Longitudinal tau aggregation, atrophy, and cognitive decline in Alzheimer's disease

Abstract

INTRODUCTION

The independent contributions of baseline and longitudinal tau positron emission tomography (PET) and magnetic resonance imaging (MRI) to cognitive decline remain unclear.

METHODS

We included n = 761 amyloid‐positive individuals from the Swedish BioFINDER‐2 study with [¹⁸F]RO948‐tau‐PET, 3‐Tesla structural‐MRI, and cognition (n = 322 with longitudinal imaging data). Linear‐mixed‐models with random‐intercepts and ‐slopes or linear‐regressions were adjusted for age, sex, education, diagnosis, and other‐imaging‐modality.

RESULTS

Tau‐PET showed stronger associations with cognitive decline than MRI, showing the strongest associations in a neocortical‐composite‐region with a cognitive composite (β = −0.25 ± 0.02, p < 0.001) for baseline and longitudinal tau‐PET (β = −0.62 ± 0.05, p < 0.001). Baseline tau‐PET explained the largest proportion of cognitive decline (54.0%–94.0%), with modest mediation effects for longitudinal tau‐PET or MRI pathways (2.0%–15.0%). Simulated reductions of tau‐PET‐slopes (up to 100%) were associated with marginally altered cognitive trajectories.

DISCUSSION

The strong associations between baseline tau‐PET and longitudinal cognition, with marginal contributions of longitudinal tau‐PET and MRI, emphasize the importance of baseline tau aggregates for prognostics and treatments in Alzheimer's disease (AD).

Highlights

• Baseline and longitudinal regional tau‐PET uptake were more closely associated than structural MRI with longitudinal cognitive decline.

• Baseline tau‐PET was a stronger determinant of longitudinal cognitive decline than longitudinal tau‐PET.

• Simulated reductions of tau‐PET accumulation showed limited alterations of cognitive trajectories, with potential implications for tau‐targeting therapies.

1. BACKGROUND

In vivo imaging measures of tau pathology and neurodegeneration, that is, tau positron emission tomography (PET) and magnetic resonance imaging (MRI), have shown strong associations with cognitive decline in Alzheimer's disease (AD). ¹ , ² Compared with MRI, tau‐PET showed higher accuracy for distinguishing AD from other neurodegenerative diseases ³ , ⁴ and stronger predictive power for cross‐sectional cognitive performance. ⁵ , ⁶ , ⁷ In addition, tau‐PET has shown strong prognostic value in AD, ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² even in early stages when no overt symptoms are present. ¹³ While stronger independent associations with longitudinal cognition have been found for neurodegeneration than tau pathology in preclinical AD by some studies, ¹⁴ superior associations for tau‐PET have been observed consistently in preclinical ⁹ and general AD. ¹⁰ , ¹⁵ , ¹⁶ However, large‐scale head‐to‐head comparisons of independent effects of tau‐PET versus MRI on cognitive decline across clinical stages of AD are relatively lacking. Insights into these associations may provide guidance to select optimal markers to prognosticate AD as well as to track disease progression, which is highly relevant for clinical trial design, both in terms of inclusion criteria ¹⁷ and outcome measures. ¹⁸

According to the temporal sequence of biomarker progression in AD, ¹⁹ tau pathology is suggested to precede atrophy, ²⁰ and to drive cognitive decline. ¹⁵ However, studies showing partial mediation effects of atrophy of the relationship between cross‐sectional tau pathology and cognition, suggested alternative tau‐mediated pathways to cognitive decline independent from neuronal loss. ⁵ , ⁶ While baseline tau pathology and atrophy have shown complementary contributions to cognitive decline, ¹² , ¹⁵ the complex interplay between baseline and changes in tau pathology and neurodegeneration remains largely unknown. Mapping these complex interactions, similar to prior work assessing contributions of baseline and longitudinal components of amyloid‐beta (Aβ) and tau pathology on cognitive decline in preclinical AD, ²¹ may enhance our understanding of specific determinants of cognitive decline in AD. In addition, these insights could inform therapeutic trials on the most optimal screening and disease progression markers. ²²

Therapeutic trials targeting tau pathology have shown relative inefficacy using tau antibodies aimed at inhibiting tau protein aggregation or spread, which has led to the general consensus that targeting tau aggregation may not be early or sufficient enough to prevent transsynaptic spread and cognitive effects. ²³ Simulations of clinical trials in AD have been proven feasible based on real‐world data from large clinical research cohorts. ²⁴ Better insights into the impact of altering tau aggregation on cognitive trajectories may inform clinical trials on the effectiveness of reducing the rate of tau production or aggregation versus targeting already accumulated tau load. With the recent developments of anti‐Aβ treatments ²⁵ and potential diversity in treatment effects among individuals with differing disease stages, combination therapies targeting amyloid in conjunction with tau may be important in AD in the future. ²²

Therefore, the aims of the current study were to investigate (a) the independent associations between baseline and rate‐of‐change in tau‐PET or MRI and cognitive decline, (b) the independent contributions of baseline and rate‐of‐change in tau‐PET and MRI to cognitive decline using serial mediation models, and (c) the effects of simulated reductions of slopes of tau aggregation on cognitive trajectories across AD clinical stages.

2. METHODS

2.1. Participants

We included participants from the ongoing prospective Swedish BioFINDER‐2 cohort (NCT03174938, http://www.biofinder.se/). All participants were recruited at Skåne University Hospital and the Hospital of Ängelholm, Sweden, and the cohort covers the full spectrum of AD, ranging from Aβ‐positive adults with intact cognition and no memory complaints (healthy controls) and subjective cognitive decline (SCD) to mild cognitive impairment (MCI) and AD dementia. The former two were labeled as cognitively unimpaired (CU), and the latter two as cognitively impaired (CI). The main inclusion criteria, as described previously, ²⁶ were age > 40 years, fluency in the Swedish language, and diagnoses were assigned based on Mini‐Mental State Examination (MMSE) scores as well as clinical evaluations including neuropsychological testing. Exclusion criteria were having significant unstable systemic illness, significant active alcohol or substance misuse, or refusing lumbar puncture or neuroimaging. MCI diagnosis was established if participants performed below 1.5 standard deviations from controls on at least one cognitive domain from an extensive neuropsychological battery examining verbal fluency, episodic memory, visuospatial ability, and attention/executive domains. ²⁷ AD dementia diagnoses were determined using criteria for dementia due to AD from the Diagnostic and Statistical Manual of Mental Disorders Fifth Edition and if positive on Aβ biomarkers based on the National Institute on Aging and the Alzheimer's Association (NIA‐AA) criteria for AD. ¹⁹ All included participants were Aβ‐positive based on cerebrospinal fluid (CSF) Aβ42/40 ratios, with a pre‐established cutoff of 0.08 using the Elecsys immunoassays (Roche Diagnostics). ²⁶ Participants were included if they had at least one visit with tau‐PET, MRI, and cognitive data. This led to the inclusion of n = 253 CU Aβ‐positive individuals (n = 125 cognitively normal controls and n = 128 individuals with SCD) and n = 508 CI Aβ‐positive individuals (n = 237 with MCI and n = 271 with dementia). All data for the current study were acquired between April 2017 and October 2023.

2.2. Image acquisition and processing of tau‐PET and MRI

MRI was performed using a Siemens 3T MAGNETOM Prisma scanner (Siemens Medical Solutions). Structural T1‐weighted MRI images were acquired from a magnetization‐prepared rapid gradient echo (MPRAGE) sequence with 1 mm isotropic voxels. PET images were acquired on digital GE Discovery MI scanners. For tau‐PET, acquisition was done 70–90 min post injection of ∼370 MBq [¹⁸F]RO948. Images were processed according to our pipeline described previously. ³ Briefly, PET images were attenuation corrected, motion corrected, summed, and registered to the closest T1‐weighted MRI processed through the longitudinal pipeline of FreeSurfer version 6.0. Standardized uptake value ratio (SUVR) images were created using the inferior cerebellar gray matter as the reference region. ²⁸ To enhance the stability of region‐of‐interest (ROI) delineation across scans, parcellation and ROI definition were conducted using FreeSurfer (version 6.0, https://surfer.nmr.mgh.harvard.edu/) on a midpoint MRI, created by merging baseline and follow‐up MRI scans. The defined ROIs were then warped back to the individual MRI scans. Finally, these ROIs were applied to PET data transformed into each participant's native T1 space to extract mean regional SUVR values. In previous work, longitudinal results were consistent using the inferior cerebellum or a composite white matter (WM) region. ²⁹

RESEARCH IN CONTEXT

1. Systematic review: Previous studies examining the longitudinal associations between tau‐positron emission tomography (PET), magnetic resonance imagi (MRI), and cognition have focused mainly on brief cognitive screening tools, and lacked investigation of the independent prognostic value of tau‐PET compared to MRI, as well as examinations of the relative importance of baseline versus longitudinal measures of tau‐PET for future cognitive decline.

2. Interpretation: Our results demonstrate that tau‐PET shows stronger independent associations with cognitive changes over time than MRI using composite cognition scores, that baseline tau‐PET was the strongest determinant of future cognitive decline, and that reductions of tau‐PET slopes in simulated scenarios showed only marginal alterations in cognitive trajectories.

3. Future directions: Future research should focus on early detection, combined biomarker strategies (e.g., PET, plasma), and evaluation of tau‐targeted therapies at earlier disease stages and their influence on cognitive outcomes.

2.3. Regional tau PET and MRI

All PET and MR images were parcellated based on the Desikan–Killiany atlas, ³⁰ using FreeSurfer (FreeSurfer version 6.0) image analysis pipelines (surfer.nmr.mgh.harvard.edu). For tau‐PET, we computed SUVRs for entorhinal (weighted average of bilateral entorhinal cortices), amygdala (weighted average of bilateral amygdala volumes), inferior middle temporal (weighted average of bilateral middle temporal and inferior temporal gyri), and neocortical (weighted average of all neocortical regions) ROIs. The entorhinal, amygdala, and inferior middle temporal regions were based on a well‐established progression of tau pathology from the medial temporal lobe into the lateral temporal cortex ³¹ , ³² as previously described. ¹³ For MRI, cortical thickness was quantified in entorhinal, inferior middle temporal, and neocortical regions using the same FreeSurfer subregions as for tau‐PET. In addition, hippocampal volume was calculated based on a weighted average of bilateral hippocampal volumes, corrected for intracranial volume.

2.4. Cognition

For cognition, both the MMSE and the modified Preclinical Alzheimer's Cognitive Composite‐5 (mPACC5) were used. A cognitive composite score analogous to the original PACC5 was calculated as a measure to capture early cognitive decline. ³³ , ³⁴ The tests included were the MMSE, Alzheimer's Disease Assessment Scale‐Cognitive Subscale (ADAS‐Cog) delayed recall, Trail‐Making Test version A (TMT‐A), and Category Fluency of animals. The original PACC5 includes two measures of memory recall: Logical Memory and the Free and Cued Selective Reminding Test. Given that in our cohort only one memory score was available, ADAS‐cog delayed recall was assigned a double weight to maintain the same proportion of memory as in the original composite score. ²⁹ , ³⁵ Further, the Digit Symbol Substitution Test used in the original PACC5 was replaced by TMT‐A, to avoid missing values, as not all patients with MCI and dementia completed the Symbol Digit Modalities Test, as described previously. ²⁹ All tests were z‐scored based on the mean and standard deviation of CU Aβ‐negative participants over 50 years old and then averaged to generate cognitive composite scores. For MMSE, scores were scaled during statistical analyses.

2.5. Statistical analysis

Demographic differences were assessed with two sample t‐tests or chi‐squared tests where appropriate. We tested associations between baseline tau‐PET or MRI and longitudinal cognition using linear mixed models fitted with random intercepts and slopes, adjusting for time, age, sex, education, and diagnosis (CU vs CI). The models included interactions for time with all other predictors. In addition, we assessed baseline tau‐PET and MRI interaction terms versus cognitive decline using these models to assess synergistic effects.

Next, we tested associations between longitudinal tau‐PET or MRI (rate‐of‐change) and cognitive decline (rate‐ofchange) using a two‐step approach. We first generated subject‐level rates‐of‐change in tau‐PET, MRI, and cognition, using linear mixed effect models with random intercepts and slopes for each brain region or cognitive test separately. These models used tau‐PET SUVR, MRI, or cognition as the dependent variables and time (from the baseline scan or test date) as the independent variable. This yielded subject‐level rates‐of‐change, for each imaging or cognitive measure which were then used in subsequent analyses. We tested associations between the slopes of tau‐PET or MRI as predictors and slopes of cognition as outcome (using the subject‐level rates calculated before) in linear regression models, adjusting for age, sex, education, and diagnosis (CU vs CI).

For models using baseline values or slopes of tau‐PET and MRI as predictors, models were fitted with and without correction for the other imaging modality to assess the independent effects of each modality on longitudinal cognition. To test significant differences between imaging predictors, models in the previous step were first bootstrapped (using n = 1000 iterations). Models were applied to the overall cohort and repeated in CU and CI groups separately in post‐hoc analyses. In sensitivity analyses, models were repeated while adjusting for baseline values of the same modality, and the average percent effect change was calculated for both modalities. In addition, we assessed rate‐of‐change tau‐PET and MRI interaction terms versus rate‐of‐change cognition using these models to assess synergistic effects. Other sensitivity analyses included repeating analyses in a subgroup of individuals with follow‐up time <2 years to assess the influence of variable follow‐up time on results.

Next, to test relative contributions of baseline and longitudinal tau‐PET and MRI to longitudinal cognition, serial mediation models were performed. Models were constructed for the regions showing the highest associations with longitudinal cognition in the previous steps, that is, the inferior middle temporal and neocortical regions. Models were constructed assessing the mediating effect of longitudinal tau‐PET and MRI on the associations between baseline tau‐PET and longitudinal cognition, adjusting for age, sex, education, and diagnostic group (CU vs CI) and bootstrapping with n = 5000 iterations.

To characterize to what degree reductions of tau aggregation impact the degree of cognitive decline, we simulated effects on longitudinal cognition according to different interventions on tau‐PET slopes. We tested hypothetical reductions of tau‐PET slopes by 15%, 30%, 50%, or 70% and 100%, similar to previously reported procedures. ³⁶ We then compared predicted cognitive slopes when reducing (i.e., simulated “treatment” arms) versus when not reducing tau‐PET slopes (i.e., simulated “placebo” group). Differences between treatment arms were assessed using analysis of variance (ANOVAs) and Tukey honestly significant difference (HSD) post‐hoc testing, with Bonferroni correction. Analyses were repeated using reductions of MRI slopes. Sensitivity analyses included repeated analyses with 15%–30% reduction arms for tau‐PET and baseline adjustments for tau‐PET and MRI models. All analyses were performed in R version 4.0.5. The main packages used were lme4 v1.1‐30 for linear mixed effect models, lavaan v0.6‐13 for mediation analyses, predict v3.1‐2 for simulation models, and ggplot2 v3.3.6 for plots.

3. RESULTS

3.1. Participants

We included n = 253 CU Aβ‐positive individuals (Aβ+, n = 125 cognitively normal controls and n = 128 individuals with SCD) and n = 508 CI Aβ‐positive individuals (Aβ+, n = 237 with MCI and n = 271 with dementia). Demographic characteristics are presented in Table 1. The average follow‐up time was 2.5 ± 1.2 years in the overall cohort, 3.1 ± 1.2 years in CU A+, and 2.3 ± 1.2 years in CI Aβ+ individuals. The mean age was 73.0 ± 7.9 years, 52% were female, and the mean length of education was 13.0 ± 4.1 years. At baseline the CI Aβ+ group had lower MMSE and mPACC5 scores, higher tau‐PET load, and lower cortical thickness and hippocampal volume (all p < 0.05, Table 1). A subset of the cohort (n = 120 CU Aβ+ and n = 202 CI Aβ+ individuals) also had longitudinal imaging data available, with similar basic demographic characteristics as the overall sample (Table S1). Imaging and cognitive slopes of this sample are shown in Table S1 and Figure S1. Assessment of subject‐level slopes as input for the models can be found in Table S2 and Figures S2–S5.

TABLE 1.

Participant characteristics.

Parameter	CU Aβ +	CI Aβ +	Overall	p‐value
n	253	508	761
Age, mean	73.0 ± 8.9	73.0 ± 7.3	73.0 ± 7.9	0.335
Sex (% female)	134 (53.0)	265 (52.0)	399 (52.0)	0.896
Education, years	13.0 ± 3.7	12.0 ± 4.3	13.0 ± 4.1	0.435
Follow‐up time, years	3.1 ± 1.2	2.3 ± 1.2	2.5 ± 1.2	<0.001
MMSE, mean	29.0 ± 1.3	24.0 ± 4.4	25.0 ± 4.4	<0.001
mPACC5, mean	−0.3 ± 0.8	−2.9 ± 1.7	−2.0 ± 1.9	<0.001
Tau PET SUVR, entorhinal	1.4 ± 0.3	1.8 ± 0.4	1.7 ± 0.5	<0.001
Tau PET SUVR, amygdala	1.1 ± 0.3	1.6 ± 0.5	1.4 ± 0.5	<0.001
Tau PET SUVR, inferior middle temporal	1.3 ± 0.3	1.9 ± 0.7	1.7 ± 0.7	<0.001
Tau PET SUVR, neocortical	1.1 ± 0.1	1.4 ± 0.4	1.3 ± 0.4	<0.001
MRI cortical thickness, entorhinal	3.1 ± 0.3	2.8 ± 0.4	2.9 ± 0.4	<0.001
MRI cortical volume, hippocampus	0.0023 ± 0.0003	0.0020 ± 0.0003	0.0021 ± 0.0003	<0.001
MRI cortical thickness, inferior middle temporal	2.5 ± 0.1	2.4 ± 0.1	2.4 ± 0.1	<0.001
MRI cortical thickness, neocortical	2.3 ± 0.1	2.2 ± 0.1	2.3 ± 0.1	<0.001

Note: Data are presented as mean ± SD unless indicated otherwise. Differences between diagnostics groups (assessed separately for Aβ‐positive and Aβ‐negative groups) were assessed using ANOVA (age, education, MMSE) and χ² tests (sex).

Abbreviations: Aβ, amyloid‐β; CI, cognitively impaired; CU, cognitively unimpaired; MMSE, Mini‐Mental state examination; mPACC5, modified Preclinical Alzheimer's disease Cognitive Composite; MRI, magnetic resonance imaging; PET, positron emission tomography; SUVR, standardized uptake value ratio.

3.2. Associations between baseline tau‐PET or MRI and cognitive decline

When assessing the associations between baseline tau‐PET or MRI and cognitive decline, the strongest predictors of longitudinal MMSE were tau‐PET in the inferior middle temporal (β = −0.18 ± 0.013, p < 0.001), meaning that the MMSE score decreased 0.18 standard deviations per year for every standard deviation higher baseline tau‐PET SUVR, and neocortical regions (β = −0.17 ± 0.01, p < 0.001). The next best predictor was cortical thickness in inferior middle temporal regions (β = −0.11 ± 0.01, p < 0.001). For mPACC5, the strongest predictors were tau‐PET in the inferior middle temporal region (β = −0.22 ± 0.02, p < 0.001), the neocortical region (β = −0.25 ± 0.02, p < 0.001), and the entorhinal region (β = −0.11 ± 0.02, p < 0.001; Figure 1 and Table S3). Results including plots for the inferior middle temporal region are shown in Figure S6.

FIGURE 1.

Associations between baseline tau‐PET or MRI and longitudinal cognitive decline in CU and CI Aβ+ individuals. (A) Associations between baseline tau‐PET and MRI and MMSE in the whole cohort (CU Aβ+ and CI Aβ+ individuals combined) while adjusting for age, sex, education and cognitive status (CU vs CI). (B) These associations for the mPACC5. Standardized coefficients for tau‐PET are depicted in blue, and standardized coefficients for MRI are depicted in green. Circles represent models without adjustment for the other modality, while triangles represent models with adjustment for the other imaging modality. ** = significant after FDR correction, * = significant at p = 0.05. (C,D) Changes in MMSE and mPACC5 by quartiles of neocortical tau‐PET (blue) or MRI (green), adjusted for age, sex and education. Error bars represent standard errors in (A) and (B) and 95% confidence intervals in (C) and (D). Aβ, amyloid‐beta; CI, cognitively impaired; CU, cognitively unimpaired; FDR, false discovery rate; MMSE, Mini‐Mental State Examination; mPACC5, modified Preclinical Alzheimer's Cognitive Composite‐5; MRI, magnetic resonance imaging; PET, positron emission tomography.

For the MMSE, the effects in most regions were significantly different from the next best region, except for MRI entorhinal versus MRI hippocampal volume (Δβ = −0.013[−0.028, 0.003], Table S4). For the mPACC5, all effects were significantly different from the next best region, except for MRI inferior middle temporal versus MRI hippocampal volume (Δβ = −0.004[−0.025, 0.017], Table S4) and MRI entorhinal versus MRI neocortical (Δβ = 0.014[−0.005, 0.034], Table S4). Similar results were obtained when repeating analyses for CU Aβ+ and CI Aβ+ groups separately (Tables S5 and S6; Figures S7 and S8).

We next adjusted the tau‐PET models for MRI in the same region, and vice versa. This reduced the effects markedly for some of the MRI measures (especially inferior middle temporal and neocortical regions), while tau‐PET effects were generally more robust (mean reduction of effect: 14.9% for tau PET vs. 46.7% for MRI, Figure 1 and Table S3). After this adjustment, tau‐PET measures were the strongest predictors for both MMSE (inferior middle temporal: β = −0.16 ± 0.01, p < 0.001, reduction: 10.5% when adjusting for MRI; neocortical: β = −0.16 ± 0.01, p < 0.001, reduction: 5.5%; entorhinal: β = −0.10 ± 0.01, p < 0.001, reduction: 17.7%) and mPACC5 (neocortical: β = −0.25 ± 0.02, p < 0.001, reduction: 2.4%; inferior middle temporal: β = −0.21 ± 0.02, p < 0.001, reduction: 3.6%; entorhinal: β = −0.10 ± 0.02, p < 0.001, reduction: 10.7%). Interactions between baseline tau‐PET and MRI predictors vs cognitive decline are shown in Table S7). Demographics of the subgroup with < 2 years of follow‐up time are shown in Table S8.

3.3. Associations between longitudinal tau‐PET or MRI and cognitive decline

We next tested associations between regional longitudinal tau‐PET or MRI (rate‐of‐change) and cognitive decline (rate‐of‐change). Change in tau‐PET levels in the neocortical region showed the strongest associations with change in MMSE (β = −0.56 ± 0.04, p < 0.001) and change in mPACC5 (β = −0.62 ± 0.05, p < 0.001), followed by inferior middle temporal cortical thickness (MMSE: β = −0.55 ± 0.04, p < 0.001, mPACC5: β = −0.48 ± 0.05, p < 0.001) and inferior middle temporal tau‐PET (MMSE: β = −0.48 ± 0.05, mPACC5: β = −0.47 ± 0.05, p < 0.001; Figure 2 and Table S9). For both tests, all predictors were significantly different from the next best predictor, for example, first versus second rank and second versus third rank, except for MRI neocortical versus MRI hippocampal volume for MMSE (Δβ = −0.025[−0.188–0.133]), and mPACC5 (Δβ = 0.084[−0.096–0.275], Table S10).

FIGURE 2.

Associations between rate‐of‐change in tau‐PET or MRI and rate‐of‐change in cognition in CU and CI Aβ+ individuals. (A) Associations between slope in tau‐PET and MRI and slope in MMSE in the whole cohort (CU Aβ+ and CI Aβ+ individuals combined) while adjusting for age, sex, education and cognitive status (CU vs CI), while (B) these associations for the mPACC5. The slopes were obtained from linear mixed models with random slopes and random intercepts, assessing either tau∼time, MRI∼time, or cognition∼time, yielding average rate‐of‐change metrics per individual. These were used as input for the linear regressions depicted in these panels. (C,D) Associations between rate‐of‐change in MMSE and mPACC5 and rate‐of‐change in tau‐PET or MRI in regions of interest. ** = significant after FDR correction, * = significant at p = 0.05. Error bars represent standard errors in (A) and (B) and 95% confidence intervals in (C) and (D). Aβ, amyloid‐beta; CI, cognitively impaired; CU, cognitively unimpaired; FDR, false discovery rate; MMSE, Mini‐Mental State Examination; mPACC5, modified Preclinical Alzheimer's Cognitive Composite‐5; MRI, magnetic resonance imaging; PET, positron emission tomography.

When adjusting these models  for rate‐of‐change in the other modality, the strongest predictors remained  similar for both MMSE (tau‐PET neocortical: β = −0.46 ± 0.05, p < 0.001, reduction: 18.6%; MRI inferior middle temporal: β = −0.42 ± 0.05, p < 0.001, reduction: 23.5%; tau‐PET inferior middle temporal: β = −0.26 ± 0.05, p < 0.001, reduction: 45.7%) and mPACC5 (tau‐PET neocortical: β = −0.54 ± 0.06, p < 0.001, reduction: 13.9%; MRI inferior middle temporal: β = −0.34 ± 0.06, p < 0.001, reduction: 30.3%; tau‐PET inferior middle temporal: β = −0.28 ± 0.06, p < 0.001, reduction: 39.9%, Figure 2 and Table S9). Similar results were obtained when repeating the analyses for CU Aβ+ and CI Aβ+ individuals separately (Tables S11 and S12; Figures S9 and S10). Overall, rate‐of‐change measures of tau‐PET and MRI showed similar reductions of effects after correction for the other modality (reduction of effect: 29.6% for tau‐PET vs. 28.1% for MRI). Next, these models were fit while also adjusting for baseline imaging levels (Table S13 and Figure S11). This had more pronounced effects for tau‐PET, where the rate‐of‐change effects were generally strongly reduced, while effects for rate‐of‐change in MRI were less affected by adjustment for baseline MRI. Repeated analyses for CU Aβ+ and CI Aβ+ individuals separately are shown in Tables S14 and S15. Interactions between rate‐of‐change tau‐PET and MRI predictors on rate‐of‐change cognition are shown in Table S16 and Figure S12. Results in the subgroup with < 2 years of follow‐up time are shown in Figure S13.

FIGURE 3.

Contributions of baseline versus change in tau‐PET and MRI to cognitive decline in CU and CI Aβ+ individuals. (A,B) Associations between (rate‐of‐change in) tau‐PET, MRI, and rate‐of‐change in MMSE in inferior middle temporal and neocortical regions, respectively, in the whole cohort (CU Aβ+ and CI Aβ+ individuals combined). (C,D) Associations between (rate‐of‐change in) in tau‐PET, MRI, and rate‐of‐change in mPACC5 in inferior middle temporal and neocortical regions, respectively, in the whole cohort (CU Aβ+ and CI Aβ+ individuals combined). Analyses were adjusted for age, sex, education, and cognitive status (CU vs CI), using n = 5000 bootstrapping iterations. Solid lines indicate significant effects, while dotted lines indicate nonsignificant associations. Aβ, amyloid‐beta; CI, cognitively impaired; CU, cognitively unimpaired; MMSE, Mini‐Mental State Examination; mPACC5, modified Preclinical Alzheimer's Cognitive Composite‐5 MRI, magnetic resonance imaging; PET, positron emission tomography.

FIGURE 4.

Cognitive trajectories according to simulated treatment‐induced reductions of tau aggregation in CU and CI Aβ+ individuals. (A,B) Predicted MMSE values over time in inferior middle temporal and neocortical regions, respectively, in the whole cohort (CU Aβ+ and CI Aβ+ individuals combined) in different arms of tau‐PET slopes, that is, placebo (no intervention) versus 15%, 30%, 50%, 70%, and 100% reductions of tau‐PET slopes, while (C) and (D) show predicted mPACC5 values over time in inferior middle temporal and neocortical regions, respectively, in the whole cohort (CU Aβ+ and CI Aβ+ individuals combined). Error bars represent 95% confidence interval. Aβ, amyloid‐beta; CI, cognitively impaired; CU, cognitively unimpaired; MMSE, Mini‐Mental State Examination; mPACC5, modified Preclinical Alzheimer's Cognitive Composite‐5; PET, positron emission tomography.

3.4. Contributions of baseline versus change in tau‐PET or MRI to cognitive decline

Next, we modeled the relationship between baseline and longitudinal imaging measures and cognitive decline using serial mediation analyses. We used the regions with the strongest associations with longitudinal cognition in previous steps, that is, the inferior middle temporal and neocortical regions. Baseline tau‐PET was modeled as the main predictor, while tau‐PET rate‐of‐change and baseline and rate‐of‐change MRI measures were implemented as mediators in associations with cognitive decline (Figure 3 and Tables S17 and S18).

Overall, the largest proportion of longitudinal cognition was explained directly by baseline tau‐PET, ranging from 90.0% to 94.0% of the total effect explained in inferior middle temporal and 54.0%–82.0% in neocortical regions (MMSE, inferior middle temporal: β = −0.48[−0.71, −0.26], p < 0.001, proportion mediated: 90.0%, neocortical, β = −0.26[−0.53, −0.01], p < 0.05, proportion mediated: 54.0%; mPACC5, inferior middle temporal: −0.47[−0.75, −0.16], p < 0.001; proportion mediated: 94.0%, neocortical: −0.47[−0.76, −0.08], p < 0.01, proportion mediated: 82.0%). In the inferior middle temporal region, the effect was partially mediated through MRI rate‐of‐change (indirect pathway 1; MMSE: β = −0.07[−0.17, −0.008], p < 0.05, proportion mediated: 0.14; mPACC5: β = −0.08[−0.19, −0.01], p < 0.05, proportion mediated: 0.15) and baseline MRI and MRI rate‐of‐change (indirect pathway 2; MMSE: β = −0.05[−0.09, −0.03], p < 0.05, proportion mediated: 0.10; mPACC5: β = −0.040[−0.076, −0.017], p < 0.05, proportion mediated: 0.08). In the neocortical region, the effect was partially mediated through tau‐PET rate‐of‐change and MRI rate‐of‐change (indirect pathway 1; MMSE: β = −0.05[−0.12, −0.01], p < 0.05, proportion mediated: 0.10; mPACC5: β = −0.05[−0.12, −0.01], p < 0.05, proportion mediated: 0.09) and baseline MRI and MRI rate‐of‐change (indirect pathway 2; MMSE: β = −0.02[−0.05, −0.01], p < 0.05, proportion mediated: 0.05; mPACC5: β = −0.01[−0.40, −0.003], p < 0.05, proportion mediated: 0.02).

3.5. Cognitive trajectories according to simulated reductions of tau aggregation

To assess the effect of interventions on tau aggregation on cognitive trajectories, we simulated interventions on tau‐PET slopes and examined the effects on cognitive trajectories. We simulated reductions of 15%, 30%, 50%, 70%, and 100% on tau‐PET slopes. Results for inferior middle temporal and neocortical regions are shown in Figure 4 and Table S19a,b. When assessing the neocortical region‐of‐interest, we observed marginal differences in cognitive slopes between placebo and treatment arms with tau‐PET slope reductions (MMSE, placebo: −1.3 ± 1.00 vs. 15%: −1.2 ± 0.95 vs. 30%: −1.2 ± 0.86, 50%:−1.1 ± 0.77, 70%:−0.96 ± 0.68, and 100%: −0.81 ± 0.64; mPACC5, placebo: −0.39 ± 0.33 vs. 15%: −0.37 ± 0.30 vs. 30%:−0.35 ± 0.28, 50%:−0.32 ± 0.25, 70%:−0.29 ± 0.23, and 100%:−0.25 ± 0.23, Figure 4 and Table S19b). Differences between treatment arms were observed in placebo versus 50% (Δ_{rate‐of‐change} = 0.25[0.06–0.44], p = 0.003), placebo versus 70% (Δ_{rate‐of‐change} = 0.35[0.16,0.54], p < 0.0001), placebo versus 100% (Δ_{rate‐of‐change} = 0.51[0.31–0.70], p < 0.0001), 100% versus 30% (Δ_{rate‐of‐change} = 0.36[0.16–0.55], p < 0.0001), and 100% versus 50% (Δ_{rate‐of‐change} = 0.25[0.06–0.44], p = 0.002), 100% versus 15% (Δ_{rate‐of‐change} = 0.43[0.24–0.62], p < 0.0001) and 70% versus 15% (Δ_{rate‐of‐change} = 0.28[0.08–0.47], p = 0.0006) for MMSE, and placebo versus 70% (Δ_{rate‐of‐change} = 0.10[0.04,0.17], p = 0.0002), placebo versus 100% (Δ_{rate‐of‐change} = 0.14[0.08–0.21], p < 0.0001), 100% versus 30% (Δ_{rate‐of‐change} = 0.10[0.03–0.17], p = 0.0003), and 100% versus 15% (Δ_{rate‐of‐change} = 0.12[0.05–0.19], p < 0.0001) for mPACC5 (Table S20a,b). Similar results were observed for the inferior middle temporal region‐of‐interest (Figure 4 and Tables S19a and S20b). Tables S21a and S22b and Figure S14 show results when assessing CU Aβ+ and CI Aβ+ individuals separately. Repeated analyses using 15%–30% reduction arms are shown in Figures S15 and S16. Repeated main analyses with MRI slopes are shown in Figures S17 and S18. Repeated analyses with adjustment for baseline imaging values are shown in Figure S19.

4. DISCUSSION

In an Aβ+ cohort including CU and CI individuals, baseline levels and rate‐of‐change tau‐PET showed superior independent associations with cognitive decline compared to MRI. In addition, baseline tau‐PET showed the largest direct contribution to cognitive decline in serial mediation models. Simulated treatment‐induced reductions of tau‐PET aggregation showed more marginal attenuations of cognitive trajectories. Taken together, this study demonstrated the utility of tau‐PET as a prognostic and disease monitoring marker relative to MRI and suggested baseline tau‐PET as a strong determinant of future cognitive decline in AD.

The baseline tau‐PET measures showing the strongest associations with longitudinal cognitive decline included inferior middle temporal and neocortical composite regions. While the entorhinal cortex ³⁷ or medial temporal cortex ³⁸ are suggested as early tau‐PET markers and the inferior middle temporal as a later stage marker, ³⁷ tau‐PET positivity in inferior middle temporal regions has shown high prognostic value in AD. ¹³ Regarding tau‐PET rate‐of‐change, our findings are consistent with recent work in a smaller sample, showing the strongest associations between neocortical tau‐PET and cognitive decline while comparing tau pathology and neurodegeneration. ¹⁵ The substantial heterogeneity in tau topography and progression patterns in AD may favor the use of whole‐brain neocortical regions rather than specific one‐size‐fits‐all regions of interest. ³⁹ Alternatively, uptake in neocortical composites may reflect more advanced tau pathology associated with faster cognitive decline. Moreover, stage‐specific or individualized approaches may provide higher sensitivity to early changes with greater potential for personalized approaches. ³² , ³⁶

Direct effects of baseline tau pathology, independent of neurodegeneration, on baseline ⁵ and longitudinal cognition ¹² have been observed in AD. Associations between rate‐of‐change tau‐PET and cognitive decline were observed in a CU population in the absence of associations between tau accumulation and structural changes. ⁴⁰ The current study represents one of the first large‐scale analyses to confirm superior independent associations between change in tau‐PET (compared to change in MRI) and cognitive decline. The mechanisms by which tau pathology may exert these direct effects remain unclear. Tau pathology may impact neuronal function either prior to neuronal loss or through neuronal loss not quantified by MRI by various mechanisms, including early neurotoxicity of neurofibrillary tangle bearing neurons, subtle morphometric neuronal changes, synapse reductions, ⁵ , ⁴¹ chronic innate neuroinflammation ⁴² , ⁴³ or vascular mechanisms. ⁴⁴ , ⁴⁵ Network alterations are associated with both baseline and rate‐of‐change measures of tau‐PET and may mediate associations between tau‐PET and cognitive decline. ⁴¹ , ⁴⁶ , ⁴⁷ In addition, these findings may be caused by variability in MRI that does not depend on AD pathology, such as premorbid differences, non‐pathological effects of aging on brain structure, and atrophy due to non‐AD processes. ⁴⁸ , ⁴⁹ , ⁵⁰ Note that the effects of tau accumulation on cognitive decline in mediation models were still partially dependent on neurodegeneration, supporting their complementary contributions. ¹² , ¹⁵ Lastly, they may represent “end products” of upstream processes leading to cognitive alterations, as described in the “tombstone” theory of AD. ⁵¹ , ⁵² , ⁵³ In that scenario, both tau aggregation and neurodegeneration may signal a general neurodegenerative state of the brain, rather than being driving forces of cognitive decline.

Baseline tau pathology was more strongly associated with cognitive decline than tau aggregation rates in serial mediation models. While similar associations between baseline and rate‐of‐change in tau‐PET and cognitive decline have been observed in a smaller AD cohort ⁵⁴ and stronger associations between change in tau pathology and cognitive decline than baseline tau pathology and cognitive decline have been observed in preclinical AD, ²¹ our study provides robust evidence for the importance of baseline versus longitudinal tau‐PET based on the systematic large‐scale cohort and comprehensive modeling. This finding may be partly due to the strong associations between tau aggregation rates and baseline tau load. ³⁷ , ⁵⁴ Baseline tau‐PET load represents the accumulation of tau pathology over the course of many years. The total amount of aggregated tau may continuously exert neurotoxic effects in the brain, which are not only due to newly formed tau aggregates. Alternatively, test–retest variability of tau‐PET may hamper the utility of longitudinal tau‐PET as a prognostic marker, which is higher for longitudinal MRI and fluorodeoxyglucose (FDG)‐PET, which were more strongly associated with cognitive decline. ⁵⁵ , ⁵⁶ Future work could incorporate head‐to‐head comparisons between tau‐PET and fluid markers specific to tau pathology, such as MTBR‐tau243 ⁵⁷ or p‐tau205 ⁵⁸ to assess the most potent and cost‐effective disease progression markers in AD. CSF MTBR‐tau243 has indeed shown decreases with lecanemab administration, ⁵⁹ suggesting its utility as a disease progression marker and potential association with cognitive decline.

Simulated clinical trials with tau aggregation reductions showed rather limited impact on cognitive trajectories. Current therapeutic strategies targeting tau include strategies aimed at (a) preventing translation of tau mRNA, (b) altering tau post‐translational modifications, or (c) inhibiting the spread of misfolded tau or inhibiting tau aggregation. ²³ The relative inefficacy of interventions using tau antibodies aimed at inhibiting tau protein aggregation, combined with the relative success of an antisense oligonucleotide (ASO), ⁶⁰ , ⁶¹ suggests that targeting tau protein aggregation may not be early or sufficient enough to prevent transsynaptic spread. ²³ Although speculative based on our simulated data, we infer that aggressive lowering of tau beyond original baseline levels may be necessary to convey cognitive effects. Bepranemab ⁶² is among the first tau‐targeting antibodies to slow tau accumulation according to longitudinal tau‐PET and showed no effect on the primary outcome, that is, Clinical Dementia Rating Scale, Sum of Boxes (CDR‐SB). Therefore, a modest slowing of tau accumulation may not translate into a meaningful clinical benefit, aligning with our current results. Tau may need to be targeted very early in the disease process, as also suggested for anti‐amyloid trials ⁶³ and by tau‐targeting ASO results. ⁶⁰ , ⁶¹ Alternatively, tau seeds may spread quickly throughout the brain in early stages, creating local aggregation and neurotoxicity, suggesting local tau seed interactions as potential therapeutic targets. ⁶⁴ Furthermore, according to the “tombstone” analogy of tau, ⁵¹ , ⁵² , ⁵³ targeting extracellularly aggregated tau may not intervene in upstream processes in AD. Lastly, it cannot be ruled out that our current results may be partly explained by the high dependence of tau aggregation on baseline tau‐PET load. It is possible that altering tau‐PET progression may still be effective in patients with low baseline tau‐PET loads.

A strength of this study is the standardized longitudinal multimodal imaging and cognition dataset across AD clinical stages. A limitation is that we only relied on structural MRI as a measure of neurodegeneration. Another study with FDG‐PET showed similar results, that is, superiority of tau‐PET in associations with cognitive decline. ¹⁵ In our imaging versus longitudinal cognition models, variable follow‐up time may influence results. Sensitivity analyses in a subgroup of < 2 years follow‐up time showed similar results, suggesting these effects are minimal. Furthermore, the influence of other co‐pathologies, for example, TAR DNA‐binding protein 43 (TDP‐43) pathology on cognitive decline cannot be ruled out. Future work could incorporate direct markers of co‐pathologies, including CSF α‐synuclein measures, ²⁷ , ⁶⁵ , ⁶⁶ as well as novel plasma measures of TDP‐43 pathology. ⁶⁷ Due to large sample size requirements (sample sizes of n = 200 or larger ⁶⁸ ), serial mediation analyses were performed in the combined cohort and were not repeated within diagnostic subgroups. Based on the similar results obtained in the first steps of analyses between CU Aβ+ and CI Aβ+ individuals, disease‐stage‐specific effects were deemed marginal. Moreover, it is unclear how our simulation models translate to real‐world clinical trial data. Cognitive effects may take longer to appear, and there might still be effects of slowing tau aggregation that are difficult to predict from observational data. Furthermore, current treatment arms ranging between 15% and 100% reductions may not match observed reductions in current tau‐targeting clinical trials. ²³ However, results suggest that even in scenarios where tau aggregation is reduced to high degrees, cognitive benefits are unlikely to manifest. In the future, with more data, it may be feasible to develop machine learning methods, that can model complex interactions between heterogeneous baseline amyloid and tau burden and aggregation indices in relation to cognitive trajectories, especially in preclinical AD. ²¹ , ⁶⁹ , ⁷⁰ , ⁷¹ Other future directions include investigating more heterogeneous cohorts, based on the predominantly white population in BioFINDER‐2, and vascular pathology in current models. ⁴⁵ Current analyses were adjusted for age, sex, education and diagnosis. In addition, we acknowledge limitations of tau‐PET compared to MRI regarding cost‐effectiveness, accessibility, and standardization challenges. Finally, in general, our current models assess associations and cannot infer causality.

In conclusion, we observed (a) stronger independent associations between tau‐PET and cognitive decline compared to MRI, both at baseline and longitudinally; (b) greater contributions of baseline tau pathology versus tau aggregation and neurodegeneration over time to cognitive decline; and (c) limited effects of simulated treatment‐induced reductions of tau aggregation on cognitive trajectories in AD. This study demonstrated the utility of tau‐PET as an independent prognostic and disease monitoring marker throughout the clinical course of AD. In addition, it suggested baseline tau load as a strong determinant of future cognitive decline and suggested the need for early interventions in therapeutic trials targeting tau pathology, or interventions that lower tau load and not only slow accumulation rates.

CONFLICT OF INTEREST STATEMENT

E.H.S., A.P.B., O.S., and E.S. have nothing to disclose. NMC has received consultancy/speaker fees from Biogen, Owkin, Merck and Eli Lilly. S.P. has received grant from Avid Pharmaceuticals and KI elements/ADDF (paid to the institution) and speaker fees from Eli Lilly, Esai, BioArctic, and Novo Nordisk. R.O. has received research funding from European Research Council, ZonMw, NWO, National Institute of Health, Alzheimer Association, Alzheimer Nederland, Stichting Dioraphte, Cure Alzheimer's fund, Health Holland, ERA PerMed, Alzheimerfonden, Hjarnfonden, Avid Radiopharmaceuticals, Janssen Research & Development, Roche, Quanterix and Optina Diagnostics, and lecture fees by GE Healthcare. O.H. has acquired research support (for the institution) from AVID Radiopharmaceuticals, Biogen, C2N Diagnostics, Eli Lilly, Eisai, Fujirebio, GE Healthcare, and Roche. He has received consultancy/speaker fees from AC Immune, Alzpath, BioArctic, Biogen, Bristol Meyer Squibb, Cerveau, Eisai, Eli Lilly, Fujirebio, Merck, Novartis, Novo Nordisk, Roche, Sanofi and Siemens. He is currently partially employed by Eli Lilly. Author disclosures are available in the supporting information.

CONSENT STATEMENT

All participants gave written informed consent. Ethical approval was given by the Regional Ethical Committee in Lund, Sweden. Approval for PET imaging was obtained from the Swedish Medical Products Agency and the local Radiation Safety Committee at Skåne University Hospital, Sweden.

Supporting information

Supporting Information

Supporting Information

Singleton EH, Mattsson‐Carlgren N, Pichet Binette A, et al. Longitudinal tau aggregation, atrophy, and cognitive decline in Alzheimer's disease. Alzheimer's Dement. 2025;21:e70435. 10.1002/alz.70435

DATA AVAILABILITY STATEMENT

Anonymized data will be shared by request from a qualified academic investigator for the sole purpose of replicating procedures and results presented in the article and as long as data transfer is in agreement with EU legislation on the general data protection regulation and decisions by the Ethical Review Board of Sweden and Region Skåne, which should be regulated in a material transfer agreement.