Figure 1.
To evaluate the accuracy of the TFA+ estimates, we compared the observed times of A+ to the estimated times of A+ values in participants who were A-at baseline and became A+ during follow-up. Observed time of A+ occurred in the interval between the last A-scan and the first A+ scan. The observed time was calculated as a weighted average of the two scan times, weighted proportionally toward the scan where the participant was closest to hitting the threshold. Observed and estimated values were compared in N=37 participants who crossed the threshold for A+ and remained A+ throughout follow-up.
Our analyses aim to model participants who are ostensibly on the AD trajectory and had calculable TFA+. Therefore, of the 963 participants with A PET, we excluded N=16 participants with negative A accumulation rates (negative rates were largely driven by one early high A PET measure), we also excluded N=6 participants with low levels of A and accumulation rates such that they were predicted to become A+ later than 120 years of age (biomarker data from these subjects are included for visual comparisons in Figures 2-4, see Figure legends). We included subjects where the TFA+ metric indicated very early accumulation of A, but for participants estimated to have become A+ before age 40 (N=24, median estimated age at A+ = 30, IQR: 24 to 34), we truncated TFA+ to age 40, based on previously described rates of A-positivity in middle age (Jansen et al., 2015).
In the second step, the relationship between TFA+ and the responses was modeled using monotone penalized regression splines. Generalized cross-validation was used to tune the smoothing parameter (Wood, 1994). Cognitive responses were covaried for age, gender and education; CSF A42, T-tau, P-tau and PET measures were covaried for age and gender.
In order to account for the uncertainty across steps 1 and 2, the entire process was repeated in 500 bootstrap samples to estimate 95% confidence intervals for the association between TFA+ and the responses.
Meaningful effect sizes of change of increase in pathology or decrease in cognition with respect to TFA+ were estimated. A Cohen's d effect size of 0.2 SD was considered small, 0.5 SD was considered medium, and a 0.8 SD effect was considered large (Cohen, 1988). A 0.2 standard deviation (SD) change from the mean response at the longest times (least pathological) from A-positivity was taken to be the initial point of meaningful change. A 0.5 SD change was also shown as a more substantial effect size of change. We also estimated change, 95% confidence intervals, and statistical significance of change for each response at TFA+ = 0, the time of A-positivity, with bootstrapestimated standard errors.
Baseline associations between demographics and TFA+ were assessed using Spearman correlation for age and education and the Wilcoxon rank-sum test for gender.
Associations between diagnosis and demographics were assessed using Wilcoxon ranksum test for continuous variables and Fisher's Exact test for categorical variables. All analyses were done in R v3.5.1 (www.r-project.org).

Cohort Characteristics
Two-hundred and twenty-eight CU (128 A-and 100 A+), 70 A+ MCI and 38 A+ AD participants were included in the analysis. The diagnostic groups varied by

A PET and Estimation of TFA+
TFA+ was estimated with a median of 3 (range: 1 to 5) A PET scans per participant. The average time between first and last scan was 3.3 years (SD=2.9) and the average time between scans was 2.2 years (SD=0.8). Across diagnoses, TFA+ ranged from -35.9 to 47.0 years, where higher (positive) TFA+ values indicate more time spent with a significant A burden. The average TFA+ was -9.3 years (SD=6.9) for CU-, 13.9 years (SD=11.2) for CU+, 21.1 years (11.7) for MCI, and 25.8 years (11.0) for AD participants (p<0.001, comparing CU+, MCI, and AD only). TFA+ was highly correlated with observed time of A+ (=0.93, p<0.001, bottom left panel of Figure 1).
TFA+ was not associated with sex (mean TFA+ = 9.5 (SD=17.4) and 6.6 (SD=16.8) in males and females, respectively, p=0.13). Within-diagnosis TFA+ distributions are shown on the bottom right panel of Figure 1. Quantile curves of the relationship between A intercepts and slopes are also shown in the top right panel of Figure 1, displaying the variation of acceleration of A deposition over different levels of baseline A.  Table 1 summarizes the values of the responses at the longest times before A+, i.e. the least pathological TFA+. Table 1 Table 1.

Regional A PET
Cognitive measures are shown in Figure 5. The MMSE showed a 0.2 SD drop nine years before A-positivity, followed by the PACC seven years before A-positivity.

Discussion
Several biological processes develop over time in sporadic AD, including accumulation of A and tau across wide areas of the brain, as well as cognitive decline.
Based on the amyloid cascade hypothesis, a relevant overarching time scale of the disease processes could be based on the development of A pathology (Koscik et al., 2020). We have therefore integrated A PET level and rate of change information to place each individual on a pathological timeline. This timeline can then be used to estimate the time of downstream events in the amyloid cascade. We estimated several major milestone events of AD progression including a small drop in CSF A42 35 years before Apositivity and a small increase in regional A PET deposition 17 years before Apositivity. Using the biomarkers tested here, the first changes in CSF A42 may define the onset of AD. Small increases in tau pathology were estimated to occur 11-12 years before A-positivity, as measured by CSF and 5 years before, as measured by PET. More substantial and statistically significant increases in CSF as well as temporoparietal tau PET were detected by the time of A-positivity. Small effects of cognitive dysfunction occurred 7-9 years before A-positivity, coinciding with previous reports (Insel et al., 2017). These findings provide a general time scale for initial changes in sporadic AD, which may inform clinical trials aimed at specific stages of the disease.
A 0.2 SD difference, a small, but meaningful increase in levels of CSF tau and temporoparietal lobe tau are observed years before the current threshold for Apositivity. In the context of secondary prevention trials where A-positivity at current thresholds is required for study inclusion, tau levels in these participants would already have been increasing for several years, likely more. The finding that temporoparietal tau starts to increase prior to other regions is in accordance with 18F-flortaucipir studies on other populations. Cross-sectional studies showed early tau deposition in cognitively healthy elderly (with or without significant A pathology) in temporal and medial parietal regions, most dominant in entorhinal and parahippocampal cortex, the amygdala and inferior temporal cortex. Longitudinal studies further suggest that cognitively healthy elderly accumulate tau in the medial temporal and medial parietal lobe, while (A positive) AD dementia patients increased in tau primarily in the frontal lobe (Harrison et al., 2018). The spread of tau beyond the MTL to the parietal lobe and other regions may be a critical milestone in the progression of AD. The early changes observed in the MPL in this study coincide with a recent report of the earliest tau deposition found in medial parietal regions (precuneus and isthmus cingulate) in autosomal dominant AD (Gordon et al., 2019). Considering that a 0.2 SD increase in MPL tau can potentially be detected several years before A-positivity (Figure 4), these data support the use of primary prevention trials against A where treatment is initiated years before the current threshold for A-positivity, if treatment efficacy relies on early intervention, prior to the development of tau pathology.
The initial descent in cognitive performance is estimated to occur 7-9 years before becoming A+ ( Figure 5). Reduced cognitive performance has repeatedly been shown to be associated with elevated levels of A Donohue et al., 2017;Insel et al., 2017Insel et al., , 2016, even within the subthreshold range (Landau et al., 2018), in cognitively unimpaired individuals. The result that CSF tau measures started to change between regional A and cognition in this study is in accordance with the theory that cognitive impairment in AD is caused primarily by tau pathology. This is also in line with other recent studies which show that cognitive impairment is more strongly related to accumulation of tau than to A (Ossenkoppele et al., 2019), and that both tau and A appear necessary for cognitive decline (Sperling et al., 2019). The ordering of the responses coincides with the magnitude of the effect sizes at the time of A-positivity (Table 1), suggesting that initial changes in the responses continue to change in parallel through to the time of A-positivity, without any major differences in acceleration.
In their 2018 draft guidance, the FDA indicated that because it is highly desirable to intervene as early as possible in AD, it follows that patients with characteristic pathophysiologic changes of AD but no subjective complaint, functional impairment, or detectable abnormalities on sensitive neuropsychological measures are an important target for clinical trials (Food and Drug Administration, 2018). If the spread of tau to the lateral temporal and parietal lobes becomes a defining characteristic of pathophysiological change in AD, the window to intervene as early as possible may shift to years before the current threshold for A-positivity. It is possible that early accelerations of tau may have contributed to recent failures of anti-A treatments in phase III clinical trials on A-positive patients (Egan et al., 2018;Honig et al., 2018).
Although selecting subjects that are A-positive ensures that only AD patients are included in trials, the use of conservative thresholds to define A-positivity may bias trial populations toward individuals where tau pathology has already accumulated, causing downstream injuries independent of A, reducing the efficacy of anti-A treatments.
This study has several limitations. Tau PET data were available for only a subsample of the data, limiting comparisons to a small cross-section of the full ADNI data set. More data, especially longitudinal data in participants in the earliest stages of A changes, will be required for more precise change point estimates. These analyses lack the power and precision to place the temporal and parietal tau regions in a particular order with confidence, but instead demonstrate that temporoparietal tau increases years before A-positivity. The ADNI CU, MCI and AD cohorts are also age matched. The AD patients, on average, have dementia by age 75, while the participants in the CU cohort who may eventually develop AD, are unlikely to do so for many years, possibly decades.
By design, these cohorts with age matched groups are therefore on systematically different disease trajectories with respect to age. If earlier onset is associated with a more aggressive form of the disease, then the AD cohort may have the most aggressive form while the CU cohort, the least aggressive. If the developing A pathology in the ADNI CU-cohort represents a less aggressive disease process compared with a more typical AD process, the estimates reported here could be conservative and biased toward later time estimates for downstream events. The ADNI MCI cohort may represent a more typical trajectory with respect to downstream events along the A pathological timeline.
These differences in disease trajectories are apparent from the cohort estimates in  (Janelidze et al., 2020;Mielke et al., 2018;Palmqvist et al., 2019;Schindler et al., 2019), longitudinal evaluations in the context of trial-ready cohorts may greatly improve early diagnosis and expedite the execution of clinical trials in early AD.

Densities
Top left panel: an example of how TFA+ is estimated. Here we have a participant with an estimated intercept of SUVR = 1.00 and an accumulation rate (slope) of 0.005 SUVR/year. We want to calculate how long it will take for them to reach the 1.10 threshold. A slope of 0.005 SUVR/year puts this participant on the 0.20 quantile (20% percentile) curve. We know this participant must accumulate 0.10 SUVR to reach the threshold and we will assume they will continue to have an accumulation rate in the 0.20 quantile. Partitioning the curve into segments from SUVR = 1.00 to 1.10 and using the formula time = distance/rate, the time to cross each segment is calculated and summed. In the figure, only two segments are shown, but in the actual calculation, the curve is partitioned into a large number of segments. Assuming a linear rate increase within each segment (shown in the dashed black line along the red quantile curve), the time to travel the distance in the 1st segment, from SUVR 1.00 to 1.05 is given by, time1 = d1 /rate1, where d1 is 0.05 and rate1 is the average rate in segment 1, which is ½(h0+h1), as shown in     Regression curves (red) and corresponding 95% CIs (shaded grey) are shown. Mean values of the response are plotted against mean TFA+ for each of the four diagnosis groups (large symbols). The 0.05 quantile (approximately -1.65 SD if normally distributed) of the response for the CU-group is also shown (short/long dashed line).

Figure 6. Summary Curves
Summary curves are shown for all modalities on a scale from zero to one. Responses are scaled such that zero is the least pathological point for each response and one is the mean response in the AD participants. The initial effect, defined by 0.2 SD change points are plotted.