Wider brain scans of tau protein buildup relate to higher blood tau levels and memory decline in people at risk for Alzheimer's

The sample for the analysis was selected from an initial pool of n = 770 possible participants (535 from the Wisconsin Registry for Alzheimer's Prevention [WRAP] and 235 from the Wisconsin Alzheimer's Disease Research Center [WADRC]) as shown in the consort diagram in the Figure S1. We first limited the inclusion based on: (1) a non‐dementia cognitive status at their first applicable prior cognitive evaluation of either mild cognitive impairment (MCI; 37(4.90%)) or cognitively unimpaired (CU; 714 (95.10%)); (2) available qualitative visual reads for amyloid and tau status from PET exams that were not more than 3 years apart from each other. For the analysis of cognitive trajectories, we only included participants with available scores from a modified three‐test Preclinical Alzheimer's Cognitive Composite (PACC3), and we limited the observations included in the analysis to those that occurred at most 3 years after the most recent FQT PET (n = 728, see Table 1). Similarly, for the analysis examining pTau217 trajectories by tau group, we limited the sample to participants who had available plasma pTau217 measurements up to 3 years after the most recent FQT PET (n = 693). The rationale for the 3‐year interval was to maximize inclusion of data from participants who volunteered for multiple procedures while also allowing latitude due to uncontrollable factors, chiefly the coronavirus disease 2019 (COVID‐19) pandemic stoppage that affected cognitive and blood collection concurrent to PET.

Participants underwent [F‐18]FQT and [C‐11]Pittsburgh compound B (PiB) PET imaging and, in most cases, also underwent T1‐weighted magnetic resonance imaging (MRI) at the University of Wisconsin‐Madison medical campus. Detailed methods for radioligand synthesis and PET and MRI acquisition, processing, and quantification, and analysis were implemented as reported previously.1, 14, 15 Briefly, MK‐6240 involved a 20‐min imaging window in four 5‐min frames starting at 70 min post‐injection of 185–370 MBq. The start of the FQT acquisition at 70 min was chosen to minimize the relative portion of PET signal resulting from off‐target binding in the skull or meninges.1 The images frames were later realigned to overcome bulk motion, co‐registered to a T1‐weighted three‐dimensional MRI, and then summed and scaled to the inferior cerebellum gray as described previously. PiB imaging was acquired dynamically 0–70 min post‐injection of 555 MBq target dose from which distribution volume ratio (DVR) images were created using the cerebellum gray matter as the reference region. When a full dynamic series was not possible (not common), the 50–70 min post‐injection time window was acquired in 5‐min frames, and standardized uptake value ratio (SUVR) was created from the frame‐realigned summed image using the cerebellum gray matter as the reference region.

The Shuping et al. method10 was adapted as follows: FQT (SUVR; 70–90 min) images were derived and overlaid on coregistered T1‐weighted MRI, and scaled from 0 to 2.5 to assess tau burden by expert readers (S.C.J. and L.E.). Specifically, tau burden was rated visually for presence or absence in each of the two search areas (the MTL and neocortex; details in next paragraph). Both expert raters independently rated a subset of 30 cases to assess the reliability of visual ratings. The inter‐rater reliability Kappa statistics for the categories of T+_MTL and T+_Neo were K = 0.87 (28/30 agreement) and 1.0 (30/30 agreement), respectively. Subsequent scans were reviewed by one of the two expert readers (S.C.J.). Visual ratings on these categories were then combined to create four overall FQT tau status categories for each scan: T− (negative in both MTL and neocortex); T+_MTL (positive in the MTL and negative in the neocortex); T+_Neo (positive in the neocortex and negative in the MTL); and T+_MTL&Neo (positive in both regions). We assigned the participant's tau status based on the most recent visual read available from them.

Further details on the tau ratings: If present, spill‐in from extra‐axial tissue was not rated positive. To investigate the source of signal in such cases, the rater had the flexibility of increasing the scale range. For all exams, images were classified as tau negative or tau positive for the MTL (T+_MTL) where the search region included the entorhinal cortex and surrounding anterior parahippocampal gyrus, amygdala, or hippocampus (which was rare). To be considered T+_MTL, the signal needed to have a quality of expansion medially or antero‐posteriorly beyond a small confined focal locus. The reason is that very small loci may be of uncertain origin and of uncertain interpretation. The neocortex tau rating: Images were classified as tau negative or positive for the neocortex (T+_Neo) in a broad search region that included any cortical areas beyond the MTL. A categorical rating of T+_MTL&Neo was derived for cases in which both the MTL and the neocortex were rated T+. Example scans with these ratings are shown in Figure 1. The case in Figure 2 started as a single subtle posterior neocortical area and progressed over four exams at approximate 2‐year intervals to include the MTL and most of the posterior neocortex.

Amyloid positivity status was assessed qualitatively with visual read as positive or negative, with a positive rating requiring substantial cortical gray matter signal in one or more of medial, lateral, superior, or ventral lobe segments from any of the parietal, temporal, frontal, or occipital lobes, whether unilaterally or bilaterally observed. Within the negative category, a label of indeterminant was occasionally assigned when the signal was not of sufficient intensity or extent to be convincingly positive; these were grouped with amyloid negative in our analyses. As with the tau rating, the scaled DVR image (or SUVR when full dynamic data were not available) was overlaid on the coregistered T1‐weighted volume when available and the scale range set from 0 to 2.5 and visualized with the actc colormap in mricron. In addition to the visual rating, the amyloid burden was further quantified for descriptive purposes with a DVR cortical composite metric and centiloid values as previously described.14 The relationship of a positive visual read outcome to quantitative DVR and centiloid corresponded to a DVR composite threshold of 1.17 and a centiloid threshold of 18 using the receiver operating characteristic (ROC = 0.99 for each; see Figure S2).

At each main cognitive study visit, samples of blood were acquired whenever possible and processed as previously described16, 17 and stored at −80°C. Banked EDTA plasma samples were assayed to determine pTau217 concentration using the ALZpath single molecule array (Simoa) assay (ALZpath, Carlsbad, CA) on a Quanterix HD‐X instrument (Quanterix, Billerica, MA) running in the Wisconsin ADRC Biomarker Core. The assay itself was described previously by Ashton et al.18

Participants in WRAP completed a comprehensive cognitive battery at each visit, including tests of memory, executive function, language ability, and other aspects of cognitive function, alongside self‐ and study partner‐based measures of everyday functioning.19 Based on these measures, participant cognitive status at each visit was determined via multidisciplinary consensus conference review.20 Similarly, participants in the ADRC completed a comprehensive cognitive battery as part of the Uniform Data Set (UDS, versions 2 and 3) in addition to select tests that were in common with WRAP. Participant cognitive status after each ADRC visit was also determined via consensus conference review. For both cohorts, the cognitive status determination was blind to biomarker information and validated diagnostic criteria were used to identify MCI21 and dementia.22 Given that cognitive assessments began in 2001 (WRAP) and 2009 (ADRC) and tau PET scans began in 2017, cognitive evaluations included in this study were predominantly completed before tau PET.

For this study, the primary cognitive outcome was a modified version of the PACC; PACC composites have been shown by others23 and our group24, 25 to be sensitive to AD‐related preclinical decline. The modified PACC included a measure of learning (the Rey Auditory‐Verbal Learning Test, sum of Trials 1‐5; AVLT);26 a separate measure of delayed recall (the Wechsler Memory Scale‐Revised27 Logical Memory II story A recall raw score or, after UDS3 was implemented in the ADRC, the cross‐walked equivalent score on the Craft Story delayed recall;28 and a measure of executive function common to both cohorts (the Trail Making Test part B time to completion; max 300 s).29 Tests were combined to create the PACC by rescaling each test and computing an unweighted average, scaled such that first observations in CU individuals were distributed ∼N(0,1) with higher scores representing better performance on each test and the PACC composite. For additional details on the two cohorts’ test batteries, harmonization, and cognitive composites, see reference.25

All analyses were performed using R 4.4.030 and the R‐packages lme431 and emmeans32 to conduct the linear mixed models and post‐hoc comparisons, respectively. Sample characteristics including sex, age at amyloid and tau scans, amyloid status, amyloid burden, baseline plasma pTau217, baseline PACC performance, baseline cognitive status, and years of cognitive follow‐up were compared across the four tau categories (T−, T+_Neo, T+_MTL, and T+_MTL&Neo) using analyses of variance (ANOVAs) or chi‐squared tests where appropriate for each variable; significant group differences were followed by pairwise comparisons to further describe how tau groups differed.

We used mixed effects models for both our cognitive and biological longitudinal outcomes of interest (i.e., PACC and plasma pTau217, respectively). In both retrospective analyses, we included the four group tau status as a categorical fixed effect (reference category = T−), and we incorporated random intercepts per participant.

For the cognitive analysis, we used the longitudinal PACC scores from each participant as the dependent variable. We included as fixed effects the tau status, linear and quadratic terms for the age at PACC assessment, as well as the interaction between tau status and the linear age term. Additionally, we also incorporated covariates for sex (female = 0, male = 1), education level (0 = no bachelor's degree [BA], 1 = BA or more), and practice. This practice covariate was a continuous predictor representing the number of times a participant was exposed to the cognitive test. It can take a value ranging from 0 (baseline) to 6 (representing 7–11 since a model using a number of exposures as a categorical variable showed plateauing of practice effects). Furthermore, since we observed a significant interaction between the tau status and the age predictor, we estimated marginal means for each tau group at ages 65, 70, 75, and 80 and performed paired comparisons between tau groups (using Tukey's adjustment), reporting the mean difference (ΔM) of the marginal effects at these ages.

For the plasma pTau217 analysis, we used the longitudinal concentration levels as the dependent variable. The analysis utilized ptau217 measures obtained from samples before tau PET in order to investigate early disease trajectories in soluble tau relative to presumed later‐stage fibrillar tau burden,33 operationalized by tau PET visual rating. In this model, instead of age at plasma visit, we operationalized time as relative to the earliest FQT scan (FQT distance = age at plasma visit–age at earliest FQT scan), to evaluate differences in the trajectories from each group relative to the time when the first tau PET exam was performed. We included as fixed effects the tau status, linear and quadratic terms for the FQT distance, as well as the interaction between tau status and the linear FQT distance. We additionally incorporated the age at tau scan, centered at 65, as a covariate. Finally, since we also observed a significant interaction between the tau status and the FQT distance, we estimated the marginal means for the pTau217 levels of each tau group at 8, 4, and 0 years from the first FQT PET, and then we performed paired comparisons between groups (also using Tukey's adjustment).

We also ran the cognitive and biofluid mixed effects models, subdividing the Tau groups by their amyloid status, resulting in eight groups instead of four: (A−/T−, A−/T+_MTL, A−/T+_Neo, A−/T+_MTL&Neo, A+/T−, A+/T+_MTL, A+/T+_Neo, A+/T+_MTL&Neo), using the same set of covariates we implemented before for each of them. Since we also observed a significant interaction between PET status and the time variables, we estimated marginal means for each group and reported the paired comparisons between groups, corrected for multiple comparisons too. In exploratory analyses, we compared incident transition rates across groups, where the incident transition was coded as moving from baseline CU to MCI or dementia at the last visit or baseline MCI to dementia at the last visit (Table S1).

Model output for the retrospective PACC mixed effects model examining tau group by age interactions is provided in Table 2 (PACC model). Both the quadratic age term and the interaction between tau group and age were significant with parameter estimates indicating cognitive performance trajectories that become steeper (worse) at later ages, particularly in the T+_MTL&Neo group.

In Figure 3, we show both the observed and model‐estimated values for each group. In panel A we can see the estimated trajectories for each tau group (with sex = female, education = less than BA, practice = 0). Panel B shows the paired comparisons of each tau group's predicted PACC at ages 65, 70, 75, and 80. Finally, panel C displays the spaghetti plot of individual observed trajectories by tau group. As shown in Figure 3, panel B, none of the pairwise simple slope comparisons or predicted PACC levels between T− and T+_Neo differed significantly at any of the ages examined. In contrast, the average cognitive trajectory in T+_MTL&Neo participants deviated from the T− and T+_MTL groups at age 65 (T+_MTL&Neo vs. T−: ΔM = −0.53, t(724.79) = −5.32, p < 0.001; T+_MTL&Neo vs. T+_MTL: ΔM = −0.58, t(737.27) = −3.16, p = 0.009), and all three T groups by age 70 (T+_MTL&Neo vs. T−: ΔM = −1.00, t(758.85) = −9.82, p < 0.001; T+_MTL&Neo vs. T+_Neo: ΔM = −0.83, t(736.91) = −4.07, p < 0.001; T+_MTL&Neo vs. T+_MTL: ΔM = −0.83, t(708.62) = −4.61, p < 0.001). At age 70, the average performance from T+_MTL&Neo (−1.32) was approximately 4.11 times worse than the performance from the T− group (−0.32). By age 80, these estimated differences were even larger (T+_MTL&Neo vs. T−: ΔM = −1.93, t(1578.92) = −14.55, p < 0.001; T+_MTL&Neo vs. T+_Neo: ΔM = −1.81, t(1522.45) = −6.97, p < 0.001; T+_MTL&Neo vs. T+_MTL: ΔM = −1.34, t(1367.01) = −5.90, p < 0.001). Additionally, at age 80, we also saw that participants in the T+_MTL group performed worse than T− participants (T+_MTL vs. T−: ΔM = −0.59, t(1373.07) = −2.89, p = 0.020).

For this analysis, a total of 693 participants were included (549 (79.2%) T−, 24 (3.5%) T+_Neo, 32 (4.6%) T+_MTL, and 88 (12.7%) T+_MTL&Neo). The group characteristics were very similar to those observed in the PACC sample, (see Table S2 for details). We found differences in baseline pTau217 plasma levels, showing a higher baseline from the T+_MTL&Neo group compared to the others, and higher baseline levels from T+_MTL compared to T− participants. We did not observe differences in the number of years of plasma follow‐up or sex composition.

Model output for the retrospective plasma pTau217 mixed effects model examining the tau group by time interactions is shown in Table 2 (pTau217 model). The significant quadratic time and significant tau group by time interaction (F(3,3469.39) = 68.26, p < 0.001) indicate that pTau217 plasma levels increased with time at different rates for the tau groups. Figure 4 depicts observed and estimated plasma pTau217. Panel A shows the estimated trajectories (and confidence intervals) by tau group (with centered FQT age = 0). In panel B, we display paired comparisons between tau groups, contrasting their estimated pTau217 levels at 8, 4, and 0 years pre‐FQT scan. Panel C shows the individual plasma pTau217 trajectories by tau group. Overall, we observed the estimated pTau217 levels at all time points were higher for T+_MTL&Neo participants compared to other tau groups. T+_MTL on the other hand showed significantly elevated pTau217 values compared to T− starting 4 years before the first FQT scan, and T+_Neo only showed higher levels than T− at the time of the FQT scan. We did not find differences between T+_Neo and T+_MTL at any of the timepoints reviewed.

Paired comparisons of estimated pTau217 levels indicated that participants in the T+_MTL&Neo group had higher plasma levels of pTau217 compared to the other groups at each of the times examined (including 0, 4, and 8 years before FQT scan). T+_MTL&Neo participants showed higher pTau217 levels compared to the other groups since 8 years prior to the FQT scan (T+_MTL&Neo vs. T−: ΔM = 0.28, t(1136.48) = 9.27, p < 0.001; T+_MTL&Neo vs. T+_Neo: ΔM = 0.25, t(1058.72) = 4.30, p < 0.001; T+_MTL&Neo vs. T+_MTL: ΔM = 0.18, t(1025.46) = 3.36, p = 0.004). These differences continued 4 years before the FQT scan (T+_MTL&Neo vs. T−: ΔM = 0.46, t(730.91) = 17.33, p < 0.001; T+_MTL&Neo vs. T+_Neo: ΔM = 0.37, t(708.01) = 7.10, p < 0.001; T+_MTL&Neo vs. T+_MTL: ΔM = 0.29, t(702.03) = 6.22, p < 0.001) and at the time of the FQT scan (T+_MTL&Neo vs. T−: ΔM = 0.64, t(752.26) = 23.96, p < 0.001; T+_MTL&Neo vs. T+_Neo: ΔM = 0.49, t(770.65) = 9.17, p < 0.001; T+_MTL&Neo vs. T+_MTL: ΔM = 0.40, t(778.38) = 8.46, p < 0.001).

We observed that, 8 years before the first FQT scan, on average, the estimated pTau217 level for T+_MTL&Neo (0.59) was 1.93 times higher than the estimated average for T− (0.30), and by the time of the FQT scan, they reached pTau217 levels 2.80 times higher than those present in T− participants (1.00 vs. 0.36). T+_MTL participants also showed higher predicted average pTau217 levels compared to T− 4 years before the FQT scan (T+_MTL vs. T−: ΔM = 0.11, t(974.70) = 2.36, p = 0.087), which continued to be present at the FQT scan (T+_MTL vs. T−: ΔM = 0.24, t(781.98) = 5.61, p < 0.001). On the other hand, T+_Neo participants only showed significantly higher pTau217 levels compared to T− at the time of the FQT scan (ΔM = 0.16, t(772.48) = 3.23, p = 0.007). We did not see differences between T+_MTL and T+_Neo at any of the timepoints analyzed.

Model output is provided in Table S3 and results are depicted in Figure 5. Briefly, in both analyses, the A/T group by time interaction was significant. In follow‐up comparisons, we only saw differences in group trajectories in the A+ participants, and the overall patterns observed shifted a bit. Overall, both the A+/T+_MTL and A+/T+_MTL&Neo groups showed lower average cognitive scores compared to A+/T− group at age 75 (A+/T+_MTL vs. A+/T−: ΔM = −0.60, t(950.67) = −2.51, p = 0.012; A+/T+_MTL&Neo vs. A+/T−: ΔM = −1.56, t(994.84) = ‐10.3, p < 0.001). However, no differences were found between the A+/T+_MTL and A+/T+_Neo at any of the timepoints tested (Figure 5 A–C).

In the pTau217 analysis, all A+/T+ groups showed elevated pTau217 levels compared to the A−/T− group 4 years before their first FQT scan (A+/T+_MTL vs. A−/T−: ΔM = 0.33, t(652.82) = 6.93, p < 0.001; A+/T+_Neo vs. A−/T−: ΔM = 0.21, t(656.9) = 3.66, p < 0.001; A+/T+_MTL&Neo vs. A−/T−: ΔM = 0.53, t(691.48) = 19.92, p < 0.001). Among these A+ groups, participants in the A+/T+_MTL&Neo group had the highest pTau217 levels 4 years before their first FQT scan (vs. A+/T−: ΔM = 0.38, t(693.94) = 11.78, p < 0.001; vs. A+/T+_MTL: ΔM = 0.2, t(660.1) = 3.87, p < 0.001; vs. A+/T+_Neo: ΔM = 0.32, t(662.26) = 5.21, p < 0.001). Additionally, the A+/T+_MTL displayed elevated biomarker levels compared to the A+/T− group, 4 years before their first FQT scan (ΔM = 0.18, t(659.02) = 3.58, p < 0.001). We did not find significant differences in pTau217 levels between the A+/T− and A+/T+_Neo groups at any time point tested.