Accuracy of Tau Positron Emission Tomography as a Prognostic Marker in Preclinical and Prodromal Alzheimer Disease

Key Points Question What is the prognostic value of tau positron emission tomography (PET) for predicting cognitive decline across the clinical spectrum of Alzheimer disease? Findings In this longitudinal, multicenter prognostic study including 1431 participants, baseline tau PET predicted change in Mini-Mental State Examination scores during a mean (SD) follow-up of 1.9 (0.8) years. Moreover, tau PET outperformed established volumetric magnetic resonance imaging and amyloid PET markers in head-to-head comparisons, especially in participants with mild cognitive impairment and cognitively normal individuals who were positive for amyloid-β. Meaning These findings suggest that tau PET is a promising prognostic tool for predicting cognitive decline in preclinical and prodromal stages of Alzheimer disease.

A n accurate prognosis for individuals with Alzheimer disease (AD) is essential for patients and families to plan for the future, reduce uncertainty, increase safety, and optimize medical decision-making. 1 Despite the development of several biomarkers for neurodegeneration and AD pathology in past decades, 2 accurately predicting rates of cognitive decline in individuals with AD remains challenging. 3 Given the strong links between tau pathology and key correlates of cognition (eg, neuronal loss and synaptic dysfunction) observed in vitro and at autopsy, 4,5 in vivo information about the magnitude of cerebral tau pathology might improve the prediction of future cognitive decline.
A variety of positron emission tomography (PET) ligands have been developed that bind with high affinity to the tau aggregates formed in AD. [6][7][8] The degree and patterns of tau PET retention strongly overlap with regions affected by brain atrophy and hypometabolism [9][10][11][12] and correlate with concurrent cognitive performance. [13][14][15][16] In addition, tau PET has shown excellent diagnostic performance for distinguishing AD dementia from non-AD neurodegenerative disorders such as frontotemporal dementia or vascular dementia. [17][18][19][20] Recently, elevated baseline tau PET levels have been associated with accelerated cognitive decline over time, [21][22][23][24][25][26][27] but most studies had relatively modest sample sizes, lacked a replication cohort, and/or focused on 1 stage of the AD clinical continuum. The objectives of this prospective, longitudinal multicenter study were to (1) examine the prognostic value of [ 18 F]flortaucipir and [ 18 F]RO948 tau PET in a large cohort (n = 1431) of individuals with AD dementia, mild cognitive impairment (MCI), or normal cognition; (2) perform a head-to-head comparison of tau PET with established magnetic resonance imaging (MRI) and amyloid PET markers for predicting future cognitive change; and (3) investigate whether age, sex, and/or APOE genotype modify the association between baseline tau PET and cognitive change over time.

Participants
From an ongoing multicenter study, 18,28-30 we included 1431 participants from the Memory Disorder Clinic of Gangnam Severance Hospital, Seoul, South Korea (n = 161); the Swedish BioFINDER-1 (n = 136) and BioFINDER-2 (n = 296) studies at Lund University, Lund, Sweden; University of California, San Francisco (UCSF [n = 44]); the Alzheimer Disease Neuroimaging Initiative (ADNI [n = 445]) Avid Radiopharmaceuticals studies (A05 [n = 160]) and the placebo arm of the Eli Lilly solanezumab Expedition-3 study [n = 79]); and the Berkeley Aging Cohort Study (BACS [n = 110]). Data were collected from June 1, 2014, to February 28, 2021. Tau PET was performed using [ 18 F]flortaucipir-PET in the discovery cohort (1135 [79.3%] of the total sample) and [ 18 F]RO948-PET in the replication cohort (296 [20.7%] of the total sample from BioFINDER-2). Following National Institute on Aging-Alzheimer's Association diagnostic criteria, 31 we only included patients with AD dementia who were positive for amyloid-β (Aβ) on PET and/or cerebrospinal fluid (CSF) (n = 315) 18,28,29 ; 34 individuals with clini-cally diagnosed AD dementia who were negative for Aβ were excluded. We also included Aβ-positive (n = 271) and Aβnegative (n = 172) participants with MCI and Aβ-positive (n = 253) and Aβ-negative (n = 420) cognitively unimpaired individuals (CU group). In addition to tau PET, all participants underwent a medical history assessment and neurological examination, MRI, and a neuropsychological test battery including the Mini-Mental State Examination (MMSE). The MMSE is a diagnostic screening tool that measures a variety of cognitive abilities-including orientation to time and place, shortterm episodic memory, attention, problem solving, visuospatial abilities, and language and motor skills-and is often used as a cognitive outcome measure in longitudinal studies and clinical trials. Inclusion criteria for this study were MMSE assessment (n = 1431), MRI scan (n = 1431), and amyloid-PET scan (n = 1329) less than 6 months from tau PET and at least 2 MMSE time points (including baseline) with a minimum follow-up duration of 12 months. Written informed consent was obtained from all participants, and local institutional review boards for human research approved the study. This study followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) reporting guideline.

PET/MRI Acquisition
We acquired PET images using the following PET/computed tomography (CT) scanners: Biograph mCT (Siemens) in Seoul, 32 Discovery 690 (GE Healthcare) in BioFINDER-1, Discovery MI (GE Healthcare) in BioFINDER-2, 13,17 Biograph 6 Truepoint (Siemens) at UCSF and BACS, 1 2 , 3 3 and multiple scanners in the multicenter ADNI 34 and Avid Radiopharmaceutic als 2 32 3.0-T T i m T r i o (S i e m e n s) o r 3 .0 -T P r i s m a (S i e m e n s) i n BioFINDER-1 and -2, 13,17 3.0-T Tim Trio or 3.0-T Prisma (Siemens) at UCSF, 33 1.5-T Magnetom Avanto (Siemens) for BACS, 12 and multiple 1.5-T and 3-T scanners in the multicenter ADNI 34 and Avid Radiopharmaceuticals 23 cohorts.

T1-Weighted MRI Processing
The MRI data were centrally processed at Lund University using previously reported procedures. 13,17,18,28,29 Briefly, cortical reconstruction and volumetric segmentation were performed with FreeSurfer, version 6.0, image analysis pipelines (https:// surfer.nmr.mgh.harvard.edu/). Magnetization-prepared rapid gradient-echo images underwent correction for intensity homogeneity, removal of nonbrain tissue, and segmentation into gray matter, white matter, and CSF with intensity gradient and connectivity among voxels. 35 Cortical thickness was measured as the distance from the gray matter-white matter boundary to the perpendicular pial surface. 36 Reconstructed data sets were visually inspected for accuracy, and segmentation errors were corrected.

PET Processing
Tau PET images were first resampled to obtain uniform image size (128 × 128 × 63 matrix) and voxel dimensions (2.0 × 2.0 × 2.0 mm) across centers. Next, [ 18 F]flortaucipir/ [ 18 F]RO948 images were centrally processed at Lund University using previously reported procedures, 18,28,29 followed by motion correction using AFNI's 3-dimensional volume registration, calculation of mean time, and rigid coregistration to the skull-stripped MRI scan. Voxelwise standardized uptake value ratio (SUVR) images were created using inferior cerebellar gray matter as the reference region. 37 To extract mean regional SUVR values, FreeSurfer parcellation of the T1weighted MRI scan was applied to the PET data transformed to participants' native T1 space. For amyloid PET, we applied computational analysis of PET by AIBL (CapAIBL) 38 and tracerspecific conversion formulas to convert PET images or SUVR values into a Centiloid scale, which is a standard framework for the quantification of amyloid PET scans across tracers and cohorts. 39

Regions of Interest
In line with previous work, 17,18,28 we calculated the mean [ 18 F]flortaucipir and [ 18 F]RO948-PET SUVR in the entorhinal cortex, 15,16 a temporal meta-region of interest (ROI) that is a weighted mean of entorhinal, amygdala, parahippocampal, fusiform, and inferior and middle temporal ROIs, 40 and Braak stages V to VI encompassing widespread neocortical ROIs. 41 For MRI, we computed hippocampal volumes (adjusted for intracranial volume), an AD-signature cortical thickness ROI consisting of bilateral entorhinal, inferior, and middle temporal and fusiform cortex 40 and whole-brain cortical thickness (adjusted for surface area). 40 The temporal meta-ROI for tau PET and ADsignature cortical thickness ROI for MRI are reported in the main text, whereas the other ROIs are presented in eFigures 2 and 4 in the Supplement.

Statistical Analyses
We first performed a head-to-head comparison between [ 18 F]flortaucipir-PET and MRI for predicting change in MMSE over time. Therefore, single-participant slopes (representing annual change) for MMSE were calculated using linear regression models adjusted for age, sex, educational attainment, and cohort. These slopes were used as dependent variables in linear regression models, including continuous tau PET, MRI, or amyloid PET measures as predictors across the whole group and in the separate diagnostic groups. We performed bootstrapping with 1000 iterations to test whether the R 2 value differed between PET and MRI models. To test whether tau PET and MRI provide complementary information, we applied linear mixed-effects models with random intercepts and fixed slopes using longitudinal MMSE as a dependent variable. Our longitudinal data set was characterized by many participants for whom only 2 MMSE measurements were available. Although linear mixed models are generally able to accommodate this, including random slopes for participants led to overfitting of our models, whereas fixed-participant slopes led to the most parsimonious model. Model 1 included age, sex, educational attainment, and cohort as predictors. In model 2, either baseline tau PET or baseline MRI was added to model 1 as a predictor. In model 3, both imaging modalities (and the predictors from model 1) were entered simultaneously in a single model. We assessed model fit (Akaike information criterion) and examined differences in Akaike information criterion between models 1 and 2 and models 2 and 3 using the χ 2 statistic. We also performed mediation analysis to examine whether associations between baseline tau PET and longitudinal change in MMSE are mediated by MRI, adjusting for age, sex, educational attainment, cohort, and APOEε4 status. All analyses described above were also performed in the [ 18 F]RO948-PET replication cohort and were repeated for a head-to-head comparison between tau PET and amyloid PET (except for the mediation analysis). Finally, we tested whether the association between baseline tau PET and change in MMSE over time across all Aβ-positive participants is moderated by age, sex, or APOE genotype using linear mixed-effect models with a 3-way interaction term (time × tau PET × age/sex/APOE), adjusted for age, sex, educational attainment, and cohort. Significance level was set at 2-sided P < .05. We used R, version 4.0.2 (R Program for Statistical Computing), for the statistical analyses.

Participants
Participant characteristics across diagnostic groups are presented in Table 1 (and stratified by discovery/replication sample and by cohort in eTables 1 and 2 in the Supplement, respectively). The mean (SD) age of the study participants was 71.

Head-to-Head Comparison: Tau PET vs MRI
When comparing [ 18 F]flortaucipir SUVR in the temporal meta-ROI against MRI-based AD-signature cortical thickness in linear regression models with annual change in MMSE as dependent variable ( Figure 1 and

Complementary Information by PET and MRI
The results presented in Table 2 indicate that the prediction of decline in MMSE over time improved with both tau PET (R 2 for all participants, 0.49; R 2 for Aβ-positive AD dementia group, 0.34; R 2 for Aβ-positive MCI group, 0.35; R 2 for Aβ-positive CU group, 0.17) and MRI (R 2 for all participants, 0.46; R 2 for Aβpositive AD dementia group, 0.38; R 2 for Aβ-positive MCI group, 0.29; R 2 for Aβ-positive CU group, 0.12) compared with a basic model including age, sex, educational attainment, and cohort (R 2 for all participants, 0.19; R 2 for Aβ-positive AD dementia group, 0.20; R 2 for Aβ-positive MCI group, 0.21; R 2 for Aβ-positive CU group, 0.08) (all P < .001). Furthermore, tau PET and MRI provide complementary information, because when adding tau PET to linear mixed-effects models assessing MRI measures, the R 2 value increased (all participants, 0.46 vs 0.55; Aβ-positive AD dementia group, 0.38 vs 0.41; Aβ-positive MCI    Figure 2 shows path diagrams assessing AD-signature cortical thickness as a potential mediator of associations between baseline

Discussion
The main finding of this multicenter prognostic study was that baseline tau PET predicts group-level changes in MMSE over time across the AD clinical spectrum. In a head-to-head comparison with established MRI and amyloid PET markers, tau PET showed stronger associations with cognitive change, especially in preclinical and prodromal stages of AD. Part of the association between baseline tau PET and cognitive decline over time was mediated by baseline cortical thickness, but tau PET and MRI also provided complementary prognostic information. We identified age as a potential moderator of the association between baseline tau PET and longitudinal cognitive change, because older individuals showed more rapid cognitive decline at similar levels of tau load compared with younger individuals. Altogether, our findings suggest that tau PET is a promising tool for predicting future cognitive change that could support the prognostic process, especially in preclinical and prodromal stages of AD. Clinicopathological studies 4,5 have identified strong associations between tau pathology and cognition as well as key correlates of cognition such as loss of neurons and synaptic activity. These observations have been replicated in vivo using PET ligands that detect neocortical AD-like tau pathology with high accuracy, 7 because increased tau PET retention was associated with worse concurrent cognitive performance 10,14-16 as well as reductions in gray matter volume, glucose metabo- lism, and synaptic density. 9,42,43 Recent studies have indicated that elevated baseline tau PET levels were associated with accelerated cognitive decline over time, 21-27 but most of these studies had relatively modest sample sizes, included retrospective cognitive time points, lacked a replication cohort focused on only 1 stage of the AD clinical continuum, and/or did not perform head-to-head comparisons against MRI and amyloid PET markers. We included a large study population with prospective longitudinal assessment of MMSE across the clinical AD spectrum and demonstrate that tau PET is a powerful predictor of cognitive change over time and outperformed MRI and amyloid PET markers. This is an important first step toward further investigation of the potential of tau PET to act as a prognostic marker, especially in the early stages of AD, when estimating rates of future decline is notoriously challenging. Future research directions include the use of more sensitive (eg, the preclinical Alzheimer cognitive composite) or domainspecific (eg, episodic memory or executive functioning) cognitive tests, functional measures (eg, Clinical Dementia Rating Scale Sum of Boxes) or diagnostic conversion (eg, from MCI to AD dementia) as clinical readouts, longer follow-up durations, assessment of individualized prognostic models, and head-to-head comparisons against fluid biomarkers (eg, plasma phosphorylated tau) that are more scalable and possibly more cost-effective. Furthermore, in a recent successful phase 2 clinical trial with the Aβ-antibody donanemab, 44 Aβ-positive individuals with MCI or mild dementia were specifically selected based on intermediate levels of tau pathology on a PET scan. This suggests that tau PET biomarkers could be used as a selection tool for trial participants, but further investigation is warranted. We found that tau PET outperformed MRI markers in predicting future cognitive decline across all participants, in the Aβ-positive MCI group, and in the Aβ-positive CU group (both in the discovery and replication cohorts). For AD dementia, the results were inconsistent, with MRI performing slightly better compared with [ 18 F]flortaucipir in the discovery cohort, whereas in the replication sample, [ 18 F]RO948-PET clearly outperformed MRI. Because there were no major demographic differences between AD cases in the discovery vs the replication cohort (eTable 1 in the Supplement), this discrepant finding may be explained by the slightly greater dynamic range of [ 18 F]RO948 compared with [ 18 F]flortaucipir that enables [ 18 F]RO948 to slightly better capture cognitive change over time in more advanced clinical stages of AD. Altogether, these findings are in line with those of a previous cross-sectional study 13 showing that tau PET is more sensitive than MRI for detecting early cognitive change, whereas at the dementia stage, tau PET and MRI perform more equally. Greater sensitivity to detect early cognitive change using tau PET compared with MRI can possibly be explained by the large variations in brain structure that preexist in the general population, which may reduce the ability of structural MRI to reliably distinguish the earliest AD-related changes from premorbid differences in brain structure accentuated by age-related brain changes. Furthermore, tau PET may be more sensitive to early changes owing to the presumed occurrence of tau pathology before onset of neurodegeneration, 45 which might affect cognition through both structural (brain atrophy) 46 and functional (network disruption) 47 pathways. Finally, we found that cortical thickness only modestly mediated the association between baseline tau PET and MMSE slopes, an effect that was diseasestage specific because it was only observed in the AD dementia and Aβ-positive MCI groups (discovery cohort only, not replicated), but not in the Aβ-positive CU group. Tau PET also outperformed amyloid PET in predicting future cognitive change, which is in accordance with previous observations of modest cognitive correlates for levels of Aβ in stark contrast to associations of pathological tau burden. 4,10,[14][15][16][23][24][25] This can be explained by differences in the temporal evolution of Aβ and tau  pathology. Widespread Aβ pathology may emerge approximately 20 years before symptom onset, but the rate of accumulation attenuates over the disease course, which reduces its clinicopathological correlates. 4,48 In contrast, neocortical tau pathology is typically only observed when the disease has clinically manifested, and rates of tau accumulation are higher in symptomatic compared with asymptomatic individuals on the AD pathological continuum. 18,49 Overall, our findings suggest that tau PET may be the most optimal biomarker to identify those Aβ-positive individuals who are at risk for future cognitive decline and to predict cognitive trajectories in clinical trial participants with preclinical or prodromal AD. Age, sex, and APOE genotype have previously been shown to affect rates of tau accumulation and cognitive performance across the AD clinical spectrum. [50][51][52] In the present study, we examined whether age, sex, and APOE genotype act as modifiers of the association between baseline tau PET and cognitive change over time. In the discovery cohort, older individuals showed more rapid cognitive decline than younger individuals with a similar tau load. This could be explained by lower resilience against tau pathology (and/or associated neurodegeneration) in older individuals or by the presence of copathological features (eg, TAR DNA-binding protein 43 or vascular pathology) that are more likely to occur with advancing age. Sex did not affect the association between baseline tau PET and cognitive change over time. Previous work has suggested that this effect may only pertain to preclinical AD, wherein women showed faster rates of cognitive decline at similar (high) levels of tau pathology compared with men. 53 The association between baseline tau PET and cognitive change over time did not differ by APOE genotype.

Strengths and Limitations
The strengths of this study include the large sample size, coverage of the full AD clinical spectrum, and availability of tau PET, MRI, amyloid PET, and prospective longitudinal MMSE scores. There are also several limitations. First, MMSE served as an out-come measure because it is the only cognitive test available across all cohorts in this study. Although MMSE is a widely used measure in clinical practice and clinical trials, it is a relatively crude measure that is characterized by a ceiling effect, and the follow-up duration of this study was relatively short. Second, inherent to multicenter studies comprising multiple cohorts that were not codesigned at inception, several challenges exist regarding data harmonization and pooling. Moreover, additional complexities exist related to use of different criteria for study entry and differences in clinical assessment at each site. Similar to previous studies using this sample, 18,28-30 we minimized variability by analyzing data centrally at Lund University using a uniform pipeline, and we adjusted for cohort effects in the statistical models. However, dissimilarities in participant selection, data acquisition, and preprocessing remain. Third, despite geographical contributions from Europe, Asia, and the US, most study participants were non-Hispanic White individuals. Future studies should test whether the study findings are generalizable to more ethnically diverse populations. Fourth, we used a different tau PET tracer in the replication cohort, informed by previous studies demonstrating good correspondence between [ 18 F]flortaucipir-PET and [ 18 F]RO948-PET for neocortical tracer uptake and tau PET positivity rates. 28,54

Conclusions
In this multicenter prognostic study, the tau PET tracers [ 18 F]flortaucipir and [ 18 F]RO948 demonstrated prognostic utility as strong predictors of cognitive change over time. Tau PET outperformed established MRI and amyloid PET markers in a head-to-head comparison, especially in the Aβ-positive MCI and Aβ-positive CU groups. Our findings suggest that although tau PET as a diagnostic marker is most valuable at the dementia stage of AD, 17,18,20 the optimal time window for tau PET as a prognostic marker is during the prodromal and preclinical stages of AD.