Baseline [18F]GTP1 tau PET imaging is associated with subsequent cognitive decline in Alzheimer’s disease

The role and implementation of tau PET imaging for predicting subsequent cognitive decline in Alzheimer’s disease (AD) remains uncertain. This study was designed to evaluate the relationship between baseline [18F]GTP1 tau PET and subsequent longitudinal change across multiple cognitive measures over 18 months. Our analyses incorporated data from 67 participants, including cognitively normal controls (n = 10) and β-amyloid (Aβ)-positive individuals ([18F] florbetapir Aβ PET) with prodromal (n = 26), mild (n = 16), or moderate (n = 15) AD. Baseline measurements included cortical volume (MRI), tau burden ([18F]GTP1 tau PET), and cognitive assessments [Mini-Mental State Examination (MMSE), Clinical Dementia Rating (CDR), 13-item version of the Alzheimer’s Disease Assessment Scale-Cognitive Subscale (ADAS-Cog13), and Repeatable Battery for the Assessment of Neuropsychological Status (RBANS)]. Cognitive assessments were repeated at 6-month intervals over an 18-month period. Associations between baseline [18F]GTP1 tau PET indices and longitudinal cognitive performance were assessed via univariate (Spearman correlations) and multivariate (linear mixed effects models) approaches. The utility of potential prognostic tau PET cut points was assessed with ROC curves. Univariate analyses indicated that greater baseline [18F]GTP1 tau PET signal was associated with faster rates of subsequent decline on the MMSE, CDR, and ADAS-Cog13 across regions of interest (ROIs). In multivariate analyses adjusted for baseline age, cognitive performance, cortical volume, and Aβ PET SUVR, the prognostic performance of [18F]GTP1 SUVR was most robust in the whole cortical gray ROI. When AD participants were dichotomized into low versus high tau subgroups based on baseline [18F]GTP1 PET standardized uptake value ratios (SUVR) in the temporal (cutoff = 1.325) or whole cortical gray (cutoff = 1.245) ROIs, high tau subgroups demonstrated significantly more decline on the MMSE, CDR, and ADAS-Cog13. Our results suggest that [18F]GTP1 tau PET represents a prognostic biomarker in AD and are consistent with data from other tau PET tracers. Tau PET imaging may have utility for identifying AD patients at risk for more rapid cognitive decline and for stratification and/or enrichment of participant selection in AD clinical trials. Trial registration ClinicalTrials.gov NCT02640092. Registered on December 28, 2015

Prior studies of tau PET versus cognitive change have examined continuous associations [11,12] or dichotomized participant cohorts into low (T-) versus high (T+) tau PET [15]. Tau PET cutoffs have the potential to be used in AD clinical trials for stratification or inclusion/exclusion criteria to enrich for participants more likely to experience a subsequent decline. Existing T-/T+ distinctions have focused on diagnostic thresholds determined from cognitively normal Aβ-negative participants [20]. However, thresholds optimized for diagnostic accuracy may differ from those optimized for the prediction of cognitive outcomes.
The primary aim of this paper was to determine the prognostic utility of [ 18 F]GTP1, a second-generation tau PET tracer that has been and is being used in multiple completed and ongoing therapeutic trials in AD, for identifying individuals at risk for more rapid clinical progression. In particular, we sought to confirm that associations between cognitive change and NFT pathology measured by [ 18 F]GTP1 are consistent with those previously reported with [ 18 F]flortaucipir [11,17], [ 18 F]MK-6240 [15], and [ 18 F]RO948 [17], which would allow for future comparisons between studies using different tau PET tracers. We analyzed the relationship between baseline [ 18 F]GTP1 tau PET indices and longitudinal cognitive change over 18 months across multiple assessments in a cohort of participants ranging from cognitively normal (CN) controls through AD patients with moderate dementia. In addition, we evaluated the relative impact of defining low tau versus high tau subgroups using distribution-based and empirically derived cut points across different tau PET ROIs for predicting clinically meaningful longitudinal cognitive change on each assessment as defined by established thresholds for minimal clinically important differences (MCIDs).

Study design
Baseline and longitudinal data from an observational study evaluating longitudinal change in [ 18 F] GTP1 tau PET in CN and AD participants (GN30009; NCT02640092) were analyzed. Baseline data from this study have been included elsewhere [3,[21][22][23].

Participants
A total of 67 CN and AD participants between 50 and 85 years of age were enrolled from 11 research centers between December 2015 and November 2017. Data from participants with baseline neuroimaging and cognitive assessments and at least one post-baseline cognitive assessment were analyzed. Inclusion criteria for the CN group (n = 10) included no subjective or objective cognitive concerns, global Clinical Dementia Rating (CDR) [24] of 0, and Mini-Mental State Examination (MMSE) [25] of 28-30. In order to capture a broader range of tau PET imaging in the CN group, we included both Aβ PET-positive (by visual read [26]; n = 3) and Aβ PET-negative (n = 7) participants. In contrast, all AD participants had positive Aβ PET scans as well as brain MRIs without significant non-AD disease likely to contribute to cognitive impairment. Prodromal AD participants (n = 26) met NIA-AA criteria for mild cognitive impairment (MCI) [27] and had global CDRs of 0.5 and MMSEs of 24-30. Mild (n = 16) and moderate (n = 15) AD participants met NIA-AA criteria for probable AD dementia [28]. Mild AD participants had global CDRs of 0.5 or 1 and MMSEs of 22-30. Prodromal and mild AD were differentiated per investigators' Keywords: Tau, PET, Cognition, Alzheimer's disease, Prognosis application of NIA-AA criteria for MCI versus dementia. Moderate AD participants had global CDRs of 0.5, 1, or 2 and MMSEs of 16-21. This study was approved by each center's Institutional Review Board and conducted in accordance with International Conference on Harmonization E6 Guidelines for Good Clinical Practice. All participants and/or their legally authorized representatives provided written informed consent.

Neuroimaging
[ 18 F]GTP1 preparation and PET were performed centrally (Invicro; New Haven, CT) as previously described [3]. Images were acquired with Siemens HR+ PET or Biograph 6 PET-CT cameras during a 30-min window 60 min post-injection after a mean (SD) bolus injection of 343 (31) MBq and reconstructed with an iterative reconstruction algorithm (OSEM 4 iterations, 16 subsets) and a post hoc 5-mm Gaussian filter. Individual PET frames were motion-corrected and average [ 18 F] GTP1 images were created and co-registered to MRI, which was then spatially normalized to standard Montreal Neurological Institute space with SPM12 (www. fil. ion. ucl. ac. uk/ spm/ softw are/ spm12). Normalization parameters were applied to the corresponding average [ 18 F]GTP1 image. Composite ROIs included whole cortical gray matter (WCG), an AD-signature temporal ROI [20], and hierarchical in vivo Braak (I/II, III/ IV, V/VI) [29,30] that correspond to neuropathological NFT staging [31]. MRI tissue segmentation was performed to define the cortical gray matter, and the Hammers atlas [32] was used to define ROIs as previously described [3]. The temporal ROI included the hippocampus, amygdala, anterior medial temporal, anterior lateral temporal, parahippocampal and ambient gyri, middle and inferior temporal gyrus, and fusiform gyrus ROIs from the Hammers atlas. [ 18 F]GTP1 standardized uptake value ratios (SUVRs) were calculated using bilateral inferior cerebellar gray as reference [3]. PET data were not partial volume corrected.
[ 18 F]florbetapir was prepared at commercial facilities and Aβ PET was performed at individual sites per manufacturer instructions (Eli Lilly; Indianapolis, IN). Aβ PET SUVRs for a global cortical ROI were calculated using the whole cerebellum as reference.
MRI was performed at individual sites on 1.5T or 3T scanners. 3D sagittal T1-weighted MPRAGE sequences were collected for volumetric analyses and PET image processing as previously described [21]. Cortical volume was quantified using Freesurfer 6.0 (http:// frees urfer. net) measurements of cortical segmentation for whole cortical gray matter and adjusted for intracranial volume.

Neuropsychological testing
Cognitive assessments at baseline and 6-, 12-, and 18-month post-baseline visits included the MMSE, CDR, 13-item version of the Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog13) [33], and Repeatable Battery for the Assessment of Neuropsychological Status (RBANS) [34]. The MMSE and ADAS-Cog13 were analyzed using total scores, the CDR was analyzed using the Sum of Boxes (CDR-SB), and the RBANS was analyzed using Total Index scores.

Statistical methods
Primary statistical analyses were performed with R (v.3.3.2) [35]. Baseline comparisons between diagnostic groups were conducted using one-way ANOVAs for continuous measures and chi-squared tests for categorical variables. Post hoc analyses were performed using Tukey's test for continuous measures and Bonferroni correction for categorical variables to correct for multiple comparisons between diagnostic groups. Longitudinal cognitive changes were analyzed using estimated slopes derived from both simple linear and linear mixed effects models. For the simple linear model, a regression line was fit to each participant's cognitive scores as a function of years post-baseline to obtain an annualized slope estimate. The linear mixed effects model estimated annualized change from baseline as a function of diagnostic cohort and time post-baseline, incorporating random effects for participants. Initially, univariate associations between baseline neuroimaging and cognitive change were assessed using Spearman correlations and annualized slopes derived from the simple linear models. These univariate analyses were then confirmed with separate multiple linear mixed effects regression models to assess associations between annualized rates of change on specific assessments and baseline [ 18  Thresholds for elevated tau on [ 18 F]GTP1 PET were determined via two methods. The first was a distribution-based approach focused on distinguishing CN and AD participants. Given the small size of the CN cohort (n = 10) and slightly skewed SUVR distribution ( Fig. 1), the upper bound of SUVR range in the WCG (SUVR = 1.245) or temporal (SUVR = 1.325) ROIs in the CN group was used to dichotomize AD participants as Tversus T+. While these distribution-based cutoffs for [ 18 [36]. The second was an empirical approach that dichotomized AD participants into decliners versus non-decliners based on whether they experienced MCIDs on cognitive assessments during longitudinal follow-up. MCIDs are patient-centered thresholds that represent the smallest changes on outcome measures that are meaningful to patients at an individual level [37]. MCIDs for clinical decline have been identified as a 3 point decrease on the MMSE [38,39], a 1 point increase on the CDR-SB [40], or a 3 point increase on the ADAS-Cog [41]. While an MCID has been proposed for the RBANS Total Index [42], it was not included in these analyses, as subsequent work has questioned its utility [43]. Receiver operating characteristic (ROC) analyses with Youden's Index were used to determine [ 18 F]GTP1 SUVR thresholds that optimally discriminated between participants whose decline did or did not meet MCID criteria on the MMSE, CDR-SB, or ADAS-Cog13. Logistic regression analyses were subsequently performed to calculate odds ratios (ORs) for each distribution-based or empiric SUVR cut point reflecting the increase in odds of cognitive decline meeting MCID thresholds for AD participants classified as T+ versus those classified as T-.

Baseline participant characteristics
Baseline data for each diagnostic group are shown in Table 1. Diagnostic groups were similar in age, racial/ ethnic background, and gender distribution (ps > 0.1). A significantly higher proportion of APOE ε4 carriers were seen in the moderate AD subgroup relative to the CN group (p = 0.029). Progressively lower MMSE and RBANS scores and higher CDR-SB and ADAS-Cog13 scores were seen with increasing disease severity (ps < 0.001). Lower proportions of AD participants had complete longitudinal cognitive data than the CN group, due to a combination of slightly greater attrition and, at later disease stages, incomplete ADAS-Cog13 and RBANS data due to the inability to complete more challenging subtests. Likewise, progressively higher [ 18 F]GTP1 WCG SUVRs and lower cortical volumes were seen with increasing disease severity (ps < 0.001). Greater cortical Aβ PET SUVRs were seen in each AD subgroup relative to the CN group (ps < 0.05). Mean cortical Aβ PET SUVRs did not differ across AD subgroups (ps > 0.10).
Baseline [ 18 F]GTP1 SUVRs in WCG, temporal, and hierarchical in vivo Braak ROIs progressively increased with disease severity (Fig. 1). Significant group effects were seen in all ROIs [WCG: F(3, 63) = 10.91, p < 0.001; ]GTP1 SUVRs in the CN group did not significantly differ from the prodromal AD group (ps > 0.10), but were significantly lower than the mild and moderate AD groups (ps < 0.05). [ 18 F]GTP1 SUVRs in prodromal AD were significantly lower than in moderate AD in all ROIs (ps < 0.05) and in mild AD in all ROIs (ps < 0.05) except Braak I/II (p = 0.12). Similar [ 18 F] GTP1 SUVRs were seen between mild and moderate AD groups across all ROIs (ps > 0.05).

Longitudinal change on cognitive indices
Longitudinal change on the MMSE, CDR-SB, ADAS-Cog13, and RBANS was determined via estimated slopes derived from simple linear modeling ( Fig. 2A) and confirmed via linear mixed effects modeling (Fig. 2B). One-sample t-tests performed on estimated slopes derived from simple linear modeling indicated that MMSE, CDR-SB, and ADAS-Cog13 scores remained stable in the CN group (ps > 0.05). However, on the RBANS, the CN group improved across assessments (p = 0.035), primarily between baseline and 6-month timepoints, likely due to practice effects [44,45], and remained stable thereafter (Supplemental Fig. 1). The prodromal AD group declined on the CDR-SB (p = 0.010) and MMSE (p = 0.039), but not on the ADAS-Cog13 or RBANS (p > 0.05). The mild AD group declined on the MMSE, CDR-SB, and ADAS-Cog13 (ps < 0.001) but not the RBANS (p > 0.05). The moderate AD group declined on all indices (ps < 0.006). One-way   . CN and prodromal AD participants were more likely to have complete longitudinal cognitive data than mild or moderate AD participants (Table 1), due to increased dropout and/or inability to complete one or more ADAS-Cog13 or RBANS subtest in the latter two groups over the course of study due to disease progression.

Continuous associations between baseline [ 18 F]GTP1 SUVR and longitudinal change on cognitive indices
Univariate associations between baseline [ 18 F]GTP1 SUVR across different ROIs versus annualized cognitive change scores estimated via simple linear models are illustrated in Fig. 3 Table 1).
Multivariate regression analyses that included all participants (   Table 2), suggesting that our overall findings were not primarily driven by inclusion of the CN group in the analyses. In analogous multivariate linear regression analyses performed using other ROIs, particularly the temporal and Braak I/II ROIs, less robust associations were seen with baseline [ 18 F]GTP1 SUVRs and subsequent cognitive performance (Supplemental Tables 2 and  3). The size of the study cohort was not large enough to perform reliable mediation analyses of the relationship  Table 2 Linear mixed effects models assessing change on cognitive indices from baseline incorporating demographic, baseline cognitive, and imaging variables (using the WCG ROI and all participants). Partial regression coefficients (B), standard errors (SE), and t statistics are reported for each predictor; r 2 values are reported for each linear mixed effect model between baseline Aβ and tau PET on the subsequent longitudinal decline.

Empirical [ 18 F]GTP1 SUVR cut points derived from MCIDs
AD participants were dichotomized as decliners versus non-decliners on each assessment based on whether they experienced declines meeting MCID thresholds during longitudinal follow-up. Across assessments (Supplemental Table 4), the temporal ROI distribution-based cut point yielded good sensitivity (0.89-0.91) but poor specificity (0.43-0.59) for predicting meaningful decline, whereas the WCG ROI distribution-based cut point resulted in a greater balance between sensitivity (0.67-0.74) and specificity (0.60-0.82). ROC analyses were used to optimize empirically determined tau PET cut points for distinguishing between decliners versus non-decliners for each cognitive assessment (Fig. 5) Table 4). For the Braak ROIs, the optimal cut points were highest in Braak I/II (1.497-1.562) and progressively lower in the Braak III/IV (1.269-1.412) and Braak V/VI (1.184-1.248) ROIs (Supplemental Table 4).

Discussion
Our data indicate that higher baseline [ 18 F]GTP1 SUVRs are independently associated with steeper cognitive declines across a spectrum of AD severity and across a range of different cognitive measures. These results establish [ 18 F]GTP1 tau PET as a potential prognostic imaging biomarker in AD and complement prior work from our group demonstrating that [ 18 F]GTP1 tau PET imaging differentiates between AD cohorts of different severity [3] and exhibits significant cross-sectional correlations with cognitive performance [21]. These results are consistent with previous studies supporting the prognostic potential of other tau PET tracers, including [ 18 F]flortaucipir [10][11][12][13]17], [ 18 F]MK-6240 [15], and [ 18 F]RO948 [17]. Together, these findings reaffirm the proposed temporal relationship between tau pathology and cognitive decline [46] and establish the potential of tau PET for identifying AD patients in clinical practice who are at risk for more rapid cognitive decline and for stratification and/ or enrichment of participant selection in therapeutic clinical trials in AD. The commonality of results across different tau PET tracers suggests similar labeling of tau NFTs [4] and similar prognostic value. Prior studies examining cross-sectional tau PET and cognitive decline have focused on specific ROIs, including inferior temporal cortex [12,13], hippocampus [12], and entorhinal cortex [12,15] in preclinical AD cohorts, or composite ROIs weighted towards occipital, parietal, and lateral and posterior temporal cortices [11] or temporal and parietal cortices [16] in broader AD cohorts. In our sample, which falls into the latter category, similar correlations between baseline [ 18 F]GTP1 PET and subsequent decline were seen irrespective of the ROIs examined (Fig. 3), arguing against the overall superiority of any specific composite ROI. These results likely reflect that much of the [ 18 F]GTP1 signal in the pre-specified ROIs (except Braak V/VI) is driven by temporal lobe tau [3] given the hierarchical pattern of tau pathology in AD (i.e., significant temporal NFT deposition prior to widespread extra-temporal NFT accumulation [31]). Therefore, while specific temporal ROIs may be critical for predicting cognitive trajectories in preclinical AD, when extra-temporal NFTs are relatively scarce [12][13][14][15], more global ROIs may be sufficient at later disease stages.
Correlations between baseline [ 18 F]GTP1 SUVRs and cognitive decline were largely consistent across assessments, though less robust correlations were seen with the RBANS. This result may be attributable to greater heterogeneity in longitudinal RBANS data relative to other assessments, floor effects in moderate AD, and/ or test-retest effects masking cognitive decline in CN or prodromal AD [44,45]. Our results are concordant with prior reports of correlations between baseline tau PET and cognitive decline across a variety of composite measures [11,13,[15][16][17]. Studies in participants with normal cognition or MCI have primarily shown correlations between tau PET and declines in episodic memory [10,12], which may be indicative of the greater sensitivity of memory tests for cognitive change in preclinical and prodromal AD. This overall pattern of results suggests that tau PET is broadly prognostic for subsequent decline across a spectrum of longitudinal assessments in AD.
Another approach that may be more applicable to patient care and clinical trials in AD is determining whether threshold-based classification of individuals as having low versus high tau burden [20] has utility for predicting rates of future cognitive decline. When using thresholds optimized for AD diagnosis, high tau PET participants demonstrate steeper rates of cognitive decline, both in the current study and prior reports [15]. However, when such thresholds were used to predict clinically meaningful cognitive decline, they exhibited excellent sensitivity but mediocre specificity, particularly with the temporal ROI. Empirically defined thresholds for optimizing sensitivity and specificity for clinically meaningful decline yielded higher SUVR cutoffs, though increased specificity was offset by reduced sensitivity. Both distribution-based and empirically defined cut points for low versus high tau burden identified participants at higher risk of clinical progression that exceeded MCIDs, supporting the potential utility of these approaches. Given the continuous distribution of [ 18 F]GTP1 SUVR with increasing disease severity [3,21], these data suggest that thresholds for distinguishing high versus low tau PET are likely to depend on the specific questions of interest. In contrast, for Aβ PET, which has a more bimodal distribution between AD patients and controls [47], the same threshold for high versus low Αβ pathology may be useful for both diagnosis and prognosis [48].

Limitations
While our results confirm and extend prior work supporting tau PET as a prognostic biomarker in AD [10][11][12][15][16][17], a number of factors may limit their interpretation. The size of the study cohort was relatively small, and there were likely too few participants within each AD subgroup to detect specific prognostic relationships between tau PET and cognitive decline at different disease stages. Nevertheless, the prognostic association seen between [ 18 F]GTP1 SUVR and cognitive decline is concordant with a larger study with [ 18 F]flortaucipir [11]. Likewise, the small size of our CN cohort may have limited our ability to demonstrate statistically significant differences in [ 18 F]GTP1 SUVR relative to the prodromal AD cohort. Furthermore, the slightly skewed distribution of baseline [ 18 F]GTP1 SUVRs in the CN group precluded more direct comparisons of distribution-derived tau PET positivity thresholds to other tracers [15,20]. Higher rates of incomplete longitudinal neuropsychological data collection were seen amongst participants with more advanced AD, particularly for the RBANS. As such, our results may represent an underestimate of the correlation between baseline [ 18 F]GTP1 SUVRs and subsequent cognitive decline. Finally, as previously noted [21], we did not collect participants' level of education so we were unable to include it in our analyses. However, in prior multivariate analyses that have included education, the association between tau PET and cognitive decline remains robust [10,16].