Establishing tau-PET cut-points for cognitive diagnosis with ¹⁸F-PI-2620 in a multi-ethnoracial cohort Open Access

The primary current biomarker model of Alzheimer’s disease (AD) proposes a sequential, non-linear series of pathological changes in which amyloid beta (Aβ) becomes abnormal first but does not produce clinical deficits, followed by the spread of tau out of the medial temporal lobe to wider neocortical regions (Jack, Wiste, et al., 2018; Jack Jr. et al., 2024; Jagust, 2018; La Joie et al., 2020; Schöll et al., 2016), although the pattern and progression of pathological changes may vary across different tau subtypes (Charil et al., 2019; Ferreira et al., 2020; Murray et al., 2011). The Aβ, tau, and neurodegeneration (AT(N)) framework (Jack, Bennett, et al., 2018; Jack et al., 2024) systematizes these biological changes for research feasibility. Staging within this framework relies on determining cut-points for biomarker abnormality, but because these thresholds are influenced by the study sample, their applicability to more diverse populations may be limited (Gleason et al., 2022).

Tau accumulation is more closely linked to cognitive changes than Aβ, with higher baseline tau PET associated with steeper rates of cognitive decline over time (Hanseeuw et al., 2019; Jack et al., 2019; La Joie et al., 2020). Accordingly, establishing cut-points for tau positivity is critical for tracking disease progression, staging, and evaluating intervention success (Jack, Bennett, et al., 2018; Jack et al., 2017). Several ongoing studies are focused on characterizing later-developed tau-PET tracers, including ¹⁸F-PI-2620, ¹⁸F-MK-6240, and ¹⁸F-RO-948 (Betthauser et al., 2019; Kroth et al., 2019; Mormino et al., 2021; Pascoal et al., 2018; Wong et al., 2018). Here, we use ¹⁸F-PI-2620, one of the newer tau-PET tracers. Although, like all tau tracers, ¹⁸F-PI-2620 does show evidence of off-target binding, most notably in the anterior cerebellum and in the meninges, it benefits from lower off-target binding in the choroid plexus than the widely-used tracer, ¹⁸F-flortaucipir (AV-1451, T807, Tauvid) (Kroth et al., 2019; Mormino et al., 2021), and improved early-stage detection of pathological tau (Aliaga et al., 2024; Bullich et al., 2022; Kroth et al., 2019; Mormino et al., 2021; Mueller et al., 2020).

Previous studies have quantified abnormal tau by calculating mean tracer uptake in a “meta region of interest (ROI),” capturing a broad range of pathological change (Jack et al., 2017; Maass et al., 2017; Wang et al., 2016), and a multi-stage approach quantifying tau in specific spatiotemporal patterns based on Braak and Braak neuropathological findings (Braak & Braak, 1991). Several studies demonstrate that temporal and temporoparietal ROIs outperform neuropathological staging approaches and are reliable across tau tracers (Jack, Wiste, et al., 2018; Leuzy et al., 2020; Villemagne et al., 2023). In key regions previously evaluated in flortaucipir PET studies (Jack, Wiste, et al., 2018; Leuzy et al., 2021), we use area under the receiver operating characteristic curve (AUROC), Gaussian mixture modeling (GMM), and two standard deviations above the mean signal in cognitively-unimpaired participants, to identify cut-points that best classify participants as cognitively-impaired using ¹⁸F-PI-2620 standardized uptake value ratios (SUVRs). We identified cut-points and evaluated their performance in a subgroup of Aβ-positive cognitively-impaired (CI) and Aβ-negative cognitively-unimpaired (CU) participants (n = 675). Because of ethnoracial differences in Aβ accumulation (Garrett et al., 2019; O’Bryant et al., 2021; Wilkins et al., 2022), we also evaluated cut-point performance across the full spectrum of participants (CI and CU), regardless of Aβ status (n = 879). We use a multi-ethnoracial cohort spanning a range of socioeconomic and comorbidity profiles that are more representative of the older population than many other cohorts (Lennon et al., 2022; Raman et al., 2022).

All participants in the Health and Aging Brain Study - Health Disparities (HABS-HD) provided informed consent in accordance with the Institutional Review Boards for human research (O’Bryant et al., 2021). Participants from Data Drop 6 of HABS-HD who completed a clinical interview, had useable ¹⁸F-florbetaben and ¹⁸F-PI-2620 PET scans within 1.5 years of their cognitive assessment, passed MRI and PET quality control, had available demographic information, and had all regions within the temporal composite ROI pass quality control were included in this study (eFig. 1): 879 participants; 729 CU, 120 with mild cognitive impairment (MCI), and 30 with dementia. Most analyses presented in the main text use a subset of these participants categorized by their Aβ PET status: 675 CU Aβ-negative and 76 CI Aβ-positive participants (48 with MCI and 28 with dementia, henceforth Alzheimer’s disease; AD). HABS-HD is a collaboration among 18 universities across the United States, led by Dr. Sid O’Bryant. All imaging is performed at the University of North Texas Health Science Center, Fort Worth, Texas. Experimental methods, enrollment, inclusion, and exclusion criteria for HABS-HD have been described previously (O’Bryant et al., 2021). Race and ethnicity were self-identified: 186 were Hispanic/Latino(a) (henceforth Hispanic), 209 were non-Hispanic African American/Black (henceforth NHB), and 280 were non-Hispanic White (henceforth NHW) (Table 1).

Consensus cognitive diagnoses were established by trained clinicians (S1) and as described previously (O’Bryant et al., 2021).

T1-weighted whole-brain volumetric spoiled Magnetization-Prepared Rapid Gradient (MPRAGE) scans were acquired (3T Siemens Magnetom Skyra, voxel size 1.1 x 1.1 x 1.2 mm, n = 46; 3T Siemens Magnetom Vida system, voxel size was 1 mm x 1 mm x 1, n = 833). MRI images were bias corrected using the Advanced Normalization Tools (ANTS) with parameters of bspline-fitting = [200], shrink factor = 1, and convergence = [50 x 50 x 45 x 40, 0]. FreeSurfer version 5.3 was used for cortical parcellation and segmentation (https://surfer.nmr.mgh.harvard.edu/). We evaluated cortical thickness quality-control (QC) outcomes across all ROIs included in this study by assessing the frequency of QC failures within each ethnoracial group (S2). Aβ- and tau-PET scans were acquired on one of two identical Siemens Biograph Vision 450 scanners at the University of North Texas Health Science Center. Tau-PET imaging (5 mCi +/- 10% ¹⁸F-PI-2620 radiotracer injection; emission scan from 45-75 minutes post-injection; six 5-minute frames) and Aβ-PET imaging (8.1 mCi +/- 10% ¹⁸F-florbetaben (Neuraceq); 20-minute emission scan from 90 minutes post-injection; four 5-minute frames) shared an identical reconstruction protocol (8 iterations/5 subsets with time of flight on, a 440 x 440 matrix, and an all-pass filter).

T1 MRPAGE MRI, ¹⁸F-PI-2620, and ¹⁸F-florbetaben PET data used for the preparation of this manuscript are stored, managed, and processed by the University of Southern California Laboratory of Neuroimaging. We analyzed the scans using the ADNI3 protocol for ¹⁸F-florbetaben (S. Landau et al., 2011) and an in-house analysis protocol for ¹⁸F-PI-2620 PET. Detailed acquisition and processing methods are outlined in S3.

FreeSurfer-defined regions (frontal, anterior/posterior cingulate, lateral parietal, lateral temporal cortex; whole cerebellum reference) were used to obtain global Aβ SUVRs, with a global SUVR ≥1.08 indicating Aβ positivity (S3).

Tau-PET summary mean SUVRs were derived from FreeSurfer-defined regions and normalized to the gray matter of the inferior cerebellum (S3, eFig. 2). To check for inflation of SUVRs due to off-target signal from the meninges and venous sinuses, we also extracted median SUVR measures for all ROIs (S4). All median and mean SUVRs were highly correlated (all r values ≥0.96), suggesting minimal influence of off-target signal on mean SUVRs. The maximum time between T1 and ¹⁸F-PI-2620 PET was 17 days.

Each participant’s T1-MRI was warped to MNI152 template space using a non-linear transformation (FSL FNIRT). ¹⁸F-PI-2620 SUVR images were created by applying these transformation parameters to the corresponding ¹⁸F-PI-2620 SUVR image in T1 space (S5).

Differences in demographics, comorbid conditions, global Aβ, and cognitive measurements across diagnostic groups were assessed using Kruskal-Wallis tests followed by Dunn’s post-hoc tests for continuous variables and chi-square tests for categorical variables (Table 1).

SUVRs in regions known to be affected by AD and frequently-examined in flortaucipir tau-PET studies were evaluated (Botha et al., 2018; Cho et al., 2018; Dodich et al., 2020; Guo et al., 2020; Jack et al., 2017, 2019; Leuzy et al., 2023; Lowe et al., 2018, 2018; Maass et al., 2017; Ossenkoppele et al., 2018; Schöll et al., 2016; Schwarz et al., 2018; Wang et al., 2016; Weigand et al., 2022): entorhinal cortex, fusiform gyrus, inferior and middle temporal gyri, parahippocampal cortex, amygdala, and hippocampus. We also evaluated tau signal in the posterior cingulate and lateral parietal regions. Two composite ROIs were created using a volume-weighted average of the mean uptake in 1) a temporal composite, which included all regions listed above except for posterior cingulate and lateral parietal cortex, and 2) regions with the highest positive correlation with diagnosis in this sample: entorhinal cortex, parahippocampal cortex, and amygdala (eFig. 3; henceforth the “medial temporal” ROI).

To assess whole-brain ¹⁸F-PI-2620 patterns between diagnostic groups, voxelwise analyses were conducted using SPM-12 independent sample t-tests. Resulting T-maps were transformed to effect-size maps (Cohen’s d) for ease of interpretation (Fig. 1).

Voxelwise ROC analyses were performed by diagnostic group using the VoxelStats toolbox (Mathotaarachchi et al., 2016) to derive voxelwise AUC values (Fig. 1).

Tau cut-points were assessed using two classification methods: (A) AUROC analyses and (B) calculation of two standard deviations above the mean (2SD+) SUVR value in CU Aβ-negative participants for the best-performing ROI identified in method A. We performed AUROC analyses in two ways: 1) using the mean SUVR in each ROI and 2) using the probabilities derived by applying a two-component (bimodal) Gaussian mixture model (GMM) to each ROI to estimate the probability of each SUVR belonging to the second Gaussian (‘abnormal’ or tau-positive) distribution. We identified the best-performing ROI and cut-point as that having the highest Youden index (sensitivity + specificity-1). A Youden index of at least 0.6 with a sensitivity or specificity of at least 0.5 was considered acceptable (Chen et al., 2015). We ran 1000 stratified bootstrap replicates to obtain a 95% confidence interval on the cut-points. Accuracy, sensitivity, specificity, Youden Index, and cut-points were calculated for each ROI.

The same AUROC analyses were performed in the full spectrum of participants regardless of their Aβ status. Continuous global Aβ was included in these AUROC analyses (eTable 2).

Participants within each ethnoracial group were classified as tau positive or negative if their SUVR value in the best-performing ROI was above or below the identified cut-point for that region. We used the cut-point with the highest Youden index to classify participants as tau positive or negative. True positives were classified as any participant with a diagnosis of MCI or AD whose SUVR value in the best-performing ROI was greater than the identified cut-point. Additionally, AUROC analyses were conducted in each ethnoracial group separately (eTable 6).

Group differences are shown in Table 1.

Both MCI and AD participants showed more tau uptake in the entorhinal cortex, fusiform gyrus, and parahippocampal cortex compared to CU participants. AD participants additionally showed more binding in the inferior and middle temporal gyri, cingulate, and precuneus compared to CU and MCI groups (Fig. 1). Comparisons between diagnostic groups irrespective of Aβ status reveal similar, albeit weaker tau differences. Mean voxelwise ¹⁸F-PI-2620 signal by diagnostic group is shown in eFigure 4.

The maximum voxelwise AUC across group comparisons was 0.88. Regions with an AUC > 0.7 for all diagnostic comparisons included the fusiform gyrus, parahippocampal cortex, amygdala, and middle and inferior temporal gyri. For CU versus AD, the cingulate, precuneus, frontal, and occipital regions also exceeded an AUC of 0.7 (Fig. 1).

AUROC results showed that the medial temporal ROI had the highest Youden index for all diagnostic comparisons, and was acceptable for differentiating CU from MCI, AD, and CI. When separating CU from both AD and CI, the cut-point for this ROI was 1.26 and when separating CU from MCI it was 1.28 (Table 2). SUVRs in all regions successfully distinguished CU from AD. See eTable 1 for comprehensive results.

AUROC results using predicted probabilities (eTable 2) also showed that the medial temporal ROI had the highest Youden index across diagnostic classifications. AUROC using predicted probabilities in all regions successfully distinguished CU from AD.

Using the 2SD+ method, a cut-point of 1.32 was determined for the medial temporal ROI (Table 2), with the only acceptable Youden index distinguishing CU versus AD (>0.60).

Of the 879 participants, 204 did not meet the Aβ-stratification criteria, including 130 Aβ-positive CU participants (14% Hispanic, 14% NHB, and 23% NHW), 72 Aβ-negative MCI participants, (73% Hispanic, 53% NHB, and 51% NHW), and 2 Aβ-negative participants with dementia (1 Hispanic; 1 NHW). Aβ status by diagnostic and ethnoracial group is shown in eTable 3.

In the full spectrum (eTable 4), AUROC analyses showed that continuous global Aβ best differentiated CU from dementia, with the highest Youden index (0.76). Discrimination for all measures was weaker for CU versus MCI and CU versus CI, with Youden indexes below 0.35, suggesting that tau cut-points are effective in explaining cognitive status predominantly in Aβ-positive individuals.

Demographics by ethnoracial group are shown in Table 3. Mean SUVR by ethnoracial group can be found in eTable 5. The 1.26 cut-point performed acceptably in Hispanic and NHW participants, but not in NHB participants (Table 4). When AUROC were fit in each ethnoracial group separately, the medial temporal composite consistently showed the highest Youden index (eTable 6).

Using ¹⁸F-PI-2620 we have identified an ROI and corresponding cut-point that differentiates cognitive status in a multi-ethnic/racial cohort. The medial temporal composite had the highest Youden index across all diagnostic classifications, indicating that this composite was sensitive to the earliest clinically-relevant tau accumulations in this sample. The high specificity of tau in this region suggests that here: 1) the presence of tau is not overestimated and 2) the presence of tau is a good indicator of cognitive impairment. However, studies have shown that tau accumulation in the entorhinal cortex slows as disease progresses (Jack, Wiste, et al., 2018; Villemagne et al., 2021), necessitating future research to evaluate the effectiveness of this ROI in predicting longitudinal changes in cognition and tau accumulation. The sensitivity observed across all ROIs was notably low, particularly in NHB participants. This may indicate the influence of other factors that contribute more to cognitive impairment in NHB participants than tau alone (Harrison et al., 2025). When all participants were included (irrespective of Aβ status) Aβ was more effective than tau at distinguishing cognitive impairment. This suggests that 1) in our multi-ethnic cohort, consistent with prior work (Robinson et al., 2024), tau in the absence of Aβ positivity does not meaningfully explain cognitive impairment and 2) tau positivity using these cut-offs should only be used to identify clinically-meaningful tau levels in Aβ-positive individuals.

In past work in NHW participants, 54% of MCI patients and 81.4% of dementia patients were Aβ positive, similar to the 49% and 92% Aβ-positive NHW MCI and dementia patients we found here, respectively. However, also as reflected in our work, Aβ-positivity rates may vary by ethnoracial group (Jang et al., 2024). In cognitively-impaired Aβ-positive individuals, the presence of tau within the normal range represents an atypical disease trajectory, with approximately 20–40% of Aβ-positive individuals remaining tau-negative depending on cohort composition and diagnostic criteria (Josephs et al., 2022; S. M. Landau et al., 2024; Spina et al., 2021) consistent with the 30% tau negativity in Aβ-positive cognitively-impaired participants in our study. For instance, past work using flortaucipir found that tau positivity in the entorhinal cortex was 80.2% in Aβ-positive dementia patients, 50.2% in Aβ-positive MCI patients, and 6.3% in cognitively-unimpaired people (Ossenkoppele et al., 2021). Our tau-positivity rate in Aβ-positive dementia patients was similar at 75%. However, tau positivity was higher in our sample for Aβ-positive MCI patients (67%) and CU participants without respect to Aβ status (12%). The relatively high number of Aβ-negative MCI cases among Hispanic and NHB participants further supports the possibility that non-amyloid, non-tau mechanisms may play a role in disease progression among these groups, warranting further investigation. Future work should also evaluate the extent to which the region of interest used, the ethnoracial composition of the sample, or the tracer used affects tau-positivity status.

Several studies have found that NHB individuals who achieve low scores on neuropsychological exams have lower Aβ-positivity rates than NHW participants (Carrillo & Mahinrad, 2024; Morris et al., 2019; Raman et al., 2021), suggesting that cognitive impairment or clinical symptoms may be more strongly influenced by non-AD pathways, including cerebrovascular injury—particularly given the higher prevalence of cardiovascular disease in NHB populations (Hines et al., 2023). However, lower neuropsychological test scores may reflect socioeconomic and educational disparities rather than AD pathology, potentially resulting in overdiagnosis of NHB individuals with MCI when a single time point is considered (Raman et al., 2021). To address this, we compared the cut-point performance between participants with higher education and those with a high-school degree and observed no significant difference in AUROC between the groups (eTable 7). NHB participants in the current study are also, on average, younger than the Hispanic and NHW participants, which may increase the likelihood that cognitive impairment is attributable to pathologies other than Aβ and tau. When available, future research in diverse populations should incorporate additional information about comorbidities and other neuropathologies to determine whether these factors influence cut-point values or their utility.

Although use of the 2SD+ method resulted in a more conservative cut-point and similar accuracy to ROC, it resulted in a Youden index that was lower for all diagnostic group comparisons. The AUROC method incorporates clinical relevance (through sensitivity and specificity) to determine the optimal cut-points. While the 2SD+ method provides a stricter threshold, AUROC may be preferable for identifying clinically-meaningful distinctions between groups. We recognize that using the Youden index prioritizes overall diagnostic accuracy by maximizing the sum of sensitivity and specificity but may overlook context-specific priorities where the costs of false positives and false negatives differ.

The strengths of this study include the multi-ethnoracial composition of the sample and the novelty of ¹⁸F-PI-2620. Although the heterogeneity of the sample affected sensitivity, the results point to a gap in knowledge as to what other factors influence the clinical phenotype in more diverse samples. We used the cut-point obtained from the full model and applied it to each ethnoracial group, acknowledging limitations due to key differences between the three groups. However, we believe this approach demonstrates the challenges of defining pathological protein levels within heterogeneous samples. Additionally, similar results were observed when cut-points were calculated independently for each group, though the smaller sample sizes in some groups limit the robustness and generalizability of these findings. Of note, the prevalence of cognitively-impaired Hispanic participants in this sub-sample was lower than in the larger HABS-HD sample. In the full spectrum (regardless of Aβ status), we included 72 new Aβ-negative participants with MCI across ethnoracial groups but only added two Aβ-negative participants with dementia. As a result, nearly all participants with dementia were Aβ positive, which likely explains why Aβ status was a stronger predictor of dementia than tau in our full sample, despite previous research suggesting that tau correlates more closely with cognitive status than Aβ. Finally, the cross-sectional nature of our study should be considered when interpreting the results.

We have established tau cut-points for cognitive diagnosis using ¹⁸F-PI-2620. Our data demonstrate the utility of the medial temporal ROI to identify clinically relevant tau levels in Hispanic and NHW individuals.

Data used in preparation of this manuscript preparation are stored, managed, and processed by the University of Southern California (USC) Laboratory of Neuroimaging (https://ida.loni.usc.edu/). All data are available upon request.

V.R.T.: Conceptualization, Data Collection, Formal Analysis, Data Curation, Writing—Original Draft, and Visualization. K.V.W.: Data Analysis, Data Curation. N.N.L., J.A.T., M.W.H., S.G., P.W., A.G., T.B., T.L.: Data Curation. R.R., R.A.R., B.T.C., M.P., A.D.C., B.M.A., K.L.M., Z.Z., R.R.N., K.Y., S.E.O., and A.W.T.: Methodology, Review & Editing. M.N.B.: Conceptualization, Methodology, Validation, Resources, Writing—Review & Editing, and Supervision.

The authors declare no competing interests.

Research reported on this publication was supported by the National Institute on Aging of the National Institutes of Health under Award Numbers R01AG054073, R01AG058533, R01AG070862, P41EB015922, U19AG078109 and by the Office of The Director, National Institutes Of Health of the National Institutes of Health under Award Number S10OD032285. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. We thank the study participants for their time and generosity and the HABS-HD staff for their hard work. We are grateful to the HABS-HD study team: HABS-HD MPIs: Sid E O’Bryant, Kristine Yaffe, Arthur Toga, Robert Rissman, & Leigh Johnson; and the HABS-HD Investigators: Meredith Braskie, Kevin King, James R. Hall, Melissa Petersen, Raymond Palmer, Robert Barber, Yonggang Shi, Fan Zhang, Rajesh Nandy, Roderick McColl, David Mason, Bradley Christian, Nicole Phillips, Stephanie Large, Joe Lee, Badri Vardarajan, Monica Rivera Mindt, Amrita Cheema, Lisa Barnes, Mark Mapstone, Annie Cohen, Amy Kind, Ozioma Okonkwo, Raul Vintimilla, Zhengyang Zhou, Michael Donohue, Rema Raman, Matthew Borzage, Michelle Mielke, Beau Ances, Ganesh Babulal, Jorge Llibre-Guerra, Carl Hill, and Rocky Vig.

Supplementary material for this article is available with the online version here: https://doi.org/10.1162/IMAG.a.41