Schindler SE, Li Y, Buckles VD, Gordon BA, Benzinger TL, Wang G, Coble D, Klunk WE, Fagan AM, Holtzman DM, Bateman RJ, Morris JC, Xiong C. Predicting Symptom Onset in Sporadic Alzheimer Disease With Amyloid PET. Neurology. 2021 Nov 2;97(18):e1823-e1834. Epub 2021 Sep 9 PubMed.

Recommends

Please login to recommend the paper.

Comments

  1. This is a great step toward translating the wealth of longitudinal amyloid PET data collected for research into a prognostic tool to project possible dementia onset in currently healthy older adults.

    It provides a solid base framework to build upon, as there are many factors (cognitive/neural reserve, cardiovascular health, comorbidities, etc.) that might impact how rapidly an individual may progress from amyloid positivity to clinical symptoms. There is a great deal of potential here, and I look forward to seeing the authors expand on this framework.

    I would love to see this data transformed to the Centiloid scale for better generalizability across tracers and scanners. It would also be very exciting in the future to see how well these “amyloid time” PET-based frameworks might correspond with more accessible and affordable plasma biomarkers.

    View all comments by Michelle Farrell

  2. This paper is a most welcome contribution to the critically important goal of developing prognostic algorithms for predicting time of onset of symptoms of AD. The authors use a reasonable number (n = 236) of subjects from the longitudinal cohort of the Knight ADRC who had been followed for an average of five years, and had had more than one Aβ PiB-PET scan. Onset of symptoms was defined by CDR of 0.5 or greater.

    The authors draw on a mechanistic model of Aβ aggregation in which a lag-phase is followed by a growth phase driven by nucleation events. Superficially, this is very attractive. Alas, we now know that nucleation is a complex affair, with primary nucleation leading to a pathway of Aβ fibril elongation and secondary nucleation with a pathway toward Aβ oligomerization (Aβo). The former may result in extracellular Aβ plaques and the latter may be the cause of synaptic damage, and tau aggregation and phosphorylation. If the PiB-PET signal reflects mainly primary nucleation events leading to Aβ plaques, it could mean that valuable information is missing if the secondary nucleation events leading to toxic Aβo are not being captured. As we move forward with biofluid (CSF and plasma) analyses, changes which occur before or independent of the PET signal may prove extremely important in deriving the prognostic algorithm.

    Be that as it may, the authors define a “tipping point,” analogous to the end of the lag phase of primary nucleation, based on their assessment of the first discernible change in the PiB-PET signal [SUVR 1.2, 7 Centiloids (CL)]. Given that the current lower threshold for clinical use is somewhere between 15-20 CL, this tipping point is very low, but still is entirely plausible if it can be confirmed in further studies. The upper-level threshold (SUVR 3.0, 88 CL) is also low, given that 100 CL is by definition the mean full-blown AD dementia level. Nevertheless, the 7-88 CL range is a useful start for beginning the statistical analysis of time of onset.

    Methodologically, the paper constructs a disease-progression model based solely on Aβ PET information, called "amyloid time." The idea has intuitive appeal—indeed, it represents how long it has been since one crossed a particular threshold, while taking into account the fact that the rate of accumulation is not constant across Aβ levels.

    All studies utilizing similar approaches follow a two-step procedure. First, individual slopes are estimated based on longitudinal data (here per-subject linear regression, but mixed models have been suggested) and are assumed to correspond to the estimated burden at midpoint of follow-up. Secondly, a population-level relationship between current levels and rates of change is examined. Despite being standard, these two steps hold somewhat contradictory assumptions, the former postulating linearity (and thus constant rate of change) at the patient level, the latter studying nonlinear association that emerges once the data have been aggregated.

    Perhaps, going forward, and making use of larger longitudinal studies, a better approach would be to explicitly assume that the dynamics of Aβ accumulation are governed by an ordinary differential equation—or, more likely, a mixture thereof corresponding to whether one is a "progressor" or not, and incorporating effects of ApoE4 coming from some parametric family, and subsequently perturbed by measurement noise. The resulting model can then be fitted using probabilistic programming frameworks, e.g., Stan, Turing.jl etc.

    As far as predicting age of onset is concerned, the results look very promising. However, to be fully convincing, external validation is required, as the authors themselves acknowledge. The outcome of testing this model in a separate dataset of PiB-based Aβ measurements would be extremely important. A study of robustness with respect to various modelling choices, e.g., bins, knot positions, etc., could also be beneficial.

    As a minor point, we note that the estimate of "age at SUVR 1.2" may potentially use information unavailable at the onset of symptoms (if, say, more PET scans were done post-onset) due to the averaging procedure employed, only including observations with SUVR>1.2 that occur prior to onset would alleviate this issue. Showing distribution of "age at SUVR 1.2" estimates per individual coming from different time points is of independent interest and would help gauge the model’s consistency at the personalized level.

    Finally, the influence of genes, environment, and age, and the poorly defined concepts of resilience/vulnerability, need to be incorporated. ApoE haplotypes and current GWAS hits are primarily “age at onset” genes, of which ApoE4 is clearly the major co-dominant risk factor. But the genes that determine “rates of progression” are still poorly defined. We and others have confirmed the polymorphic BDNF gene plays a significant role in rates of progression from preclinical to prodromal stages. This type of genetic influence also needs to be incorporated in any prognostic algorithm. Environmental factors, particularly manifest through cerebrovascular co-morbidities, together with other age-related neurodegenerative conditions (such as TDP43–related FTD), also need to be factored in.

    Our AD field has indeed reached a “tipping point” in which new improved technologies of neuroimaging, biofluid analyses, genetics, and the cognitive neurosciences will all soon deliver useful prognostic algorithms.

    View all comments by Colin Masters

  3. This work adds to a growing literature examining the amount of time an individual has been amyloid+, referred herein as Amyloid Time and referred to as Amyloid Chronicity by Koscik et al., 2020.

    I love this approach. It is intuitive and provides an important dimension to explore regarding amyloid burden in our studies. In particular, can a duration measure provide additional information above and beyond simply characterizing individuals as A-/A+? Further, are there factors that modify the impact of amyloid duration? 

    The work by Schindler suggests that the impact of duration is modified by chronological age, such that individuals with a younger onset of amyloid positivity can bear that burden for longer than their older counterparts. This is consistent with the idea that older subjects may have a greater degree of comorbid pathology, thus a smaller amount of amyloid pathology may be necessary for clinical impairment. However, it is potentially at odds with early onset AD, which tends to have a faster progression. The predictive value of amyloid burden as a function of age will require larger datasets, and potentially progression outcomes that are not as focused on typical presentations of AD (for instance, the initial deficits in younger cases may be non-amnestic).

    Somewhat related to the comment above regarding early age of onset of amyloid+, I find the distribution of age of onset in Figure 2D among the APOE noncarriers very exciting. This plot highlights that a decent chunk of noncarriers became amyloid+ between age 50-55. This group in particular could be extremely insightful to discover novel drivers of amyloid deposition.

    References:

    Koscik RL, Betthauser TJ, Jonaitis EM, Allison SL, Clark LR, Hermann BP, Cody KA, Engle JW, Barnhart TE, Stone CK, Chin NA, Carlsson CM, Asthana S, Christian BT, Johnson SC. Amyloid duration is associated with preclinical cognitive decline and tau PET. Alzheimers Dement (Amst). 2020;12(1):e12007. Epub 2020 Feb 13 PubMed.

    View all comments by Elizabeth Mormino

  4. A key contribution of this paper to the literature is the individual-level assessment of the interval between the onset of elevated amyloid levels and the onset of clinical symptoms of sporadic Alzheimer’s disease. Previous studies that addressed this question relied mainly on datasets where the age at onset of elevated amyloid and the age at onset of symptoms were not both known (or inferable) for each individual. Because of this, previous studies were able to estimate population-average intervals only.

    This paper confirms the estimated intervals from these previous studies, and expands upon them by demonstrating that the interval between the onset of elevated amyloid levels and onset of clinical symptoms varies substantially across individuals (as illustrated by the left-most red dot in each row in Fig 3A). As shown by Schindler and colleagues, a factor that explains a substantial proportion of this variability is the age at onset of elevated amyloid levels. Compared to individuals who developed symptoms earlier in life, individuals who developed symptoms later also had developed amyloid later (Fig 3C and 3D). Interestingly, however, those with a later onset of elevated amyloid had a shorter time between amyloid onset and symptom onset compared to those who developed elevated amyloid earlier in life. 

    It is important to remember that not every individual who has elevated amyloid levels develops cognitive impairment or dementia. This makes it difficult to obtain unbiased estimates of the interval between onset of elevated amyloid levels and onset of symptoms. The reported interval estimates in this paper are for only the individuals who did go on to develop impairment. Because of this, the findings of this paper may not be generalizable to every individual with elevated levels of amyloid. Further analyses are necessary to provide insight into who among the amyloid accumulators might develop future cognitive impairment.

    Nevertheless, the strong correlation observed between the age at elevated amyloid onset and age at symptom onset suggests that the former might be a promising outcome measure in clinical trials targeting the preclinical stage of Alzheimer’s disease.

    View all comments by Murat Bilgel

  5. In a beautifully outlined paper, the authors conclude that at around 1.2 SUVR PiB PET, a fairly steady rate of amyloid accumulation kicks in. After that, it takes about 17 years to reach 3.0 SUVR. This steady accumulation rate toward reaching a specific amount of amyloid burden gives an interesting insight into the pathophysiology of AD.

    So does the finding that a predictable rate (correlation between SUVR and accumulation rate) exists in APOE4 carries, but not in noncarriers, below 1.2 SUVR.

    The identified cutoff at 1.2 SUVR for PiB PET could be very useful in several settings, ranging from primary prevention trials to studies on preclinical AD, if replicated (preferably using transformed cutoffs for other amyloid PET tracers). The paper is slightly technical at times, but for readers looking for raw data I recommend Fig. 2E, which captures the essence without modeling.

    Regarding the part about estimating the age of symptom onset, I am a bit hesitant about the generalizability when considering the following: 1) The diagnostic assessment, where clinicians were blinded to both amyloid PET and the assessment at other visits, must have been a challenging task and naturally contains diagnostic uncertainty. 2) These models were mostly derived from small subsamples. 3) Age at symptom onset partly depends on one’s level of cognitive reserve. The present population was much more highly educated (mean years of education of around 15-16) than the average elderly population. 4) They tended also to be somewhat younger than the typical preclinical/early symptomatic AD population (mean age 66 years) and a typical older population often has more co-pathologies that could affect the estimated time to symptom onset.

    The approach and effort are nonetheless extremely important, since these kinds of predictions of symptom onset may be very valuable when deciding whom to treat with anti-amyloid drugs in the future (i.e., preferable those with preclinical AD expected to develop symptoms within the next couple of years and not the 45-year-old example case in the paper with an amyloid PET of 1.2 SUVR who was predicted to develop symptoms in 21 years).

    View all comments by Sebastian Palmqvist

  6. There has been a lot of interest in using amyloid age as the time scale. Terry Therneau in our group recently published his algorithm based on the same general idea, see, for example, Fig. 2 in Therneau et al., 2021.

    How to line different people up on a common x axis was identified as a problem for AD disease modeling a long time ago. We discuss this in Jack et al., 2010. Schindler et al. have done a nice job with their approach, and their results are believable.

    References:

    Therneau TM, Knopman DS, Lowe VJ, Botha H, Graff-Radford J, Jones DT, Vemuri P, Mielke MM, Schwarz CG, Senjem ML, Gunter JL, Petersen RC, Jack CR Jr. Relationships between β-amyloid and tau in an elderly population: An accelerated failure time model. Neuroimage. 2021 Nov 15;242:118440. Epub 2021 Jul 29 PubMed.

    Jack CR, Knopman DS, Jagust WJ, Shaw LM, Aisen PS, Weiner MW, Petersen RC, Trojanowski JQ. Hypothetical model of dynamic biomarkers of the Alzheimer's pathological cascade. Lancet Neurol. 2010 Jan;9(1):119-28. PubMed.

    View all comments by Clifford R. Jack

  7. This study is an important addition to our field for all the reasons others have commented on so far. I want to draw attention to the “order from chaos” plot in the Alzforum article, which is from Figure 2 in the paper. Everyone who is amyloid-positive is accruing signal at roughly the same rate (within measurement error). The heterogeneity is largely in the time domain (age at which the trajectory crossed the x-axis and became positive).

    This is consistent with group observations from Villemagne et al., 2013, and several others, including recent work from Jagust and Landau, 2021, that amyloid accumulation PET signal is predictable.

    From our work in Koscik et al., 2020, on the Wisconsin Registry for Alzheimer’s Prevention, and the updates we showed at AAIC 2021, and now this important paper from Schindler et al., we can be reasonably confident not only that it's linear, but also that the rate of change is largely the same for all the A+ participants in the plot. Bilgel et al., 2016, were the first to show some of these ideas at the individual level in the BLSA cohort. As was demonstrated nicely by Schindler et al., this is going to enable individual-level estimates of age of onset from a single PET scan! This should greatly inform our understanding of the biomarker cascade and may be informative clinically.

    I also want to point out an important implication of these results. The rate of amyloid change in A+ subjects looked to be about the same for ApoE4+ and ApoE4-. Similarly, we showed at AAIC 2021 that the rate of change in A+ subjects looked the same for men and women, and for people who got on the amyloid time escalator at younger or older ages. We can reasonably hypothesize, based on the curves we are seeing, that since rate of change is impervious to these potent non-modifiable risk factors, it is likely also impervious to modifiable risk factors. This is now testable. 

    Also, as we begin to more broadly apply this class of methods and extract the temporal information from amyloid PET in cohort studies, we can begin to actually quantify the many factors that might influence cognitive decline, anchored to person-level amyloid onset age. This new capability will greatly help our field understand vulnerability and resilience to cognitive decline and conversion to clinical AD stages, anchored to amyloid time.

    References:

    Bilgel M, An Y, Zhou Y, Wong DF, Prince JL, Ferrucci L, Resnick SM. Individual estimates of age at detectable amyloid onset for risk factor assessment. Alzheimers Dement. 2016 Apr;12(4):373-9. Epub 2015 Nov 14 PubMed.

    Jagust WJ, Landau SM, Alzheimer's Disease Neuroimaging Initiative. Temporal Dynamics of β-Amyloid Accumulation in Aging and Alzheimer Disease. Neurology. 2021 Mar 2;96(9):e1347-e1357. Epub 2021 Jan 6 PubMed.

    Koscik RL, Betthauser TJ, Jonaitis EM, Allison SL, Clark LR, Hermann BP, Cody KA, Engle JW, Barnhart TE, Stone CK, Chin NA, Carlsson CM, Asthana S, Christian BT, Johnson SC. Amyloid duration is associated with preclinical cognitive decline and tau PET. Alzheimers Dement (Amst). 2020;12(1):e12007. Epub 2020 Feb 13 PubMed.

    Villemagne VL, Burnham S, Bourgeat P, Brown B, Ellis KA, Salvado O, Szoeke C, Macaulay SL, Martins R, Maruff P, Ames D, Rowe CC, Masters CL, Australian Imaging Biomarkers and Lifestyle (AIBL) Research Group. Amyloid β deposition, neurodegeneration, and cognitive decline in sporadic Alzheimer's disease: a prospective cohort study. Lancet Neurol. 2013 Apr;12(4):357-67. Epub 2013 Mar 8 PubMed.

    View all comments by Sterling Johnson

Make a Comment

To make a comment you must login or register.

This paper appears in the following:

News

  1. Can a Single Amyloid PET Scan Predict Time to Symptom Onset?