Lifetime and 10-year absolute risk of cognitive impairment associated with amyloid PET burden: a retrospective, longitudinal cohort study

Summary

Background:

A key knowledge gap in the field concerns the lifetime risk of developing cognitive impairment among cognitively unimpaired (CU) individuals with abnormal Alzheimer’s disease (AD) biomarkers. Our objective was to compute lifetime and 10-year absolute risk of cognitive impairment as a function of continuous amyloid PET.

Methods:

This was a retrospective, longitudinal cohort study from the population-based Mayo Clinic Study of Aging (Olmsted County MN). We computed lifetime and 10-year absolute risk of cognitive impairment in 5158 CU men and women age 50 plus at enrollment. The primary predictor of interest was biological AD severity, based on amyloid PET Centiloid (CL) value. Starting age, sex, and APOE ε4 carriership were also predictors. Outcomes were incident mild cognitive impairment, dementia, and death which were ascertained or estimated via Multi-State Hidden Markov modeling both in- and out-of-study.

Findings:

Lifetime risk of MCI and dementia increased monotonically with increasing CL value [p<0.001], which was the predictor with the largest effect. Lifetime risk of MCI for male CU APOE ε4 carriers at starting age 75 years was 56.2% for CL 5, 60.2% for CL 25, 71.0% for CL 50, 75.2% for CL 75, and 76.5% for CL 100. Within each CL group, and for both men and women, lifetime and 10-year absolute risk for MCI and dementia was greater for APOE ε4 carriers than non-carriers [p<0.001]. Biological AD severity was a predictor of 10-year absolute risk of MCI and dementia [p<0.001]; however, starting age [p<0.001] had a more prominent effect. The rate of incident dementia was 2 times greater among individuals who had previously left the study than those who remained in the study.

Interpretation:

Lifetime and 10-year absolute risk for MCI and dementia among currently CU individuals increase with increasing biological severity. This information should be important for risk/benefit evaluation of therapeutic interventions in the future. The high lifetime risk in participants with higher Centiloid values address academic controversies concerning risk of future impairment associated with biomarkers of Alzheimer’s disease among CU individuals. Ascertainment and modeling of out-of-study outcomes are necessary for accurate lifetime risk estimates.

Funding:

National Institutes of Health, GHR foundation, the Alexander Family

Introduction

The presence of Alzheimer’s disease neuropathologic changes (ADNC) can be detected by modern biomarkers 15 or more years prior to the onset of overt symptoms. ^1–3 However, major controversies ^4–6 center on the appropriate diagnostic label for cognitively unimpaired (CU) individuals with abnormal Alzheimer’s disease (AD) biomarkers, primarily because some proportion will never experience symptoms in their lifetime. ⁷ To date, this controversy has largely been academic, since no disease modifying treatments have yet been approved for this population and biomarker testing is therefore not currently recommended. However, ongoing trials of anti-Aβ immunotherapy (NCT04468659, NCT05026866) raise the possibility that disease modifying treatments for AD, which are now only available for symptomatic individuals ^{8, 9}, could in coming years be approved for biomarker positive CU individuals. ^10–13 When considering treatment for CU individuals, the estimated lifetime/absolute risk of developing cognitive impairment will be a key factor in weighing risks and benefits. In the future, measures of biological AD severity should play a central role in estimating the potential benefits of treatment because greater biological severity increases the likelihood of clinical progression. A major knowledge gap, however, is the limited information available about the lifetime/absolute risk of cognitive impairment for CU individuals with abnormal AD biomarkers. Essentially all information on risk of future cognitive impairment associated with biomarkers is in the form of relative risk, or risk without considering the competing risk of death ^14–17 which does not address the probability that a CU individual will ever experience symptoms in their lifetime. To address that question, estimates of lifetime/absolute risk are necessary which requires accounting for the competing risk of death. ⁷

An additional shortcoming in observational research concerns survivor bias among participants who remain in longitudinal cohort studies over many years. Potential survivor bias arises when participants discontinue further study participation. Out-of-study events (i.e. incident MCI or dementia that occur after a participant has left a study) are not captured in typical observational cohort studies and this is a significant potential source of bias.

To address these knowledge gaps, we computed lifetime and 10-year absolute risk of MCI and dementia in CU participants from the Mayo Clinic Study of Aging (MCSA). The primary predictor of interest was biological AD severity denoted by amyloid PET Centiloid values.

Methods

This was a retrospective, longitudinal cohort study from the Mayo Clinic Study of Aging (MCSA). The MCSA is a longitudinal population-based study of cognitive aging among an age and sex stratified random sample of Olmsted County, Minnesota residents with a high retention rate (Figure s1).¹⁸ The study was approved by the Mayo Clinic and the Olmsted Medical Center Institutional Review Boards. All participants provided written informed consent.

Participants

All participants were enrolled in the MCSA and were age 50 years plus at enrollment. A clinical diagnosis of mild cognitive impairment (MCI) ¹⁹, dementia, or cognitively unimpaired (CU, defined as not MCI or dementia) was determined for each participant at enrollment and for all subsequent visits (see supplement). Critiera for MCI were normal functional activities; cognitive concern expressed by a physician, informant, participant, or study nurse; cognitive impairment in one or more domain; not demented ¹⁹. DSM-IV or DSM-V criteria were used for the diagnosis of dementia. Clincal diagnoses were made blinded to biomarker information and prior clincal diagnoses. A baseline cardiovascular/metabolic conditions score was calculated for each participant. This score is the sum of the presence or absence of seven vascular-health related conditions. A higher score indicates worse cardiometabolic health²⁰ .

The primary focus of this study was estimating lifetime/absolute risk of progressing from CU to MCI or to dementia as a function of biological disease severity. However, to improve estimates of state-to-state transition rates, information from all MCSA visits was used in participants who were CU or MCI at enrollment. Information from those with and without amyloid PET studies was used in modeling.

Predictor variables

The primary predictor of interest was amyloid PET Centiloid (CL) value. ²¹ Other predictor variables were age, sex, and APOE ε4 (carrier vs non carrier). Risk was assessed with CL as a continuous variable. In addition, to illustrate effects of combinations of predictors along with amyloid PET, we selected 5 exemplar CL values: CL 5 representing a clearly normal amyloid PET value that is common in elderly persons (Figure s2); CL 25 representing a commonly used normal/abnormal threshold ²²; CL 50 representing the midpoint in the CL scale; CL 75; and CL 100 representing the upper benchmark value in the CL scale). ²¹

Procedures

Amyloid PET imaging was performed with Pittsburgh Compound B using previously described methods (see supplement). ^{23, 24}

Outcomes

Outcomes were incident MCI, dementia, and death. Participants were followed from enrollment through all MCSA visits until censoring or death, with MCI and dementia as intermediate states. Incident MCI and dementia ascertained at a study visit are termed “in-study” events. For participants with no contact at a sufficient delay from their last scheduled visit, incident dementia and death were ascertained from a semiannual review of the electronic medical record using the Rochester Epidemiology Project medical records-linkage system. ²⁵ These are termed “out-of-study” events. We highlight that while dementia can be reliably ascertained via the medical record, MCI cannot be directly ascertained out of study, but rates can be estimated as described in the following section. ²⁶ We did not attempt to distinguish AD vs non-AD dementia; thus, the dementia outcomes were “all-cause”. MCI is a syndrome, not an etiological diagnosis and to remain consistent, the same logic should apply to dementia.

Statistical Methods

Each participant’s progression through time was modeled using a Multi-State Hidden Markov model (MSHMM) with four latent states: CU, MCI, dementia, and death (Figure 1A). We used all MCSA participants who entered the study in either the CU or MCI state at age 50 or greater, because dementia rates are essentially zero before age 50. Each participant was followed until either censoring or death. MCSA diagnostic assignments are performed by a consensus panel that is blinded to information from prior visits and biomarker information. To account for potential diagnostic misclassification at any given visit, we employed a misclassification matrix (see supplement, Table s1).

Figure 1: Modelling illustration.

(A) The multistate hidden Markov model. The four latent states are denoted by boxes: alive and cognitively unimpaired, alive with MCI, alive with dementia, and deceased. The five possible transition paths are denoted by arrows: progression from cognitively unimpaired to MCI, from MCI to dementia, and from cognitively unimpaired, MCI, or dementia to death. (B) An example estimated state distribution by age relative to three amyloid PET values (5, 50, and 100) for a female APOE ε4 carrier starting from age 65 years. Estimates are based on transition rates denoted by the arrows in panel A. At any age, the proportions of the original group in the four different states add to 100%. A few individuals lived past age 100 years, thus the proportion of each centiloid group that is deceased is not exactly 100% at age 100. MCI=mild cognitive impairment.

Our MSHMM implementation is a combination of a multi-state hazard model with misclassification, which was fit using our implementation of the MSHMM based on the msm package, version 1.8.2. We augmented the optimization based on the Levenberg–Marquardt algorithm ^{27, 28} to allow for stable convergence. All computations were done using R version 4.4.1. The MSHMM directly estimates the transition rates between each of the pairs of states, e.g., CU to MCI, as a function of current age, sex, APOE ε4 carrier status, and current CL level. These rates allow the prediction of the progression path for a participant with a given starting age, clinical state, and covariate set, i.e., the probability of being in each state at subsequent ages. These absolute risks are the primary analysis outcomes. Confidence intervals for the absolute rates were computed using standard errors obtained via grouped jackknife resampling. See supplemental methods for further details.

Our primary risk analysis used amyloid PET as a continuous measure but as a supplement we also summarized lifetime/absolute risk as a function of amyloid PET binarization. A cut point of Centiloid 25 was used to label PET scans as abnormal (≥25) or normal (<25). ²² Lifetime and 10-year absolute risk were then estimated using the refitted HMM with binarized amyloid instead of continuous amyloid PET, the other predictors were unchanged.

Role of the funding sources

The National Institute on Aging, Alexander family Foundation, and GHR Foundation had no role in the study design, data collection, data analysis, data interpretation, or writing of the report.

Results

Demographics

The sample consisted of 5,158 CU and 700 MCI at enrollment. Of these 2,067 participants were CU and 265 were MCI at their first amyloid PET visit (Table 1). Overall age at enrollment was 7 years younger for CU vs MCI. MCI was slightly more prevalent in men than women. CU participants had a lower proportion of APOE ε4 carriers, slightly higher education, an average of one fewer cardiovascular/metabolic condition, and lower baseline CL values compared with MCI. Most CU participants were APOE ε4 non-carriers and had CL values under 25 (Figure s2, Table s2). Those who underwent PET had on average one additional year of education compared with the whole sample, while the cardiovascular/metabolic conditions score did not differ.

Table 1:

Characteristics of all participants and the subset with amyloid PET, by baseline clinical diagnosis

	All		Subset with amyloid PET*
	Cognitively unimpaired (n=5158)	MCI (n=700)	Cognitively unimpaired (n=2067)	MCI (n=265)
Age, years	73 (66–80)	80 (74–85)	71 (63–78)	80 (74–85)
Sex
Female	2623 (51%)	307 (44%)	999 (48%)	114 (43%)
Male	2535 (49%)	393 (56%)	1068 (52%)	151 (57%)
Education, years	14 (12–16)	12 (12–15)	15 (13–16)	13 (12–16)
APOE genotype
ε4 non-carrier	3801 (74%)	477 (68%)	1500 (73%)	171 (65%)
ε4 carrier	1357 (26%)	223 (32%)	567 (27%)	94 (35%)
Cardiovascular and metabolic conditions score	2 (1–3)	3 (2–4)	2 (1–3)	3 (2–4)
Amyloid PET, centiloid	NA	NA	14 (8–25)	26 (12–91)
Number of on-study visits
Median (IQR)	5 (2–8)	3 (2–5)	5 (3–8)	3 (2–5)
Range	1–16	1–13	1–15	1–12
Total follow- up on study, years
Median (IQR)	5·4 (1·8–10·1)	2·7 (1·2–5·5)	6·2 (2·5–10·0)	2·7 (1·2–5·6)
Range	0·0–19·5	0·0–17·3	0·0–17·8	0·0–14·7
Total follow-up including medical record review, years†
Median (IQR)	8·3 (4·1–11·4)	5·1 (2·6–8·0)	8·0 (3·8–10·5)	4·8 (2·6–7·6)
Range	0·0–19·5	0·0–18·0	0·0–17·8	0·0–16·8

Data are n (%) or median (IQR). unless otherwise specified. Data are summarised at the Mayo Clinic Study of Aging enrolment visit for all participants and the first visit with amyloid PET for that subset. MCI= mild cognitive impairment. NA= not applicable.

^*18 participants with diagnoses other than cognitively unimpaired or MCI at the first amyloid PET visit are not shown in the subset of participants with amyloid PET, but are included in the analyses. The analysis included all clinical visits for every participant, but the first amyloid PET scan was not always at the first clinical visit; these 18 participants had progressed to dementia or to a censored diagnosis at the first amyloid PET scan (appendix pp 1–2).

^†Does not include timepoints at which only the alive or deceased status was ascertained (appendix p 2).

Transition events

A total of 985 incident dementia diagnoses were recorded; 332 of these were in-study and 653 out-of-study (the out vs in study incident dementia rate was 1.97, 95% CI: 1.72, 2.24). We recorded 1,257 transitions from CU to MCI (all in-study by definition) (Table s3). A total of 2,614 deaths were recorded. Among 277 incident dementias (in-study and out-of-study) that occurred within 10 years of the last PET study, the CL value was <25 in 80 (29%) of these individuals.

Predicted state by age and Centiloid value

Figure 1B illustrates properties of the MSHMM method where groups defined by amyloid PET CL are projected through advancing age. We selected, female APOE ε4 carriers who were alive and CU at age 65 as an example demographic in Figure 1B. The proportion of this risk group who were alive and CU fell monotonically with advancing age; however, at every age the probability of remaining CU was inversely related to higher CL value. The peak frequency of those with MCI or dementia (as opposed to CU) was shifted to younger age with higher CL value. For a given CL value, the proportion of the cohort in the MCI state increased to a maximum then decreased as individuals transition from MCI to dementia or death. The same pattern is seen with dementia but with a 3–4 year offset. The likelihood of being deceased increased monotonically to slightly under 100% at age 100 while the proportion in other states approached 0.

Lifetime risk

Figure 2 illustrates lifetime risk of MCI and of dementia for a starting age of 65 by the four sex and APOE ε4 groups as a function of continuous CL values. For all groups, increasing amyloid burden is associated with increasing lifetime risk of both MCI and dementia [p<0.001], although the effect trends toward a plateau at CL values above 75. Figure 3 illustrates lifetime risk for 3 discrete CL values (5, 50, 100) by starting age, sex and APOE status. For male CU APOE ε4 carriers at starting age 75 years, lifetime risk of MCI was 56.2% for CL 5, 60.2% for CL 25, 71.0% for CL 50, 75.2% for CL 75, and 76.5% for CL 100 and lifetime risk of dementia was 31.6% for CL 5, 35.7% for CL 25, 48.6% for CL 50, 54.5% for CL 75, and 56.5% for CL 100 (Table s4). At any given CL value, female APOE ε4 carriers exhibit the highest lifetime risk for both MCI and dementia, while male non-carriers demonstrate the lowest. Within each CL group for all starting ages and for both men and women, lifetime risk of MCI or dementia was greater for APOE ε4 carriers than non-carriers [p<0.001]. Within each CL group, for all starting ages, for APOE ε4 carriers and non-carriers, lifetime risk of MCI or dementia was greater for women than men [p=0.01]. Within each CL group lifetime risk of dementia was slightly greater for younger than older starting age [p<0.001], but not MCI [p=0.94]. Lifetime dementia risk showed similar patterns as MCI but were approximately 15% to 30% lower.

Figure 2: Lifetime and 10-year absolute risks of MCI and dementia as a function of continuous amyloid PET centiloid value.

Both lifetime absolute risk plots for MCI and dementia show four lifetime risk curves for the four sex and APOE ε4 status (carrier vs non-carrier) groups given at a starting age of 65years. Both 10-year absolute risk plots for MCI and dementia show three 10-year absolute risk curves for each starting age for a female APOE ε4 carrier. Vertical dotted lines indicate centiloid values 25 and 75 for reference.

Figure 3: Lifetime and 10-year absolute risks of MCI and dementia for individuals alive and cognitively unimpaired, by PET centiloid value, sex, APOE ε4 status, and starting age.

95% Cl bars are shown. MCI=mild cognitive impairment.

The Figures and Tables illustrate exemplar predictor combinations which were selected from many possible permutations. To fully explore all possible predictor permutations, we have created an online app for Calculating Absolute Cognitive Risk version 1.0 last updated August 17, 2025 which can be accessed at https://rtools.mayo.edu/CACR/. The reliability of risk estimates is limited at lower age by few incident MCI/dementia events and at extreme upper age by the small number of individuals still alive. Updates to the app will be noted with incremental version numbers.

10-year absolute risk

CL value was a predictor of 10-year absolute risk of MCI and dementia [p<0.001]; however, starting age [p<0.001] had a more prominent effect on 10-year absolute risk (Figure 3). For this reason, we plotted 10-year absolute risk in Figure 2 by starting age rather than sex and APOE. Ten-year absolute risk of MCI among female CU APOE ε4 carriers with CL of 50 was 8.9% for starting age 65, 36.3% for starting age 75, and 69.4% for starting age 85. Similarly, 10-year absolute risk of dementia among female CU APOE ε4 carriers with CL of 50 was 2.4% for starting age 65, 19.2% for starting age 75, and 42.3% for starting age 85 (Figure 3, Table s4). Effects of APOE ε4 and sex on 10-year absolute risk of MCI and dementia were similar to those for lifetime risk.

Lifetime and absolute 10-year risk of MCI and dementia summarized as a function of binarized amyloid PET are found in Figure 4.

Figure 4: Lifetime and 10-year absolute risks of MCI and dementia for individuals alive and cognitively unimpaired, by binarised amyloid PET status, sex, APOE ε4 status, and starting age.

95% CI bars are shown. Amyloid PET positive equates to a centiloid value ≥25 and amyloid PET negative is a centiloid value <25.

Discussion

Our central findings were the following. Lifetime and 10-year absolute risk of MCI and dementia increased continuously with increasing CL value for all sex-APOE combinations. Of all predictors evaluated, amyloid PET CL value was the predictor with the largest effect for lifetime risk of both MCI and dementia. Within each CL group, and for both men and women, lifetime and 10-year absolute risk for MCI and dementia was greater for APOE ε4 carriers than non-carriers. The rate of incident dementia was 2 times greater among individuals who had previously left the study than those who remained in the study.

An unmet need for clinical trials and future clinical care is the ability to more accurately predict the risk of progression from CU to MCI or dementia. While a short prognostic time horizon (a few years) may be more impactful for individuals who are already impaired, onset of impairment at any point in the future (i.e. lifetime risk) is highly relevant to individuals who are presently unimpaired. The risk of MCI is relevant because any degree of impairment, not just dementia, represents a significant decline in quality of life. MCI also denotes the current clinical impairment threshold that is sufficient to qualify for disease modifying treatment. ²⁹ To our knowledge there is no previous literature addressing lifetime risk of incident MCI associated with AD biomarkers.

Much emphasis in the AD biomarker field has been placed on binarizing biomarkers into positive/negative. ^{14, 30} Fewer studies have assessed risk associated with biomarker severity. ^31–33 However, measures of biological disease severity are far more informative for predicting lifetime risk of MCI or dementia than a binary positive/negative determination. Comparison of lifetime risk of MCI or dementia as a function of continuous versus binarized of amyloid PET illustrates that the latter fails to capture the continuous increase in risk with increasing amyloid PET severity. In addition, a potentially serious problem with binarization is non-transportability. If the distribution of amyloid PET Centiloid values (or any AD biomarker) above and below the cut point differs between two populations, then the risk estimates are not directly transportable from one to the other. An “average” amyloid negative individual in a population with more participants near Centiloid 20 would have higher risk than in a different sample with negative values centered near 5. A fixed threshold would underestimate risk in the former relative to the latter.

Increasing risk with increasing CL value is intuitive because a CU individual with more advanced AD pathobiology should be closer in time to incident MCI. ^{34, 35} Finding that lifetime and 10-year absolute risk of MCI were greater than risk of dementia is also intuitive because everyone who develops dementia must first pass through MCI, but some individuals with MCI will not progress to dementia either due to death or to a non-progressing variant of MCI.

The lifetime risk estimates for MCI may seem high. This is likely because the statistic most clinicians are familiar with is prevalence (i.e. the proportion of living individuals with the condition of interest in a defined population in a specific time period). For example, an American Academy of Neurology evidence review identified MCI prevalence among non-demented individuals to be 6.7% for ages 60–64, 8.4% for 65–69, 10.1% for 70–74, 14.8% for 75–79, and 25.2% for 80–84. ¹⁹ However, population prevalence at a given age is a much different statistic than the probability that a person will experience MCI at any point in their lifetime. Lifetime risk accumulates over time beyond what is captured in cross-sectional prevalence in a single age range. An additional point to consider regarding validity of our results is that all transition rates are interdependent in the MSHMM. It would be unlikely that estimates of lifetime risk of MCI would be incorrect if estimates of lifetime risk of dementia are plausible. To expand on this point, it is first important to note that most individuals in the MCSA cohort fall into lower risk groups – i.e. CL values under 25 and APOE ε4 non carriers. Prior external epidemiological estimates (without biomarker or APOE ε4 information) of lifetime risk of dementia for a 75-year-old female range from 19% - 35% and for a 75-year-old male range from 10% to 27% ^{7, 36–38}. Comparing those estimates with our lifetime dementia risk estimates for female (32.5%) and male (26.5%) APOE ε4 non carriers with CL of 25 shows that our lifetime dementia risk estimates are consistent with independent epidemiological reference values. This in turn gives us confidence that our estimates for risk of MCI, although seemingly high, are valid.

Finding that for a given CL value the lifetime risk of MCI or dementia was greater for APOE ε4 carriers than non-carriers is consistent with prior evidence that APOE ε4 confers deleterious effects beyond those measured by amyloid PET severity. ³⁹ Only 85 (6%) of the 1,357 APOE ε4 carrier CU participants in our sample were homozygous. Thus, the risk estimates associated with APOE ε4 mostly reflect ε4 heterozygotes. Finding that lifetime and 10-year absolute risk of MCI and dementia were higher for women than men overall and the APOE ε4 effect was slightly more pronounced in women, is also consistent with prior evidence. ⁴⁰

Finding that the effect of starting age on lifetime risk was negligible, while in contrast, the 10-year absolute risk of MCI or dementia was much greater at older starting ages may on the surface seem contradictory. An explanation may be that while rates of incident MCI, incident dementia, and mortality increase with age, these competing risks tend to cancel each other with a wide time window (lifetime). Older CU individuals have a higher annual rate of progressing to MCI or dementia than younger individuals but have fewer remaining years in which to develop impairment. Conversely with a narrower 10-year time window, the faster rates of progression to MCI or dementia in old vs young individuals dominate absolute risk.

While higher CL values increase lifetime and 10-year risk of MCI and dementia, risk in those with low CL values is not negligible. For example, estimated lifetime risk of MCI for a 75-year-old APOE ε4 non-carrier woman with a CL value of 5 is 55%. This is consistent with the well-established fact that pathologies other than AD are common contributors to cognitive impairment. Theoretically, the risk reduction achievable by therapeutic amyloid removal could be estimated by subtracting lifetime or absolute 10-year risk in an individual with high CL from the estimates of risk in a matched covariate group (age, sex, APOE ε4) with CL of 5. This assumes, however, that removing amyloid now will reduce the risk of future impairment to that of someone who has never had high levels and thus will likely be an upper bound on the actual effect.

This study differs from prior studies evaluating the risk of progression from CU to MCI or dementia associated with AD biomarkers in several important ways. Most prior studies ^{15–17, 41} have assessed relative risk (i.e. hazard ratios), or risk without considering attrition bias or the competing risk of death, rather than lifetime/absolute risk (percent likelihood). It is the latter that gives patients information about the likelihood of experiencing MCI or dementia in their lifetime. A predictor with a high relative risk for a condition that has low population prevalence may mean a low lifetime/absolute risk. Conversely, a predictor with a moderate relative risk for a condition with high population prevalence may mean a clinically meaningful lifetime risk.

Lifetime risk of dementia has been estimated using demographic predictors ^36–38; however, to our knowledge only three prior studies have assessed absolute risk of dementia associated with AD biomarkers, specifically amyloid PET. There are several major differences from our current study, however. Brookmeyer and Abdalla ⁷ used the 2011 National Institute on Aging – Alzheimer’s Association preclinical AD stages ¹ as biomarker predictors of dementia in CU individuals. Hartz et al ⁴² estimated 5-year absolute risk (not lifetime) of dementia. Both studies binarized amyloid PET (positive/negative) rather than employing severity, neither modeled MCI as an outcome, and neither included out-of-study outcomes in the modeling.

The MSHMM method used in the present study enables modeling of rates of progression to MCI in individuals who left the study. Accounting for attrition bias may be the most important methodological aspect of this study. An unavoidable and substantial potential source of bias in all longitudinal cohort studies is drop out, even in studies with excellent retention rates. For example, the annual retention rate among CU individuals in the MCSA is 93%. But after 10 years only 48% of a starting cohort (defined by time-in-study) would remain. In our study, the observed incident dementia rate was 2 times greater out-of than in-study indicating that drop out was not random. People who find cognitive testing progressively more uncomfortable are more likely to drop out while those that remain in-study for long periods are healthier and less likely to become impaired. Studies that ascertain only in-study events thus have significant potential survivor bias due to informative censoring. ⁴³ To our knowledge, with one exception ³², all existing biomarker-clinical risk literature is based on in-study outcomes only. These may be adequate for shorter time horizons ³³ where the number of dropouts is not large, but ascertaining and modeling out-of-study outcomes is necessary for long time horizon (i.e. lifetime) risk assessment. This is a reason we focused on lifetime and 10-year rather than shorter time horizons (e.g. 5 years) in this study. The appropriate denominator of absolute/lifetime risk is all who were ever enrolled in a study, not only those who remain in-study over long periods. Assessing only in-study outcomes will underestimate risk.

A third study to estimate lifetime risk of dementia associated with biomarkers was ours. ³² A major limitation of that study however was that the methods to estimate out-of-study MCI transitions were not yet implemented. For that reason, the outcomes in our prior study ³² were limited to dementia and death, and the starting group was non-dementia (CU or MCI). Thus, in that study we were unable to answer the fundamental question that lies at the heart of much of the academic controversy in the field which hinges on the lifetime risk of progression from CU to MCI or dementia associated with AD biomarkers. ^4–6

This study had limitations. We are not aware of other studies with all the methodological features we employed which may limit replication of our results. This study can, however, serve as a template for future studies that couple more accessible plasma biomarkers ⁴⁴ (vs PET) with comprehensive capture of outcomes from electronic medical records in large health care networks. Such an approach could address issues of both biomarker accessibility and outcomes in real world settings. ⁴⁴ Our prediction model could be improved by including plasma biomarkers, tau PET ^{16, 30}, APOE ε4 gene dose, and expansion to more diverse cohorts. Residents of southeast Minnesota in the older age range have higher education and socio-economic status than US averages and are mostly white (98% in this sample). All else being equal, our model would overestimate lifetime risk associated with AD biological severity in populations with higher mortality rates or more prevalent non-AD comorbidities and underestimate this risk in populations with lower mortality rates or less prevalent non-AD comorbidities.

In summary, even with high retention rates, over long periods, substantial survivor bias (informative censoring) may enter longitudinal observational cohort studies. However, survivor biases in long term longitudinal cohort studies can be ameliorated by assessing/modeling out-of-study outcomes. The likelihood of being deceased approaches 100% at age 100 which points to the importance of accounting for the competing risk of death when evaluating the prognostic implications of AD biomarkers. Continuous amyloid PET values in combination with age, sex and APOE e4 status are useful predictors of lifetime and 10-year absolute risk of MCI and dementia in participants who are currently cognitively unimpaired. We anticipate that lifetime and 10-year absolute risk estimates from amyloid PET (or plasma surrogates) will be important in the future for clinicians when weighing risk vs benefits of therapeutic interventions in patients who are cognitively unimpaired.

Research in Context.

Evidence before this study

We searched PubMed for terms: predicting incident mild cognitive impairment with amyloid PET, predicting incident dementia with amyloid PET, lifetime risk of Alzheimer’s disease dementia, lifetime risk of mild cognitive impairment from 1/1/2010 – 5/1/2025. The search criteria above did not identify any studies with all the following features: absolute/lifetime rather than relative risk, biological AD severity as a predictor rather than positive/negative amyloid PET binarization, a population-based cohort, modeling both incident MCI and dementia as outcomes, ascertainment of all outcomes (MCI, dementia and death) in the same cohort, and perhaps most importantly, ascertainment/modeling of out-of-study events.

Added value of this study

By including all the aforementioned features we were able to provide estimates lifetime/absolute risk of MCI and dementia associated with amyloid PET. By employing continuous amyloid PET values we demonstrate that lifetime/absolute risk varies considerably among individuals who would be considered amyloid PET positive. Thus, binarizing AD biomarkers into positive/negative discards valuable predictive information and inadequately addresses lifetime/absolute risk associated with AD biological severity. We also found that the rate of dementia ascertained out of study was 2 times greater than rates ascertained in study which provides clear evidence of informative censoring. Thus, accurate estimates of lifetime risk of cognitive impairment require ascertaining or modeling events that occur in participants who leave observational studies.

Implications of all the available evidence

Lifetime and 10-year absolute risk for MCI or dementia increase with increasing AD biological severity. Lifetime risk of MCI can exceed 80% in highest risk groups. However, those at highest risk constitute a small proportion of the population, which underscores the importance of improving individual estimates of lifetime risk of MCI and dementia using biological severity as a key predictor. Estimated lifetime/absolute risk of developing cognitive impairment with biological AD severity as a key predictor should be important in future clinical care.

Data availability

The Mayo Clinic Study of Aging makes data available to qualified researchers via an online request form at (https://ras-rdrs.mayo.edu/Request/IndexRequest).