Amyloid and tau PET-positive cognitively unimpaired individuals are at high risk for future cognitive decline

1Alzheimer’s disease (AD) is neuropathologically characterized by the presence of amyloid-β (Αβ) plaques and tau neurofibrillary tangles. Although the two most well-established diagnostic criteria for AD both acknowledge the importance of Αβ and tau pathology in AD pathogenesis, an important distinction is that the National Institute on Aging and Alzheimer’s Association (NIA-AA) criteria define AD by its biological features (that is, the presence of Αβ and tau pathology) irrespective of the clinical syndrome, whereas the International Working Group (IWG) criteria require the presence of objective cognitive impairment (mild cognitive impairment (MCI) or dementia) in conjunction with positive AD biomarkers. Consequently, there is fundamental disagreement between the criteria about the nomenclature for cognitively unimpaired individuals who harbor one or both AD hallmark neuropathological features. For example, a cognitively unimpaired individual with positive Αβ (Α+) and tau (T+) biomarkers is classified as ‘preclinical AD’ by the NIA-AA criteria, while the IWG criteria would label such an individual ‘at risk for progression to AD’ (Fig. 1).

2Aside from philosophical differences (for example, according to the IWG criteria the term AD should be restricted to symptomatic individuals), the discrepancy between the NIA-AA and IWG can be explained by at least two factors. First, according to the IWG criteria, currently available Αβ and tau biomarkers show ‘low predictive accuracy’ for development of cognitive symptoms. For Αβ biomarkers alone this may indeed be the case,, although PET studies with long follow-ups (approximately 10 years) indicate substantial Αβ-associated risk for cognitive decline, dementia and death. For tau biomarkers, however, the clinicopathological correlates are much stronger than for Αβ biomarkers,, even in asymptomatic individuals,. Second, most prognostic studies to date are based on cerebrospinal fluid (CSF) biomarkers of soluble phospho-tau levels, which is an early marker of AD pathology,. In contrast, the more recently introduced tau PET technique measures more advanced pathological changes since it captures insoluble tau aggregates,. Although tau PET positivity in the neocortex is relatively rare among cognitively unimpaired individuals (approximately 5–10%–), its presence is strongly associated with worse cross-sectional,– and longitudinal,– cognitive outcomes. However, large-scale longitudinal studies with combinations of amyloid PET and tau PET biomarker profiles as predictors of clinical progression among cognitively unimpaired individuals are lacking.

3In addition to addressing the differences between the NIA-AA and IWG criteria, early detection of AD pathological changes may be key for future interventions with disease-modifying treatments since these will most likely be cost-effective and show the most favorable benefit versus risk ratio when specifically targeting preclinical populations with AD that are most likely to experience substantial cognitive deterioration in the short term (that is, 3–5 years). Therefore, the aim of the current multicenter study was to examine clinical progression to MCI or dementia and assess cognitive decline in cognitively unimpaired individuals with different Αβ (A) and tau (T) biomarker profiles as defined by PET at baseline. We divided A+T+ individuals into medial temporal lobe (MTL) only (A+TMTL+) and temporal neocortical (A+TNEO-T+) T+ groups to additionally investigate the impact of more advanced tau pathological changes on clinical progression.

4We included 1,325 cognitively unimpaired participants from 7 cohorts, of whom 843 (63.6%) were A−T−, 328 (24.8%) A+T−, 55 (4.2%) A+TMTL+ and 65 (4.9%) A+TNEO-T+ (see Table 1 and Supplementary Table 1 for a breakdown by cohort). All biomarker-positive groups were older and had lower baseline Mini-Mental State Examination (MMSE) scores compared to the A−T− group (all P < 0.001). There were no sex differences between groups. The average follow-up duration was 41.8 ± 18.9 months. The A−T+ group was considerably smaller than the other groups (n = 34; Extended Data Table 1), hence their results are only reported in Extended Data Fig. 1.

5During the clinical follow-up, 26 out of 781 (3.3%) of A−T−, 26 out of 292 (8.9%) of A+T−, 25 out of 51 (49.0%) of A+TMTL+ and 32 out of 60 (53.3%) of A+TNEO-T+ participants progressed to MCI or dementia. Among A+TNEO-T+ individuals, the progressors (27.8 ± 1.5) had worse baseline MMSE scores compared to stable individuals (28.6 ± 1.0, P = 0.02; Extended Data Table 2) and tended to have higher tau PET retention at baseline (Supplementary Fig. 1), suggesting that progressors were already in a slightly more advanced disease stage at the start of this study. Cox proportional-hazards models, adjusted for age, sex, education and cohort, showed an increased risk for future progression to MCI in the A+TNEO-T+ (HR = 19.2, 95% CI = 10.9–33.7, P < 0.001), A+TMTL+ (HR = 14.6, 95% CI = 8.1–26.4, P < 0.001) and A+T− (HR = 2.4, 95% CI = 1.4–4.3, P = 0.002) groups compared to the A−T− (reference) group (Fig. 2a,b). Both A+TMTL+ (HR = 6.0, 95% CI = 3.4–10.6, P < 0.001) and A+TNEO-T+ (HR = 7.9, 95% CI = 4.7–13.5, P < 0.001) groups also showed faster clinical progression to MCI than the A+T− group (Fig. 2c). Pairwise log-rank tests showed that the A+TMTL+ and A+TNEO-T+ groups did not differ from each other (P = 0.19). Fifty percent of the A+TNEO-T+ and A+TMTL+ groups had progressed to MCI after 42.8 and 43.6 months, respectively.

6During clinical follow-up, 21 participants progressed to all-cause dementia: 4 out of 781 (0.5%) in A−T−, 3 out of 292 (1.0%) in A+T−, 2 out of 51 (3.9%) in A+TMTL+ and 12 out of 60 (20%) in A+TNEO-T+. Of those, 14 progressed to clinically defined AD-type dementia and 7 to non-AD dementias (see Extended Data Table 3 for dementia type). Cox proportional-hazards models, adjusted for age, sex, education and cohort, demonstrated an increased risk for future progression to all-cause dementia in the A+TNEO-T+ (HR = 41.3, 95% CI = 16.7–101.9, P < 0.001) and A+TMTL+ (HR = 6.2, 95% CI = 1.4–26.8, P = 0.01) groups compared to the A−T− (reference) group (Fig. 2d,e). There was no difference between the A+T− and the A−T− group (HR = 1.6, 95% CI = 0.5–5.4, P = 0.53). The A+TNEO-T+ (HR = 26.7, 95% CI = 10.4–68.1, P < 0.001) group showed faster clinical progression to all-cause dementia than the A+T− group, while there was no significant difference between the A+TMTL+ (HR = 3.8, 95% CI = 0.9–16.2, P = 0.08) and the A+T− group (Fig. 2f). Pairwise log-rank tests showed that the A+TNEO-T+ group progressed significantly faster to all-cause dementia than the A+TMTL+ group (P = 0.01). Similar results were found when using progression to AD-type dementia (11 A+TNEO-T+, 2 A+TMTL+ and 1 A+T−) as outcome instead of all-cause dementia (Extended Data Fig. 2).

7Linear mixed-effect models adjusting for age, sex, education and cohort indicated that the A+TNEO-T+ (standardized β (stdβ) of interaction with time in months ± s.e. = −0.020 ± 0.002, T = −10.14, P < 0.001), A+TMTL+ (stdβ = −0.017 ± 0.002, T = −8.84, P < 0.001) and A+T− (stdβ = −0.005 ± 0.001, T = −5.26, P < 0.001) groups showed faster decline over time on the modified preclinical Alzheimer cognitive composite 5 (mPACC5 (ref.)) compared to the A−T− (reference) group (Fig. 3a). Additionally, the A+TNEO-T+ (stdβ = −0.16 ± 0.002, T = −7.53, P < 0.001) and A+TMTL+ (stdβ = −0.13 ± 0.002, T = −6.21, P < 0.001) groups progressed faster than the A+T− group but there was no difference between the A+TNEO-T+ and A+TMTL+ groups (stdβ = −0.003 ± 0.002, T = −1.13, P = 0.26). Exploratory study of the mPACC5 subcomponents showed that A+TNEO-T+ and A+TMTL+ groups did not differ on delayed episodic memory (stdβ = −0.002 ± 0.003, P = 0.47; Fig. 3c) but the A+TNEO-T+ group showed faster decline on timed executive function (stdβ = −0.007 ± 0.002, P = 0.003; Fig. 3d) and semantic memory (stdβ = −0.009 ± 0.003, P = 0.007; Fig. 3e).

8On the MMSE, the A+TNEO-T+ (β = −0.056 ± 0.005, T = −11.55, P < 0.001), A+TMTL+ (β = −0.024 ± 0.005, T = −4.72, P < 0.001) and A+T− (β = −0.008 ± 0.002, T = −3.46, P < 0.001) groups showed faster decline over time compared to the A−T− (reference) group (Fig. 3b). The A+TNEO-T+ (stdβ = −0.49 ± 0.005, T = −9.51, P < 0.001) and A+TMTL+ (stdβ = −0.16 ± 0.005, T = −3.04, P = 0.002) groups progressed faster than the A+T− group and the A+TNEO-T+ group declined faster than the A+TMTL+ group (stdβ = −0.033 ± 0.007, T = −4.82, P < 0.001). Cognitive trajectories on the MMSE and mPACC5 for each cohort are displayed in Extended Data Fig. 3.

9Based on the lower age of the A−T− group compared to the biomarker-positive groups (Table 1), we performed the Cox proportional-hazards models and linear mixed models (1) stratified by different age groups (that is, 50–69, 70–79 and 80+ years; Tables 2 and 3) and (2) restricting the A−T− group to individuals >65 years so that groups were age-matched (Extended Data Tables 4 and 5). All analyses yielded highly similar results as the primary analyses, suggesting that age did not explain the observed differences in clinical progression rates between AT groups and that the findings can be generalized across different age groups.

10To account for the potential effect of white matter lesions on clinical progression, we additionally adjusted the Cox proportional-hazards models and linear mixed models for white matter hypointensity volumes (Extended Data Tables 6 and 7). These analyses yielded highly similar results as the primary analyses, suggesting that the observed differences in clinical progression rates between AT groups were independent of white matter pathology.

11To examine whether amyloid and tau PET-positive cognitively unimpaired individuals are destined to decline, we performed a multicenter study in 1,325 participants with, on average, approximately 3.5 years of clinical follow-up data available. We found that A+TNEO-T+ and A+TMTL+ cognitively unimpaired individuals had clearly increased risk for future development of MCI and all-cause dementia and showed steep trajectories of cognitive decline. Hence, evidence of advanced AD pathology provided by amyloid and tau PET is strongly associated with short-term clinical progression in initially cognitively unimpaired individuals. This supports the NIA-AA criteria-based classification of A+T+ cognitively unimpaired individuals as ‘preclinical AD’, especially when ‘T’ is defined by PET. To consider A+T+ merely as a risk factor, and not manifest disease, may be an underestimation of its malignancy.

12Although the A+TNEO-T+ group was at increased risk for progression to all-cause dementia compared to the A+TMTL+ group, there were no differences in risk for progression to MCI. A potential explanation is that tau pathological changes in the MTL can cause severe-enough memory loss leading to an individual being classified as MCI (but not dementia), while widespread tau pathology into the neocortex might be needed to produce a dementia syndrome,,. Supporting this hypothesis, we found that the A+TNEO-T+ group exhibited faster decline in global cognition, semantic memory and timed executive function but not in delayed episodic memory function compared to the A+TMTL+ group (Fig. 3c–e).

13Another important finding was the higher clinical progression rate for both A+T+ groups compared to the A+T− group, with HRs up to 7.9 and 26.7 for progression to MCI and all-cause dementia, respectively. This indicates that among A+ cognitively unimpaired individuals, who are presumably already on the AD pathological continuum, the coexistence of tau pathological changes in the MTL and/or the neocortex represents a ssubstantial additional relative risk for short-term cognitive decline. However, a proportion of the A+T+ group remained cognitively intact after approximately 3.5 years of follow-up, highlighting the variable rates of cognitive decline even among individuals who are at the highest risk of deterioration based on their AD biomarker profile–,,,. This possibly represents highly resilient individuals due to a favorable genetic makeup and/or a healthy lifestyle–. Alternatively, these nonprogressors may harbor fewer additional factors on top of Aβ and tau pathology, for example, synaptic loss, copathologies or neuroinflammation, resulting in an attenuation of their cognitive decline–. Hence, research into both resilience and risk factors is necessary to optimize future prediction models.

14The main strengths of this study include the large-scale multicenter dataset with available amyloid PET, tau PET and longitudinal clinical and cognitive data, which allowed us to accurately estimate the relative risk of the different AT groups in terms of clinical progression and cognitive decline. This study also has several limitations. First, there are inherent challenges due to the multicenter study design, such as data pooling and harmonization, different amyloid and tau PET tracers and dissimilarities in PET acquisition protocols. Also, different neuropsychological tests were used as mPACC5 subcomponents across cohorts and these differences could impact the overall mPACC5 score. Relatedly, the varying rates of clinical progression in the A+TNEO-T+ and A+TMTL+ groups across cohorts (Supplementary Table 1) could be due to chance given the relatively small number of T+ participants but might also be a result of different ascertainment and recruitment methods. Second, the number of events (that is, progression to MCI and especially dementia) was relatively low. Third, we may have underestimated the actual risk of A+TNEO-T+ and A+TMTL+ cognitively unimpaired individuals due to consent or volunteer bias (lower study participation among individuals at risk) and informative censoring (the tendency of people to drop out when experiencing onset or worsening of symptoms). Fourth, we acknowledge that our design only allowed establishing relative risk and not lifetime risk and we did not control for the competing risk of death in our survival analyses.

15Future studies should test whether our findings are generalizable to more diverse populations in terms of ethnicity, socioeconomic status and medical comorbidities. Furthermore, studies with longer follow-ups and larger samples of A+TNEO-T+ and A+TMTL+ cognitively unimpaired individuals will help refine the current findings. This may, in turn, aid in enriching clinical trials for fast progressors and developing algorithms for a personalized prognosis that reliably estimates the risk for future cognitive decline at an individual level. Of particular interest will be the performance of head-to-head studies between tau PET and high performing plasma p-tau assays– to investigate the potential added value and cost-effectiveness of tau PET as a prognostic tool.

16We included 1,325 participants from the Mayo Clinic Olmsted Study of Aging (MCSA, n = 680), the Swedish BioFINDER-1 (n = 56) and BioFINDER-2 (n = 228) studies at Lund University,, the Berkeley Aging Cohort Study (BACS, n = 109), the Harvard Aging Brain Study (HABS, n = 162, data obtained in March 2022 from data release 2.0 via https://habs.mgh.harvard.edu), the Australian Imaging, Biomarker & Lifestyle Flagship Study of Ageing (AIBL, n = 48) and the SCIENCe project, which is part of the Amsterdam Dementia Cohort (ADC, n = 42). A brief description of each cohort is provided in Supplementary Table 2. All participants were (1) cognitively unimpaired at baseline defined by neuropsychological test scores within the normative range given an individual’s age, sex and educational background, (2) had amyloid PET available to determine Αβ status, (3) underwent a tau PET scan before 1 January 2019, to allow for sufficiently long follow-up and (4) had at least one clinical follow-up visit available. Follow-up data were collected until 1 April 2022. Amyloid and tau PET scans included in the study were acquired at the same time point in most cases and always within a maximum of 1 year of each other. Written informed consent was obtained from all participants and local institutional review boards for human research approved the study. This includes the Mayo Clinic and Olmsted Medical Center institutional review boards for MSCA, regional ethics committee at Lund University for BioFINDER-1 and BioFINDER-2, institutional review boards at Lawrence Berkeley National Laboratory and the University of California, Berkeley, for BACS, institutional human research ethics committees of Austin Health, St. Vincent’s Health, Hollywood Private Hospital and Edith Cowan University for AIBL, partners human research committee for HABS and the medical ethics review committee of the Amsterdam University Medical Center for ADC.

17Amyloid PET scans were performed and analyzed at each respective cohort site. Αβ status was determined using center-specific cutoffs or visually read metrics using [18F]flutemetamol PET for BioFINDER-1 and BioFINDER-2, [11C]Pittsburgh compound-B PET for MCSA, BACS and HABS, [18F]florbetapir PET for ADC and AIBL (n = 47 out of 48) and [18F]NAV4694 for AIBL (n = 1 out of 48). Each cohort provided the Αβ status for their participants (see Supplementary Table 3 for details).

18Tau PET was performed using [18F]flortaucipir across all cohorts, except BioFINDER-2 where [18F]RO948 was used, and data were processed according to previously described procedures (Supplementary Table 4). BioFINDER-1, BioFINDER-2 and BACS (part of a previous multicenter study) tau PET scans were analyzed at Lund University. For the other cohorts, tau PET scans were processed at the respective sites; standardized uptake value ratios and region-of-interest (ROI) volumes were sent to the statistical analysis team (R.O. and A.P.B.) at Lund University. Based on these data, we computed the tau PET status for an MTL (an unweighted average of bilateral entorhinal cortex and amygdala) and an NEO-T (a weighted average of bilateral middle temporal and inferior temporal gyri) ROI. We used an unweighted MTL ROI because we intended the entorhinal cortex to have a relatively higher contribution since this region is involved in the earliest stages of tau accumulation yet it is somewhat smaller compared to the amygdala. The MTL and NEO-T ROIs were modified from a previously described temporal meta-ROI, based on a well-established stereotypical progression of tau pathology from the MTL into the lateral temporal cortex,,. The threshold was determined for each cohort separately, based on the mean + 2 × s.d. across all Αβ-negative participants in each cohort (see cohort-specific cutoffs in Supplementary Table 4). Based on amyloid and tau PET status, we generated four different biomarker groups: A−T−; A+T−; A+TMTL+ (defined as tau PET-positive in the MTL but not in the neocortex); and A+TNEO-T+ (defined as tau PET-positive in the NEO-T and/or MTL; 49 out of 65 were also TMTL+). The A−T+ group was considerably smaller than the other groups (n = 34; Extended Data Table 1), hence their results are only reported in Extended Data Fig. 1.

19We used both binary and continuous measures of clinical progression. First, we examined progression from cognitively unimpaired to MCI (Fig. 2a), all-cause dementia (Fig. 2d) or AD-type dementia (Extended Data Fig. 2). MCI was established using the Petersen criteria and is defined as significant cognitive symptoms as assessed by a physician, in combination with cognitive impairment on one or multiple domains (for example, memory, executive functioning, attention, language) that is below the normative range given an individual’s age, sex and educational background but not sufficiently severe to meet the diagnostic criteria for dementia. A large systematic review assessing 11,000 studies showed convergence across practices when validated diagnostic tools were used, such as in the current study. AD-type dementia was diagnosed using established criteria. Both MCI and dementia diagnoses were made by clinicians who were blinded for any PET or CSF outcome. For BACS, no formal diagnosis of MCI or dementia was made during the study; hence, the cohort was excluded from this analysis. Second, we examined cognitive trajectories using a sensitive composite measure specifically developed to detect cognitive changes in preclinical stages of AD (that is, the mPACC5 (refs.,); Fig. 3a) and a screening tests of global cognition (that is, the MMSE, which is frequently used in clinical practice and in trials; Fig. 3b). The mPACC5 consists of tests capturing episodic memory, executive function, semantic memory and global cognition. Individual tests were z-transformed using the baseline test scores of Αβ-negative participants in each cohort as the reference group and then averaged to obtain a composite z-score. The composition of mPACC5 is described for each cohort in Supplementary Table 5. Note that we used a modified version of the mPACC5 because, although we measured the same cognitive domains, the specific tests used in this study are not consistent with the original mPACC5 and they differ by cohort. However, a direct comparison between the original PACC5 and the mPACC5 in HABS (where the original PACC5 was developed) showed a strong correspondence between the 2 versions (r = 0.89, P < 0.001; Supplementary Fig. 2).

20All statistical analyses were performed in R v.4.0.5. Differences in baseline characteristics between groups were assessed using analysis of variance (ANOVA) with post-hoc t-tests with Bonferroni correction for continuous variables and chi-squared and Kruskal–Wallis with post-hoc Mann–Whitney U-tests for categorical or ordinal variables. First, we examined progression from cognitively unimpaired to MCI (Fig. 2a), all-cause dementia (Fig. 2d) or AD-type dementia (Extended Data Fig. 2) using Cox proportional-hazards models, adjusting for age, sex, education and cohort using A−T− as the reference group. We additionally repeated the Cox proportional-hazards model analysis while using the A+T− group as the reference group. Furthermore, we compared the A+TMTL+ and A+TNEO-T+ groups using pairwise log-rank tests with false discovery rate correction. For individuals who progressed to MCI and subsequent dementia, we used the respective times at conversion to MCI and dementia for the analyses presented in Fig. 2a,d. Second, we examined differences in cognitive trajectories between groups on the mPACC5 (Fig. 3a) and on global cognition (that is, the MMSE; Fig. 3b) using linear mixed-effect models with random intercepts and slopes, adjusting for age, sex, education and cohort. Finally, we examined whether our findings would generalize across different age groups and whether our results would be consistent when accounting for white matter pathology. Therefore, we performed two sets of analyses. First, we performed the main analyses when (1) stratifying the participants into different age groups (that is, 50–69, 70–79 and 80+ years old) and (2) restricting the A−T− group to individuals older than 65 years, so that all groups were age-matched. Second, we additionally adjusted the statistical models for a measure of white matter pathology, that is, white matter hypointensity volumes derived from T1-weighted magnetic resonance imaging scans using the standard FreeSurfer pipeline. Statistical significance for all models was set at two-sided P < 0.05.

21Further information on research design is available in the Nature Research Reporting Summary linked to this article.

22Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41591-022-02049-x.