22 Novembre 2026

Psicologia Punto One | Psicologia e Psicologi nel Web

Menu

Menu

SCHEDA SCIENTIFICA2026Altra ricerca

A model of brain function to predict tau across Alzheimer's disease and related disorders.

Nick Corriveau-Lecavalier, Gemeng Zhang, Hugo Botha, Leland R Barnard, Heather J Wiste, Ellen Dicks, Venkatsampath R Gogineni, Jeyeon Lee, Irene Sintini, Matthew L Senjem, Christopher G Schwarz, Petrice M Cogswell, Derek R Johnson, Jeffrey L Gunter, Michael G Kamykowski, Jonathan Graff-Radford, Vijay K Ramanan, Julie A Fields, Mary M Machulda, Nikki H Stricker, Bradley F Boeve, Val J Lowe, Kejal Kantarci, David S Knopman, Jennifer L Whitwell, Keith A Josephs, Ronald C Petersen, Clifford R Jack, David T Jones, or the Alzheimer’s Disease Neuroimaging Initiative

Nature communications · 2026

Abstract originale

We developed a machine learning framework grounded in a computational model of brain function and FDG-PET imaging to predict Alzheimer's disease (AD)-type tau across AD and related disorders (ADRD). The model was trained on 1088 research participants spanning normal aging and ADRD with FDG-PET and flortaucipir-PET (FTP-PET). Validation was conducted in: (1) a longitudinal subset of the training cohort (n = 215), (2) Alzheimer's Disease Neuroimaging Initiative (n = 295), and (3) a heretogeneous clinical cohort (n = 264; n = 28 with FTP-PET). Results showed that the model accurately predicted FTP-PET when applied at the participant-level (R2 ~ 0.34-0.71). It performed well in clinical scenarios, including classification of low versus high tau stages (79-95% accuracy) and prediction of tau in a heterogeneous clinical cohort without FTP-PET. These findings advance the understanding of the relationship between neurodegeneration and AD-type tau and provide a framework for estimating AD-type tau across ADRD.

Parole chiave

HumansAlzheimer Diseasetau ProteinsPositron-Emission TomographyBrainFemaleAgedMaleAged, 80 and overFluorodeoxyglucose F18Machine LearningCarbolinesPredictive Learning ModelsNeuroimaging
Leggi il testo integrale originale

Testo estratto dalla pubblicazione: sono esclusi immagini, media, supplementi e parti soggette a diritti distinti.

A model of brain function to predict tau across Alzheimer’s disease and related disorders

We developed a machine learning framework grounded in a computational model of brain function and FDG-PET imaging to predict Alzheimer’s disease (AD)-type tau across AD and related disorders (ADRD). The model was trained on 1088 research participants spanning normal aging and ADRD with FDG-PET and flortaucipir-PET (FTP-PET). Validation was conducted in: (1) a longitudinal subset of the training cohort (n = 215), (2) Alzheimer’s Disease Neuroimaging Initiative (n = 295), and (3) a heretogeneous clinical cohort (n = 264; n = 28 with FTP-PET). Results showed that the model accurately predicted FTP-PET when applied at the participant-level (R2 ~ 0.34-0.71). It performed well in clinical scenarios, including classification of low versus high tau stages (79–95% accuracy) and prediction of tau in a heterogeneous clinical cohort without FTP-PET. These findings advance the understanding of the relationship between neurodegeneration and AD-type tau and provide a framework for estimating AD-type tau across ADRD.

Researchers developed a machine learning model using FDG-PET imaging to accurately predict Alzheimer’s-type tau pathology across spectrums of aging and dementia, providing a novel framework to estimate tau burden non-invasively.

Introduction

The trans-modal prediction of Alzheimer’s disease (AD)-type tau across aging and AD and related disorders (ADRD) has become a major focus of research since the advent of positron emission tomography (PET) radioligands sensitive to this proteinopathy1. This effort is motivated by both scientific and practical impetuses. Scientifically, linking imaging modalities that capture distinct biological processes may clarify the mechanisms driving AD-type tau pathology and its downstream neurodegenerative effects. Practically, accurate prediction of this pathology may offer a cost-effective approach to enhance the efficiency of biological staging, prognostication, and longitudinal tau assessment.

Most trans-modal imaging models of AD-type tau propagation and neurodegeneration have relied on network-based frameworks, commonly referred to as the “prion-like” spreading hypothesis2. This hypothesis posits that functional and structural brain networks provide conduits through which pathogenic proteins propagate. In this framework, regions that are strongly connected to areas of high pathology (“epicenters”) are more likely to accumulate pathology, whereas weakly connected regions are less vulnerable. Ample empirical evidence supports this model in AD3–7, including its atypical variants8–10, as well as in primary tauopathies6,11, and neurodegenerative processes in non-AD syndromes, such as frontotemporal dementia (FTD)12–14. However, these studies are most often designed to simulate tau propagation dynamics at the region-level and typically not intended to be applied at the participant-level. Moreover, most of these investigations have relied on functional resting-state MRI, which, though valuable for probing brain physiology, is a noisy measure that constrains predictive accuracy and reliability at the participant-level15. It has also been suggested that the prion-like spreading hypothesis, while informative, may provide only a partial mechanistic account of tau propagation and neurodegeneration16.

Emerging computational models of brain function may offer additional insight into disease mechanisms and potentially improve predictive accuracy and clinical applicability17. It has been proposed that the brain’s functional architecture is structured along a hierarchy of large-scale gradients, which are continuous, low-dimensional axes that capture dominant patterns of spatial variation in brain function18. These gradients align with multiple biological features, including gene expression profiles19,20, neurotransmitter systems21, and metabolic regulation22. Importantly, large-scale gradients dynamically coordinate cognition17,23,24 and map onto the anatomical systems that support specific mental functions that are selectively degenerated across dementia syndromes25,26. Therefore, selective vulnerability to age-related proteinopathies such as AD-type tau may be embedded within these gradients25,27,28. We have previously demonstrated25 that a small set of gradients derived from fluorodeoxyglucose (FDG)-PET could map most common dementia syndromes within a global functional state space, with subspaces within this latent projection corresponding to the anatomical systems targeted in these disorders. Subsequent gradient-based approaches have since been applied in AD, helping to explain inter-individual variability in AD-type tau deposition and clinical symptoms27,29.

In this study, we developed a scalable and interpretable model of AD-type tau prediction based on FDG-PET. In this study, we developed a scalable and interpretable model of AD-type tau prediction based on FDG-PET. We selected FDG-PET as our primary imaging modality because it is currently utilized in clinical practice for the evaluation of degenerative cognitive disorders to clarify differential diagnosis, index disease severity, and track progression30,31. Our model builds on StateViewer32, a clinical decision support system for the classification of degenerative dementia syndromes. StateViewer is derived from a principal component decomposition of 1000 randomly sampled FDG-PET scans and has demonstrated high classification accuracy across aging and nine neurodegenerative dementia syndromes (n = 3671), including robust out-of-sample performance. Here, we leveraged this existing framework to train our FDG-PET model of AD-type tau prediction in a large database of participants drawn from three Mayo Clinic research programs spanning the aging, AD, dementia with Lewy bodies (DLB), and FTD spectra (referred to as the “Mayo research cohort”) with Flortaucipir (FTP)-PET. The model is best framed as a population-level regression framework that leverages shared variance across a large cohort to estimate participant-level spatial distributions of FTP-PET burden. We validated the model in (1) participants from a Mayo Clinic research cohort with available longitudinal FDG-PET and FTP-PET data, (2) an external research dataset from the Alzheimer’s Disease Neuroimaging Initiative (ADNI)33, and (3) a clinically heterogeneous cohort of patients drawn from the Mayo Alzheimer’s Disease Treatment Clinic (ADTC)34, including a subset with available research FTP-PET. Figure 1 provides an overview of the analytic workflow.

In addition, we evaluated model performance across several clinically relevant tasks: (1) prediction of tau staging according to the Alzheimer’s Association’s revised criteria for the diagnosis and staging of AD35 in the training and validation cohorts; (2) prediction of tau stages in clinical syndromes with a high likelihood of co-pathologies (i.e., TDP-43 in amnestic mild cognitive impairment [aMCI] and amnestic AD [aAD])36–38 or that can result from either primary AD or non-AD pathologies (e.g., TDP-43 or tauopathies in behavioral variant FTD [bvFTD] or corticobasal syndrome [CBS])39–42. and (3) generation of biologically plausible tau patterns in clinically heterogeneous patients from the Mayo ADTC regardless of FTP-PET availability.

Results

Characterization of study cohorts

Study cohorts are described in Table 1. The Mayo research cohort (n = 1088) was used as the training set and included participants from research programs at Mayo Clinic Rochester (Mayo Clinic AD Research Center [ADRC], Mayo Clinic Study of Aging [MCSA], Neurodegeneration Research Group [NRG]) with amyloid-, FTP-, and FDG-PET. This cohort covered the AD, FTD, and DLB spectrum, in addition to aMCI, non-amnestic MCI (naMCI), and cognitively unimpaired (CU) participants. Overall, the cohort recapitulates known phenotype-specific features, for instance, a younger age in atypical AD phenotypes (dysexecutive AD [AD], logopenic variant primary progressive aphasia [lvPPA], posterior cortical atrophy [PCA]) and bvFTD, a male predominance in DLB, and a higher proportion of APOE4 carriership in amnestic AD phenotypes (e.g., aMCI and amnestic AD [aAD]). A subset of the Mayo research cohort with longitudinal concurrent FTP-PET and FDG-PET was used as a validation cohort (n = 215, 478 time points total).Table 1Characterization of the Mayo Clinic research, ADNI, and ADTC cohorts

Group n Age Female (%) Education (years) CDR-SB APOE4 (%) A+ (%) T+ (%) Longitudinal PET (n)

Mayo research cohort

CU A-T- 477 66 [57–73] 48 16 [14–17] 0 [0–0] 26 0 0 111

CU A + T– 144 75 [69–81] 50 15 [13–17] 0 [0–0] 40 100 0 30

CU A + T+ 65 79 [73–85] 49 16 [14–16] 0 [0–0] 34 100 100 12

aMCI 29 75 [72–82] 31 16 [15–19] 1.5 [0.5–3.5] 69 100 76 10

naMCI 24 77 [72–81] 46 14 [12–16] 1 [0–2.6] 42 92 21 5

aAD 68 76 [71–80] 53 16 [14–18] 4 [1.5–6] 68 100 75 20

dAD 21 61 [57–65] 62 14 [14–16] 7 [5–9.5] 86 100 100 8

lvPPA 35 69 [61–74] 69 16 [13–16] 2.5 [1.5–13] 37 100 97 5

PCA 44 65 [61–72] 68 16 [14–16] 6.5 [3.3–7.5] 46 100 100 3

DLB 32 72 [67–80] 9 16 [12–17] 3.5 [2–5.6] 34 53 47 7

bvFTD 18 68 [61–74] 44 16 [13–18] 6.8 [2.6–10] 29 29 29 1

agPPA 41 70 [61–76] 61 16 [14–18] 1.8 [0.9–2.6] 32 24 17 1

PSP 51 71 [66–76] 41 17 [17–17] 0 [0–0] 24 14 14 0

SD 17 68 [62–73] 41 18 [16–18] 6.5 [4.4–8] 29 35 41 1

CBS 22 68 [64–73] 41 16 [12–16] 1 [0.3–2.8] 14 32 14 1

ADNI cohort

CU 13 68 [66–72] 54 16 [14–18] 0 [0–0] 46 39 0 –

MCI 211 74 [68–79] 44 16 [14–18] 1 [0.5–2] 39 46 42 –

Dementia 71 77 [71–81] 42 16 [14–18] 4.5 [3.5–5.8] 65 76 70 –

ADTC cohort

AD A+ 138 73 [69–76] 59 16 [14–18] 2 [1.3–3.5] 80 100 – –

Atypical AD A+ 64 62 [59–68] 55 16 [14–18] 3 [1.5–5] 55 100 – –

Non–AD A– 11 75 [72–78] 54 16 [14–18] 1 [0.5–3.5] 43 0 – –

Non–AD A+ 28 75 [72–78] 64 16 [13–16] 1 [0.6–2.1] 55 100 – –

AD A+ co–path 23 79 [76–81] 44 16 [14–18] 3 [1.3–4.5] 70 100 – –

Values are expressed as median with IQRs 25–75, percentage, or counts. A Amyloid, aAD amnestic Alzheimer’s disease, ADNI Alzheimer’s Disease Neuroimaging Initiative, ADTC Alzheimer’s Disease Treatment Clinic, agPPA Agrammatic primary progressive aphasia, aMCI amnestic mild cognitive impairment, bvFTD Behavioral variant frontotemporal dementia, CBS Corticobasal syndrome, CDR-SB clinical dementia rating sum of boxes, CU cognitively unimpaired, dAD dysexecutive Alzheimer’s disease, DLB Dementia with Lewy Bodies, IQR interquartile range, lvPPA Logopenic variant primary progressive aphasia, MCI Mild cognitive impairment, naMCI Non-amnestic mild cognitive impairment, PET Positron emission tomography, PCA posterior cortical atrophy, PSP Progressive supranuclear palsy, SD Semantic dementia, T +  Tau, PET positron emission tomography.

The ADNI cohort (n = 295) included CU, MCI, and Dementia participants with available amyloid-PET, FTP-PET, and FDG-PET imaging. This cohort was further classified according to amyloid and tau status for some of the analyses.

The ADTC cohort (n = 264) included amyloid-positive (A + ) patients with early symptomatic AD (“AD A + ”), early symptomatic AD with a high suspicion of co-pathology (e.g., α-synucleinopathy or TDP-43; “AD A+ co-path”), atypical AD variants (dAD, lvPPA, PCA, motor AD, or behavioral AD; “Atypical AD A + ”), and non-AD syndromes (e.g., normal pressure hydrocephalus, Parkinson’s disease, vascular dementia, frontotemporal dementia; “Non-AD”). The Non-AD group was further stratified by amyloid status into “Non-AD A + ” (i.e., non-AD symptom-driving etiology with positive amyloid testing) and “Non-AD A–” (i.e., non-AD symptom-driving etiology with negative amyloid testing). A subset of the ADTC cohort (n = 28) had research FTP-PET acquired as part of concomitant research programs (MCSA, ADRC; not included in the training cohort).

FDG- and FTP-PET distribution across the ADRD spectrum in the Mayo research cohort

FDG-PET standard uptake value ratios (SUVR) distributions (Fig. 2) followed expected patterns, with CU groups reflecting natural patterns of age-related metabolism with higher SUVR values in metabolically demanding areas, including the precuneus, and lower values in regions with low metabolic demands such as the temporal pole and sensorimotor stip22. All symptomatic groups showed relatively diffuse hypometabolism throughout the cortex (i.e., lower SUVR values), with a predominance towards lateral temporal areas in aMCI and aAD, heteromodal cortex in dAD, left temporoparietal areas in lvPPA, occipito-temporal in PCA and DLB, temporal poles bilaterally in semantic dementia (SD), superior frontal and temporal areas in agrammatic PPA (agPPA), medial frontotemporal areas in bvFTD, medial frontal areas in progressive supranuclear palsy (PSP), and sensorimotor areas in CBS.

The extent and distribution of group-level observed FTP-PET SUVR values (Fig. 3) also followed expected patterns, with greater uptake in the lateral temporal and medial parietal areas with more advanced disease stage from CU A + T+ to aMCI to aAD groups, whereas tau-negative CU groups had minimal uptake throughout the cortex. Atypical AD groups had the highest uptake that followed phenotype-specific patterns: the dAD group showed extensive uptake across the heteromodal cortex with relative sparing of the primary sensory and motor areas, the lvPPA group showed uptake in lateral temporoparietal areas with a leftward hemispheric predominance, and the PCA group showed prominent occipito-temporal uptake bilaterally. The naMCI and non-AD syndromes (DLB, SD, agPPA, bvFTD, PSP, CBS) showed minimal uptake throughout the cortex.

Training of FDG-PET eigenbrain model in the Mayo research cohort

We projected the Mayo research cohort onto all principal components (eigenbrains; n = 150) from StateViewer32, an independent latent space model derived from 1000 randomly sampled FDG-PET images from Mayo Clinic. The first ten eigenbrains are shown in Supplementary Fig. 1. This yielded participant-level eigenvalues for each eigenbrain across the cohort. We then fit a series of linear regression models relating these eigenvalues to FDG-PET uptake in each region of the Schaefer-400 atlas. Models were iteratively tested with an increasing number of eigenbrains, and the optimal model was selected based on the highest overall leave-one-out R2pred value. This was applied at the participant level, and we computed participant-level spatial R2 values across the atlas. Linear models applied across participants were benchmarked against five non-linear and regularized machine learning models (detailed in the Methods section), and a base model using age, sex, APOE4 carriership, clinical diagnosis, and clinical dementia rating sum-of-boxes (CDR-SB) as predictors of FTP-PET instead of FDG-PET eigenbrains.

Additional regression models were performed across participants between eigenvalues and CenTauR values (mesial temporal tau, temporoparietal tau, temporal tau) to determine predicted tau stage based on the number of eigenbrains used in the final Schaefer-400 model. Clinically relevant case examples are presented, where observed versus predicted tau stages/burden were compared in aMCI, aAD, bvFTD, and CBS groups.

The best-performing model included 50 eigenbrains as predictors of regional FTP-PET uptake across participants, achieving median R2 = 0.71 [IQR 0.61 to 0.77], R2adj = 0.69 [IQR 0.59 to 0.76], R2pred = 0.65 [IQR 0.54 to 0.73], mean absolute error (MAE) = 0.09 [IQR 0.08 to 0.09], and mean squared error (MSE) = 0.01 [0.01 to 0.02] across the Schaefer-400 atlas. Participant-level spatial comparisons of observed and predicted FTP-PET values across the 400 regions of the atlas achieved a median value of R2 = 0.79 [0.73 to 0.84].

Figure 3 depicts predicted values from the model, which show relatively similar patterns across groups compared to observed FTP-PET uptake values. Figure 4 (panel A) shows observed and predicted FTP-PET values in six region examples. Supplementary Fig. 2 shows performance metrics across Yeo’s networks43, and Supplementary Fig. 3 shows three-dimensional brain renderings of R2 values for each group, with markedly higher values in groups with a wider FTP-PET dynamic range (e.g., dAD, PCA) than groups with minimal FTP-PET signal (e.g., CU groups, SD, PSP). Supplementary Fig. 4 shows Bland-Altman plots across all diagnostic groups. While the overall mean predicted values generally fall within the 95% limits of agreement of the observed values, high-tau groups (i.e., aAD, dAD, and lvPPA) exhibit a slight degree of proportional bias, characterized by overestimation at lower mean values and underestimation at higher mean values.

Repeating the analyses accounting for age, sex, and APOE4 status did not improve model performance, achieving median R2 = 0.71 [IQR 0.62 to 0.78], MAE = 0.09 [IQR 0.08 to 0.09], and MSE = 0.1 [IQR 0.01 to 0.2]. Therefore, further analyses were performed without covariates. However, in order to explicitly assess performance across sexes, R2 values from the main model were compared between males and females across the Schaefer-400 atlas. This yielded higher values in females (R2 = 0.75 [IQR 0.66–0.81]) compared to males (R2 = 0.61 [IQR 0.5–0.70]).

Comparisons with alternative non-linear and regularized models revealed lower performance, with R2 values ranging from 0.59 to 0.65. The base model including age, sex, APOE carriership, diagnosis, and CDR-SB also achieved lower performance, with an R2 value of 0.51. Full performance metrics can be found in Supplementary Table 1.

The regression model between FDG-PET eigenvalues and CenTauR values yielded R2 = 0.59 [IQR 0.54 to 0.60], MAE = 0.23 [IQR 0.22 to 0.25], and MSE = 0.11 [IQR 0.10 to 0.12] across the mesial temporal, temporoparietal, and temporal meta-ROI regions across participants. A confusion matrix excluding A- participants (as this stage is determined by amyloid status only) achieved an overall accuracy of 88%. In total, 290 cells were correctly labeled as A/B, 134 cells were correctly labeled as C/D, 20 cells were incorrectly labeled as A/B, and 36 cells were incorrectly labeled as C/D. Re-running the model including A- individuals achieved an overall accuracy of 95%. Full performance metrics are shown in Supplementary Table 2.

The case examples in aMCI, aAD, bvFTD, and CBS groups showed an overall accuracy of 82% when excluding A- individuals, with 42 cells correctly labeled as A/B, 47 cells correctly labeled as C/D, 7 cells incorrectly labeled as A/B, and 13 cells incorrectly labeled as C/D. Including A- individuals yielded an overall accuracy of 85%. Full performance metrics are shown in Supplementary Table 3.

Group-wise comparisons of predicted CenTaur values achieved significant results for the mesial temporal lobe F(1, 107) = 41.3, P < 0.001, temporoparietal area F(1, 107) = 42.3, P < 0.001, and temporal meta-ROI F(1, 107) = 46, P < 0.001. Tukey post-hoc analyses indicated higher predicted values in individuals with an observed C/D stage compared to observed A- and A/B stages (P < 0.001 in all cases), and higher values in A/B stage compared to A- (P < 0.05). Results from this analysis are shown in Fig. 4 (panel B). Supplementary Fig. 5 visually depicts predicted CenTaur values across all groups, stratified by predicted tau stage.

Validation in the longitudinal Mayo research cohort

The first set of validation analyses involved the participants from the Mayo research cohort with longitudinal FTP-PET and FDG-PET data. Predicted Schaefer-400 and CenTauR values were extracted using the model trained on baseline data and StateViewer-derived eigenvalues at each time point for each participant. Mixed-effects models between observed and predicted values across the Schaefer-400 atlas yielded a marginal R2 = 0.51 [IQR 0.43 to 0.67], MAE = 0.03 [IQR 0.03 to 0.03], and MSE = 0.001 [IQR 0.001 to 0.002] across participants. Participant-level spatial comparisons of observed and predicted FTP-PET values across visits achieved a median value of R2 = 0.78 [0.71 to 0.83].

The regression model between FDG-PET eigenvalues and CenTauR values yielded a marginal R2 = 0.57 [0.49 to 0.58], MAE = 0.06 [0.06 to 0.07], and MSE = 0.01 [0.01 to 0.01] across participants. A confusion matrix at baseline that excludes A- participants achieved an overall accuracy of 86%, with 64 cells correctly labeled as A/B, 18 cells correctly labeled as C/D, 3 cells incorrectly labeled as A/B, and 10 cells incorrectly labeled as C/D. Including A- individuals resulted in an accuracy of 94%. A confusion matrix comparing observed and predicted tau stage transitions between the first and last FTP-PET scans for each participant while excluding A-participants achieved an overall accuracy of 80%. A confusion matrix including A- participants achieved an accuracy of 91%. Longitudinal observed versus predicted CenTauR temporal meta-ROI and confusion matrices are shown in Fig. 5. Performance metrics for all confusion matrices can be found in Supplementary Table 4.

Analyses were repeated using a baseline training model without participants with longitudinal data to assess whether overlap between the training and validation set influences performance. Performance of mixed-effect linear models across the Schaefer-400 atlas across participants yielded lower performance, with a marginal R2 = 0.39 [IQR 0.33–0.46], MAE = 0.03 [IQR 0.03–0.03], and MSE = 0.002 [IQR 0.001 to 0.002]. Performance of mixed-effect linear models across the CenTauR regions yielded marginal R2 = ~ 0.33–0.47, MAE = ~0.06–0.07, and MSE = ~0.007–0.009. However, this had no impact on tau staging classification results, which remained identical to the original model.

Finally, we evaluated whether baseline prediction residuals between predicted and observed FTP-PET in the CenTauR temporal meta-ROI carry prognostic value by assessing their relationship with change (Δ) in CDR-SB and the change (Δ) in CenTauR temporal meta-ROI between the baseline visit and the final follow-up visit for each participant. This yielded a small yet significant relationship between baseline residuals and Δ CDR-SB, ρ = 0.231, P = < 0.001 and between baseline residuals and Δ temporal meta-ROI, ρ = 0.125, P = 0.04. This indicates that individuals for whom FTP-PET burden was underestimated at baseline tended to clinically worsen and accumulate more tau over time. These relationships are visually depicted in Supplementary Fig. 6.

Validation in the ADNI cohort

The second set of validation analyses involved out-of-sample prediction in the ADNI cohort. Predicted Schaefer-400 and CenTauR values were extracted using the model trained in the Mayo research cohort and StateViewer-derived eigenvalues in ADNI participants. Mixed-effect models between observed and predicted FTP-PET across the Schaefer-400 atlas achieved R2 = 0.33 [IQR 0.21 to 0.41], MAE = 0.11 [IQR 0.10 to 0.12], and MSE = 0.02 [IQR 0.02 to 0.03] across participants. Individual spatial R2 across the cohort achieved a median value of R2 = 0.72. Figure 6 shows renderings of predicted FTP-PET values (panel A), group-wise comparisons (panel B), and observed and predicted FTP-PET values in six region examples (panel C).

The regression model between FDG-PET eigenvalues and CenTauR values yielded R2 = 0.44 [IQR 0.42 to 0.44], MAE = 0.27 [IQR 0.26 to 0.30], and MSE = 0.12 [IQR 0.11 to 0.15]. Comparisons between observed and predicted tau burden with combined stages A/B and C/D without A- individuals achieved an accuracy of 79%, with 82 cells correctly labeled as A/B, 41 cells correctly labeled as C/D, 12 cells incorrectly labeled as A/B, and 22 cells incorrectly labeled as C/D. Analyses including A- individuals achieved an accuracy of 89%. Performance metrics for all analyses can be found in Supplementary Table 5.

Comparisons of predicted CenTauR regions across clinically- and biomarker-defined groups showed a significant effect for the mesial temporal area, F(9, 285) = 21.5, P < 0.001, temporoparietal area, F(9, 285) = 12.2, P < 0.001, and temporal meta-ROI, F(9, 285) = 15.3, P < 0.001. Post-hoc analyses revealed that, across all regions, the Dementia A + T+ group had higher values than the CU A-T-, CU A + T-, Dementia A-T-, and all MCI groups, and that the MCI A + T+ group had higher values than the MCI A-T- and MCI A + T- groups. In addition, post-hoc analyses for the mesial temporal lobe revealed higher values in the Dementia A + T+ group compared to the Dementia A + T- group, and higher values in the Dementia A-T+ group compared to the CU A-T-, CU A + T-, MCI A-T-, MCI A-T + , and MCI A + T- groups. Figure 6 shows predicted FTP-PET values across groups.

Validation in the ADTC cohort

The third set of validation analyses involved out-of-sample prediction in ADTC patients with available FTP-PET (n = 28). Predicted Schaefer-400 and CenTauR values were extracted using the model trained in the Mayo research cohort and StateViewer-derived eigenvalues in ADTC participants. Mixed-effect models between observed and predicted FTP-PET across the Schaefer-400 atlas achieved R2 = 0.41 [IQR 0.27 to 0.52], MAE = 0.12 [IQR 0.09 to 0.14], and MSE = 0.02 [IQR 0.01 to 0.04] across participants. Individual spatial R2 across the cohort achieved a median value of R2 = 0.80. Figure 7 shows brain renderings of predicted FTP-PET values (panel A), group-wise comparisons (panel B), and observed and predicted FTP-PET values in six region examples (panel C).

The regression model between FDG-PET eigenvalues and CenTauR values yielded R2 = 0.31 [IQR 0.23 to 0.42], MAE = 0.30 [IQR 0.27 to 0.30], and MSE = 0.16 [IQR 0.13 to 0.17] across participants. Comparisons of predicted CenTauR regions across clinical groups regardless of FTP-PET availability showed a significant effect for the mesial temporal area, F(4, 259) = 14.7, P < 0.001, temporoparietal area, F(4, 259) = 23.4, P < 0.001, and the temporal meta-ROI area, F(4, 259) = 20.1, P < 0.001. Post-hoc analyses revealed that, across all regions, the Atypical AD A+ group had higher values than all other groups. In addition, the AD A+ group had higher values than the Non-AD A- group for the mesial temporal lobe and temporal meta-ROI regions. Figure 7 shows predicted FTP-PET values across groups.

Discussion

The trans-modal prediction of AD-type tau across the ADRD spectrum is critical for advancing our understanding of this pathology and its relationship to large-scale functional dynamics and neurodegeneration. In this study, we developed a highly interpretable model of FTP-PET prediction based on FDG-PET imaging and a population-based regression framework, implemented within StateViewer32, a clinical decision support system for dementia syndrome classification based on a computational model of brain function. The model demonstrated high accuracy in the training set in predicting the extent and distribution of FDG-PET as well as tau staging across the ADRD spectrum, including CU individuals with and without positive AD biomarkers, clinical AD phenotypes, and non-AD syndromes spanning the DLB and FTLD spectra. Moderate-to-high performance in longitudinal data and independent datasets from research and clinical practice further highlight its potential for generalization and scalability.

Predictions from the FDG-PET eigenbrain model closely matched observed FTP-PET when applied across and within participants. It accurately captured high FTP-PET burden in atypical AD phenotypes, moderate levels in aMCI and aAD, and lower levels in CU, naMCI, and non-AD groups. The prediction of low FTP-PET in non-AD primary tauopathies (e.g., PSP, CBS) was expected, given the lower binding sensitivity of existing FTP-PET tracers to non-AD tau44,45. This also supports the model’s specificity to AD-type tau in that predictions are driven by specific patterns of whole-brain metabolism rather than simply translating regional FDG-PET signal into spatially equivalent FTP-PET patterns.

This work carries several practical implications. Our FDG-PET–based model can approximate FTP-PET burden across the ADRD spectrum using FDG-PET imaging as a single input, consistent with prior work from our group46. Generalization to longitudinal data and independent ADTC and ADNI cohorts suggests potential scalability to unseen data. Moreover, our model achieved high accuracy (78–95%) in predicting tau staging cross-sectionally and longitudinally both within the training and validation cohorts. This highlights its potential ability to support biological staging under revised diagnostic criteria for AD35, enabling stratification of participants by low versus high tau burden. To this point, clinically relevant case examples in amnestic phenotypes of AD (aMCI/aAD) and syndromes that can be caused either by AD or non-AD etiologies (bvFTD, CBS) demonstrated that the model accurately predicted low tau burden in those with observed low tau (A/B stages). Similar results were found in the ADNI cohort by comparing predicted tau values across CU, MCI, and Dementia patients according to observed biomarker status. This suggests that our model could potentially (1) identify “off-diagonal” symptomatic individuals with possible symptom-driving co-pathologies35,37, and (2) identify rare AD phenotypes that are traditionally caused by non-AD etiologies. FDG-PET could also be used to infer longitudinal FTP-PET burden when the latter is unavailable or onerous. Overall, our model provides a highly scalable surrogate that maximizes the clinical and research yield of FDG-PET, and supports the role of this imaging modality as a close proxy of AD-type tau accumulation47–50, related neurodegeneration, and clinical symptoms25,30,32,

Our findings are also relevant to broader conceptual frameworks of neurodegeneration. The FDG-PET model is grounded in the principle of shared vulnerability across large-scale brain systems18,25. In this view, functional gradients that support specific cognitive domains shape the topography of neurodegeneration across dementia syndromes25,28. Such vulnerability likely arises from commonalities and differences in genetic expression, metabolic demands, and cellular properties within and across gradients18–20,51–54. Importantly, these results do not contradict the prion-like spreading hypothesis in non-prion diseases. Instead, gradient-based and epicenter-based frameworks evaluate similar questions at different scales, and converge on the notion that large-scale functional systems are central to disease biology and may even act not only as mere conduits but also drivers of proteinopathies in later life, an aspect often underappreciated in molecular pathology models25,55–57. From this perspective, tau and other pathogenic drivers may accumulate preferentially along syndrome-specific gradients rather than arising from isolated focal sites, rendering the gradients themselves “disease epicenters.” Our framework helps explain why some large-scale systems are disproportionately vulnerable to AD-type tau while others remain relatively resistant. This is supported by the model’s capacity to predict an absence of FTP-PET signal in dementia syndromes where tau is not the primary driver of neurodegeneration, even in the presence of significant FDG-PET abnormality. Future work linking molecular physiology, genetic expression profiles, and large-scale network organization will be essential to clarify mechanisms of pathologic AD-type tau genesis and propagation1.

It is also noteworthy that while our model explains a high proportion of variance in patterns of FTP-PET, additional factors may influence inter-individual variability. For instance, post-hoc analyses on baseline residuals and rates of change in CDR-SB and temporal FTP-PET support the notion that variance unexplained by our model carries biological meaning and prognostic value. A potential contributing factor is biological sex, as supported by distinct differences in the proportion of variance explained between females and males in our post-hoc stratified analyses. This aligns with extant literature demonstrating sex-specific effects on trajectories of AD-type tau accumulation that disfavor women58. In addition, the model showed a slight degree of proportional bias in high-tau groups (i.e., dAD, PCA, lvPPA). However, this type of bias for extreme values was expected given the linear assumptions of the model. Future work exploring such factors is needed not only to improve the accuracy of data-driven tools, but also to clarify the mechanistic underpinnings of AD-type tau progression.

These findings should be interpreted in light of several limitations. First, the model has not yet been evaluated in non-degenerative conditions that can clinically mimic neurodegenerative syndromes, such as autoimmune disorders, cerebrospinal fluid dynamics, and primary psychiatric disorders. Second, model training was limited to the Mayo research cohort, which is predominantly non-Hispanic white and highly educated. Although out-of-sample performance supports generalizability, further validation in racially, ethnically, and socioeconomically diverse populations is warranted. Third, while we used CenTauR based on prior work59, FTP-PET thresholds for biological stages are not well defined in the literature. In addition, there was evidence of class imbalance in tau stages, with a minimal proportion of cases presenting with intermediate tau stages (B/C). This warrants validation in future cohorts featuring a more evenly distributed representation of tau stages. Fourth, despite acceptable out-of-sample performance in both the ADNI and ADTC cohorts, further cross-center validation is needed. Fifth, the baseline data from the Mayo longitudinal validation cohort overlapped with the training cohort. Moreover, only a small number of individuals transitioned to a higher tau stage. While sensitivity analyses support robust generalization, further replication in completely independent longitudinal datasets is needed. Sixth, our model was only trained on FTP-PET imaging, and it remains to be tested whether it would perform similarly using different ligands with a broader dynamic range, or even recent ligands that may bind to primary tauopathies such as PI262060. Seventh, it is important to note that the ADNI cohort is heavily pre-screened for AD-centric symptoms and biology, and is minimally representative of real-world clinical populations. While the inclusion of the ADTC cohort mitigates this to some degree, replication in larger, clinically heterogeneous cohorts is warranted. Finally, an important future direction of this work is the assessment of integrating FDG-PET markers and plasma biomarkers into a minimal, cost-effective clinical workflow for diagnostic purposes. While our standalone imaging model provides robust utility when fluid biomarkers are missing, combining spatial metabolic predictions with highly accessible plasma assays could further refine differential diagnosis in complex clinical settings61.

In summary, we developed a multimodal model of tau prediction based on a framework using FDG- and FTP-PET imaging and machine learning within a functional gradient paradigm. The model achieved high accuracy across the ADRD spectrum, shows strong potential for scalability and generalization, and may provide a practical tool to improve the efficiency of research and clinical efforts that rely on tau-PET staging.

Methods

Mayo clinic research cohort

Patients and/or their legal representative provided written consent for their data to be used for research purposes. This study met HIPAA guidelines and was approved by the Mayo Clinic Institutional Review Board. Research participants were financially compensated for their participation in the ancillary studies to this research.

The Mayo research cohort included 1088 participants enrolled through research programs at Mayo Clinic Rochester, including the ADRC, the MCSA, and the NRG. The cohort spanned the spectrum of normal aging, AD, DLB, and FTLD. AD phenotypes had to be A+ based on PET imaging and included: aMCI62, aAD63, dAD64, lvPPA65, and PCA66. FTD phenotypes included bvFTD67, agPPA, SD65, CBS39 and PSP68. naMCI participants did not have to be A+ to be included, as they often reflect the prodromal stage of non-AD disorders. CU participants covered the AD biomarker spectrum (A-T-, A + T-, A + T + )35,69. All participants had FDG-PET, amyloid-PET, and FTP-PET imaging acquired within a 6-month window. A subset of participants had at least one pair of follow-up FDG-PET and FTP-PET (n = 215, total of 478 time points) and was used as a validation cohort.

Clinical diagnoses were determined by dementia specialists using standardized protocols established by the National Institute on Aging–funded ADRC program70. This included a structured clinical interview with the participant and an informant, focused on core diagnostic features of the relevant syndromes described above, along with a neurological examination, review of medical records, and ancillary data (e.g., neuroimaging, neuropsychological testing), when appropriate.

ADNI cohort

Participants from ADNI (adni.loni.usc.edu) were used as a second validation cohort (n = 295). This cohort included participants with available amyloid-PET, FTP-PET, and FDG-PET imaging, including participants labeled as CU (n = 13), MCI (n = 211), or dementia (n = 71). Participants were further classified according to biomarker status, with CU A-T- (n = 8), CU A + T- (n = 5), MCI A-T- (n = 94), MCI A + T- (n = 28), MCI A + T+ (n = 70), MCI A-T+ (n = 19), Dementia A-T- (n = 13), Dementia A + T- (n = 8), Dementia A + T+ (n = 46) and Dementia A-T+ (n = 4). Clinical diagnoses were assigned according to standardized protocols described in the ADNI methods33.

ADTC cohort

Clinical patients from the Mayo Clinic ADTC were used as a third validation cohort (n = 264)34. All participants underwent a comprehensive clinical evaluation, including a clinical interview, neurological examination, amyloid-PET, MRI, FDG-PET, neuropsychological testing, and APOE genotyping, as part of the ADTC diagnostic workup. Amyloid status was based on visual reads by nuclear medicine radiologists with expertise in dementia imaging. Only patients with available pre-treatment FDG-PET were included, and none had received amyloid-reducing therapies at the time of data collection. Clinical diagnoses were established via consensus among neurologists, neuropsychologists, and radiologists with expertise in aging and dementia.

The cohort was clinically heterogeneous and included patients with early symptomatic AD (aMCI or mild AD-type dementia; “AD A + ,” n = 138), early symptomatic AD with a high suspicion of co-pathology (e.g., α-synucleinopathy or TDP-43; “AD A+ co-path”, n = 23), atypical AD variants (dAD, lvPPA, PCA, motor AD, or behavioral AD; “Atypical AD A + ”, n = 64), and non-AD syndromes (e.g., normal pressure hydrocephalus, Parkinson’s disease dementia, vascular dementia, frontotemporal dementia; “Non-AD”). The Non-AD group was further stratified by amyloid status into “Non-AD A + ” (n = 11) and “Non-AD A–” (n = 28), where Non-AD A+ patients were considered to have a non-AD etiology as the primary driver of symptoms despite having biological AD46. A subset of participants (n = 28) also had research FTP-PET acquired within six months of FDG-PET as part of concurrent research protocols.

The study was approved by the Mayo Clinic Institutional Review Board and complied with HIPAA privacy regulations. All participants, or their legally authorized representatives, provided written informed consent for the use of their data in research.

MRI acquisition and processing

MRI scans from the Mayo research cohort and a subset of the ADTC cohort with research MRI (n = 28) were acquired on a 3 T GE scanner using a magnetization-prepared rapid gradient echo (MPRAGE) sequence. Acquisition parameters included: repetition time (TR) = 2300 ms, echo time (TE) = 3 ms, flip angle = 8°, field of view (FOV) = 26 cm, reconstructed in a 256 × 256 mm in-plane matrix with a phase FOV of 0.94, and slice thickness = 1.2 mm. MRI scans from the ADNI cohort were acquired on 3 T scanners at individual ADNI sites using standardized MPRAGE protocols (see: https://adni.loni.usc.edu/data-samples/adni-data/neuroimaging/mri/). Parameters included: 1 mm isotropic resolution, TR = 2300 ms, minimum full TE, inversion time = 900 ms, and FOV reconstructed in a 208 × 240 × 256 mm matrix. All MRI images were segmented using the Unified Segmentation algorithm in SPM12, incorporating Mayo Clinic Adult Lifespan Template (MCALT) population-optimized priors and settings. MCALT atlases were subsequently normalized to each individual scan using ANTs.

Amyloid- and FTP-PET imaging acquisition and processing

Amyloid-PET imaging was performed using [¹¹C]PiB for the Mayo research cohort and [¹⁸F]AV45 (florbetapir) or [18F]FBB (florbetaben) for the ADNI cohort. Clinical amyloid-PET imaging in the ADTC cohort was performed using [¹⁸F]AV45 (florbetapir). FTP-PET imaging across all cohorts employed the [¹⁸F]AV1451 (FTP) tracer. Detailed acquisition and processing protocols have been previously described71,72 and are available online for the ADNI cohort (https://adni.loni.usc.edu/data-samples/adni-data/neuroimaging/pet/). Research PET images were co-registered to each participant’s corresponding MRI and spatially normalized to the Mayo Clinic Adult Lifespan Template (MCALT) using SPM12. SUVR images were generated by normalizing tracer uptake to the cerebellar crus, and mean SUVR values were extracted from validated meta-regions of interest (meta-ROIs) based on the MCSA202 atlas72. Amyloid positivity thresholds were set at >25 centiloid across amyloid ligands73–76, except for the ADTC clinical amyloid-PET scans, for which positivity was determined by expert visual reads as described above (and therefore assumed to be >25 centiloid). FTP-PET positivity was defined as SUVR > 1.2977

Analyses involving FTP-PET data utilized two parcellations. The Schaefer-400 atlas was applied to train the FDG-PET-based FTP-PET prediction model at the whole-brain level72. The CenTauR atlas was used to (1) determine tau stage based on mesial temporal and temporoparietal areas78,79 and (2) gather a proxy of tau burden based on the temporal meta-ROI. SUVR values in the CenTauR regions were extracted using the cerebellar reference region and converted to CenTauR values59. Tau stages were defined according to the revised criteria for AD diagnosis and staging35, as follows: A- (amyloid centiloid ≤25), A (amyloid centiloid ≥25; mesial temporal lobe CenTauR ≤ 39), B (amyloid centiloid ≥25; mesial temporal lobe CenTauR ≥ 39; temporoparietal CenTauR ≤25), C (amyloid centiloid ≥25; temporoparietal CenTauR ≥25 and ≤63), and D (amyloid centiloid ≥25; temporoparietal CenTauR ≥63).

FDG-PET imaging and processing

FDG-PET imaging in the Mayo research and ADTC cohort was performed using a PET/CT scanner (GE Healthcare) following a 30-minute uptake period during which participants rested in a dimly lit room. Image acquisition lasted 8 min and consisted of four dynamic frames (2 min each), preceded by a low-dose CT transmission scan. In the ADNI cohort, FDG-PET images were acquired as a 30-minute dynamic emission scan composed of six 5-minute frames, following a 30- to 60-minute uptake period. Detailed acquisition protocols for the ADNI cohort are available online (https://adni.loni.usc.edu/data-samples/adni-data/neuroimaging/pet/). FDG-PET images from all cohorts were processed using an MRI-independent pipeline32. This included non-linear symmetric diffeomorphic registration to the MCALT, normalization of uptake values to the pons to generate SUVR images, and spatial smoothing with a 6 mm full width at half maximum (FWHM) Gaussian kernel. Regional SUVR values were additionally extracted across the Schaefer-400 atlas for visualization purposes only.

FDG-PET eigenbrain model

FDG-PET images were analyzed using StateViewer, a machine learning–based latent space framework previously developed at Mayo Clinic32. This approach applies principal component decomposition to 1000 randomly sampled FDG-PET scans from a large database of clinical and research individuals at Mayo Clinic, and that is non-overlapping with the cohorts of interest in this study. Cohorts of interest (Mayo, ADNI, ADTC) were projected into this latent space, defined by the first 150 components, referred to as “eigenbrains,” to extract participant-level eigenvalues. Each eigenbrain represents an orthogonal axis of spatial metabolic variation across the cortex, ordered by the proportion of variance explained, and each eigenvalue quantifies the degree to which an individual’s metabolic pattern aligns with the spatial topology of that eigenbrain. Each eigenbrain represents an orthogonal axis of spatial metabolic variation across the cortex, ordered by the proportion of variance explained, and each eigenvalue quantifies the degree to which an individual’s metabolic pattern aligns with the spatial topology of that eigenbrain. As such, each participant-level eigenvalue for each eigenbrain represents a mapping onto a specific FDG-PET gradient, resulting in a 1088 ×150 (participant X eigenvalue) table as input.

Model training in the Mayo research cohort

The FDG-PET-based model of tau prediction is configured as a population-level regression framework. Population-derived regression coefficients are applied to individual participant FDG-PET features to estimate their corresponding regional FTP-PET burden across a standardized brain atlas. We first trained a mixed-effects linear regression framework in the Mayo research cohort to predict regional FTP-PET SUVR across the Schaefer-400 atlas using participant-level FDG-PET eigenvalues. All variables were log-transformed. Model performance was quantified using the coefficient of determination (R2), adjusted R2 (R2adj), predicted R2 (R2pred) obtained via leave-one-out cross-validation across participants, MAE, and MSE. Regressions were iteratively performed with increasing numbers of eigenbrains, and the optimal model was defined as that achieving the highest atlas-wide R2pred. Visual inspection of model performance included brain renderings of R2 values across participants in each group, and Bland-Altman plots across the Schaefer-400 atlas to assess systematic over- and under-estimation of FTP-PET values at the group- and region-level. In addition, we computed spatial R2 values between observed and predicted FTP-PET across the Schaefer-400 atlas within each participant, and report the median value aggregated across the cohort. Models were re-run including age, sex, and APOE status as covariates to assess potential improvement in prediction, and atlas-wide R2 values were compared across male and female participants.

To evaluate whether non-linear or regularized machine learning architectures could optimize regional FTP-PET prediction relative to our primary linear model, we benchmarked performance against five alternative regression frameworks: Ridge regression (α = 0), Lasso regression (α = 1), Elastic Net regression (tuning both α and λ) Random Forest (implemented via the fast conditional inference forest engine ranger), and Support Vector Machines (SVM) utilizing a Radial Basis Function (RBF) kernel. Each algorithm was trained to predict regional log-transformed FTP-PET SUVR values across participants using the optimal number of FDG-PET eigenbrain eigenvalues based on the linear model. Model optimization and hyperparameter tuning were executed via a 5-fold cross-validation. The optimal model was additionally benchmarked against a base model using age, sex, APOE4 carriership, clinical diagnosis, and CDR-SB as predictors of FTP-PET across participants.

Additional regression models were trained to predict CenTauR mesial temporal, temporoparietal, and temporal meta-ROI values, with the number of eigenbrains determined based on the optimal Schaefer-400 model. CenTauR values were then converted back from a log distribution into their original scale using a Duan’s smearing estimator. They were then submitted to a confusion matrix to derive balanced accuracy, sensitivity, specificity, positive predictive value, and negative predictive value between observed and predicted tau stages. Due to class imbalance, we combined stages A/B (absence of significant tau/mesial temporal tau) and C/B (moderate-to-severe temporoparietal tau). These analyses were performed with and without A- individuals, since this stage is influenced by amyloid status only and not tau.

Additional analyses were performed for clinically relevant case examples, where observed versus predicted tau stages were compared in aAD, aMCI, bvFTD, and CBS, since these syndromes can be either influenced by co-pathologies (e.g., TDP-43 in AD)36–38 or result from either AD or other primary proteinopathies (e.g., TDP-43 or tauopathies)39–42. Predicted CenTauR values (mesial temporal, temporoparietal, temporal) were also compared between observed stages (A-, A/B, C/D).

Model validation in the longitudinal Mayo research cohort

Baseline-trained models were applied across participants from the Mayo research cohort with longitudinal FDG-PET and FTP-PET data to compare observed and predicted follow-up Schaefer-400 and CenTauR values. Linear mixed-effects models accounted for within-participant variability and time from baseline. In addition, we computed spatial R2 values between observed and predicted FTP-PET across the Schaefer-400 atlas for each participant and report the cohort-level median value aggregated across visits. Model performance was summarized using marginal R2, reflecting variance explained by the fixed effects (predicted tau, time from baseline, and their interaction), marginal MSE, and marginal MAE.

Baseline-trained models on CenTauR values (mesial temporal, temporoparietal, temporal meta-ROI) were applied to extract performance metrics. CenTauR cut-offs were used to determine observed and predicted tau stages for each participant at each time point. First, a confusion matrix was computed for baseline data. Then, another confusion matrix was computed for tau stage transitions (e.g., A/B to C/D) between the first and last visits for each participant. Performance was assessed using balanced accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. Here again, analyses were performed with and without A- individuals.

Sensitivity analyses were performed where baseline-trained models of FTP-PET prediction were re-run after excluding participants who had longitudinal data available. These models were subsequently used to replicate the longitudinal validation analyses listed above. This was done to assess whether any overlap between the training and validation cohorts influenced predictive performance in the longitudinal analyses.

Finally, we conducted a longitudinal validation analysis to determine whether baseline model prediction errors carry prognostic value for future clinical deterioration and disease progression. For each participant, prediction residuals (observed FTP-PET minus predicted FTP-PET) were calculated for the temporal meta-ROI at the baseline visit. Longitudinal clinical progression and downstream tau accumulation were quantified by calculating the change (Δ) in CDR-SB and the change (Δ) in temporal meta-ROI FTP-PET SUVR between the baseline visit and the final follow-up visit. Because clinical scales and residuals often yield non-normal distributions and tied ranks, the association between baseline prediction residuals and subsequent longitudinal changes (Δ CDR-SB and Δ FTP-PET) was evaluated using partial Spearman rank correlation coefficients.

Model validation in the ADNI cohort

Mayo-derived models were applied to ADNI participants using their StateViewer eigenvalues. Linear regression assessed correspondence across the Schaefer-400 and CenTauR values, summarized using median R2, MAE, and MSE. We computed spatial R2 values between observed and predicted FTP-PET across the Schaefer-400 atlas for each participant and report the cohort-level median value. A confusion matrix was used to assess agreement between observed and predicted tau stages as described above. Predicted CenTauR temporal meta-ROI values were compared across clinically- and biomarker-defined groups (CU, MCI, Dementia; A + /A-, T + /T-) using ANOVA with Tukey post hoc analyses.

Model validation in the ADTC cohort

Mayo-derived models were applied to ADTC participants using their StateViewer eigenvalues. Linear regression assessed correspondence across the Schaefer-400 and CenTauR values in the subset with available PET-PET, summarized using median R2, MAE, and MSE. We computed spatial R2 values between observed and predicted FTP-PET across the Schaefer-400 atlas for each participant and report the cohort-level median value. Agreement between observed and predicted tau stage was not assessed in the ADTC cohort due to the small number of participants with available FTP-PET and significant class imbalance.

Predicted CenTauR values (mesial temporal, temporoparietal, temporal meta-ROI) were compared across clinical groups (AD A + , AD A+ co-path, Atypical AD A + , Non-AD A–, Non-AD A + ) using ANOVA with Tukey post hoc analyses, regardless of FTP-PET availability.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Supplementary information

Source data

References

1.Vogel JW Connectome-based modelling of neurodegenerative diseases: towards precision medicine and mechanistic insight Nat. Rev. Neurosci. 2023 24 620 639 10.1038/s41583-023-00731-8 37620599 Vogel, J. W. et al. Connectome-based modelling of neurodegenerative diseases: towards precision medicine and mechanistic insight. Nat. Rev. Neurosci.24, 620–639 (2023).10.1038/s41583-023-00731-8 37620599

2.Walker LC Jucker M The prion principle and Alzheimer’s disease Science 2024 385 1278 1279 39298592 10.1126/science.adq5252 Walker, L. C. & Jucker, M. The prion principle and Alzheimer’s disease. Science385, 1278–1279 (2024).10.1126/science.adq5252 PMC11492928 39298592

3.Frontzkowski L Earlier Alzheimer’s disease onset is associated with tau pathology in brain hub regions and facilitated tau spreading Nat. Commun. 2022 13 4899 10.1038/s41467-022-32592-7 35987901 Frontzkowski, L. et al. Earlier Alzheimer’s disease onset is associated with tau pathology in brain hub regions and facilitated tau spreading. Nat. Commun.13, 4899 (2022).10.1038/s41467-022-32592-7 PMC9392750 35987901

4.Franzmeier N Functional connectivity associated with tau levels in ageing, Alzheimer’s, and small vessel disease Brain 2019 142 1093 1107 30770704 10.1093/brain/awz026 Franzmeier, N. et al. Functional connectivity associated with tau levels in ageing, Alzheimer’s, and small vessel disease. Brain142, 1093–1107 (2019).10.1093/brain/awz026 PMC6439332 30770704

5.Franzmeier N Patient-centered connectivity-based prediction of tau pathology spread in Alzheimer’s disease Sci. Adv. 2020 6 eabd1327 33246962 10.1126/sciadv.abd1327 Franzmeier, N. et al. Patient-centered connectivity-based prediction of tau pathology spread in Alzheimer’s disease. Sci. Adv.6, eabd1327 (2020).10.1126/sciadv.abd1327 PMC7695466 33246962

6.Franzmeier N Tau deposition patterns are associated with functional connectivity in primary tauopathies Nat. Commun. 2022 13 1362 10.1038/s41467-022-28896-3 35292638 Franzmeier, N. et al. Tau deposition patterns are associated with functional connectivity in primary tauopathies. Nat. Commun.13, 1362 (2022).10.1038/s41467-022-28896-3 PMC8924216 35292638

7.Vogel JW Spread of pathological tau proteins through communicating neurons in human Alzheimer’s disease Nat. Commun. 2020 11 2612 10.1038/s41467-020-15701-2 32457389 Vogel, J. W. et al. Spread of pathological tau proteins through communicating neurons in human Alzheimer’s disease. Nat. Commun.11, 2612 (2020).10.1038/s41467-020-15701-2 PMC7251068 32457389

8.Sintini I Tau and Amyloid Relationships with Resting-state Functional Connectivity in Atypical Alzheimer’s Disease Cereb. Cortex 2021 31 1693 1706 33152765 10.1093/cercor/bhaa319 Sintini, I. et al. Tau and Amyloid Relationships with Resting-state Functional Connectivity in Atypical Alzheimer’s Disease. Cereb. Cortex31, 1693–1706 (2021).10.1093/cercor/bhaa319 PMC7869088 33152765

9.Therriault J Intrinsic connectivity of the human brain provides scaffold for tau aggregation in clinical variants of Alzheimer’s disease Sci. Transl. Med. 2022 14 eabc8693 36001678 10.1126/scitranslmed.abc8693 Therriault, J. et al. Intrinsic connectivity of the human brain provides scaffold for tau aggregation in clinical variants of Alzheimer’s disease. Sci. Transl. Med.14, eabc8693 (2022).10.1126/scitranslmed.abc8693 36001678

10.de Bruin H Connectivity as a universal predictor of tau progression in atypical Alzheimer’s disease Brain 2025 148 3893 3912 40810361 10.1093/brain/awaf279 de Bruin, H. et al. Connectivity as a universal predictor of tau progression in atypical Alzheimer’s disease. Brain148, 3893–3912 (2025).10.1093/brain/awaf279 PMC12588720 40810361

11.Cope TE Tau burden and the functional connectome in Alzheimer’s disease and progressive supranuclear palsy Brain 2018 141 550 567 29293892 10.1093/brain/awx347 Cope, T. E. et al. Tau burden and the functional connectome in Alzheimer’s disease and progressive supranuclear palsy. Brain141, 550–567 (2018).10.1093/brain/awx347 PMC5837359 29293892

12.Brown JA Patient-tailored, connectivity-based forecasts of spreading brain atrophy Neuron 2019 104 856 868.e5 31623919 10.1016/j.neuron.2019.08.037 Brown, J. A. et al. Patient-tailored, connectivity-based forecasts of spreading brain atrophy. Neuron104, 856–868.e5 (2019).10.1016/j.neuron.2019.08.037 PMC7012373 31623919

13.Seeley WW Crawford RK Zhou J Miller BL Greicius MD Neurodegenerative diseases target large-scale human brain networks Neuron 2009 62 42 52 19376066 10.1016/j.neuron.2009.03.024 Seeley, W. W., Crawford, R. K., Zhou, J., Miller, B. L. & Greicius, M. D. Neurodegenerative diseases target large-scale human brain networks. Neuron62, 42–52 (2009).10.1016/j.neuron.2009.03.024 PMC2691647 19376066

14.Raj A Kuceyeski A Weiner M A network diffusion model of disease progression in dementia Neuron 2012 73 1204 1215 22445347 10.1016/j.neuron.2011.12.040 Raj, A., Kuceyeski, A. & Weiner, M. A network diffusion model of disease progression in dementia. Neuron73, 1204–1215 (2012).10.1016/j.neuron.2011.12.040 PMC3623298 22445347

15.Wu C Clinical applications of magnetic resonance imaging based functional and structural connectivity NeuroImage 2021 244 118649 34648960 10.1016/j.neuroimage.2021.118649 Wu, C. et al. Clinical applications of magnetic resonance imaging based functional and structural connectivity. NeuroImage244, 118649 (2021).10.1016/j.neuroimage.2021.118649 34648960

16.Walsh DM Selkoe DJ A critical appraisal of the pathogenic protein spread hypothesis of neurodegeneration Nat. Rev. Neurosci. 2016 17 251 260 10.1038/nrn.2016.13 26988744 Walsh, D. M. & Selkoe, D. J. A critical appraisal of the pathogenic protein spread hypothesis of neurodegeneration. Nat. Rev. Neurosci.17, 251–260 (2016).10.1038/nrn.2016.13 PMC6701169 26988744

17.Jones, D. T. & Breakspear, M. The Hippocampal latent diffusion engine: a computational framework for memory, perception, and cognitive dysfunction. Preprint at https://osf.io/preprints/psyarxiv/tnygx_v1/ (2026).

18.Margulies DS Situating the default-mode network along a principal gradient of macroscale cortical organization Proc. Natl. Acad. Sci. USA 2016 113 12574 12579 27791099 10.1073/pnas.1608282113 Margulies, D. S. et al. Situating the default-mode network along a principal gradient of macroscale cortical organization. Proc. Natl. Acad. Sci. USA113, 12574–12579 (2016).10.1073/pnas.1608282113 PMC5098630 27791099

19.Goulas A The natural axis of transmitter receptor distribution in the human cerebral cortex Proc. Natl. Acad. Sci. USA 2021 118 e2020574118 33452137 10.1073/pnas.2020574118 Goulas, A. et al. The natural axis of transmitter receptor distribution in the human cerebral cortex. Proc. Natl. Acad. Sci. USA118, e2020574118 (2021).10.1073/pnas.2020574118 PMC7826352 33452137

20.Vogel JW Deciphering the functional specialization of whole-brain spatiomolecular gradients in the adult brain Proc. Natl. Acad. Sci. USA 2024 121 e2219137121 38861593 10.1073/pnas.2219137121 Vogel, J. W. et al. Deciphering the functional specialization of whole-brain spatiomolecular gradients in the adult brain. Proc. Natl. Acad. Sci. USA121, e2219137121 (2024).10.1073/pnas.2219137121 PMC11194492 38861593

21.Hansen JY Mapping neurotransmitter systems to the structural and functional organization of the human neocortex Nat. Neurosci. 2022 25 1569 1581 10.1038/s41593-022-01186-3 36303070 Hansen, J. Y. et al. Mapping neurotransmitter systems to the structural and functional organization of the human neocortex. Nat. Neurosci.25, 1569–1581 (2022).10.1038/s41593-022-01186-3 PMC9630096 36303070

22.Castrillon G An energy costly architecture of neuromodulators for human brain evolution and cognition Sci. Adv. 2023 9 eadi7632 38091393 10.1126/sciadv.adi7632 Castrillon, G. et al. An energy costly architecture of neuromodulators for human brain evolution and cognition. Sci. Adv.9, eadi7632 (2023).10.1126/sciadv.adi7632 PMC10848727 38091393

23.Brown JA Lee AJ Pasquini L Seeley WW A dynamic gradient architecture generates brain activity states NeuroImage 2022 261 119526 35914669 10.1016/j.neuroimage.2022.119526 Brown, J. A., Lee, A. J., Pasquini, L. & Seeley, W. W. A dynamic gradient architecture generates brain activity states. NeuroImage261, 119526 (2022).10.1016/j.neuroimage.2022.119526 PMC9585924 35914669

24.Shine JM Human cognition involves the dynamic integration of neural activity and neuromodulatory systems Nat. Neurosci. 2019 22 289 296 10.1038/s41593-018-0312-0 30664771 Shine, J. M. et al. Human cognition involves the dynamic integration of neural activity and neuromodulatory systems. Nat. Neurosci.22, 289–296 (2019).10.1038/s41593-018-0312-0 30664771

25.Jones D A computational model of neurodegeneration in Alzheimer’s disease Nat. Commun. 2022 13 1643 10.1038/s41467-022-29047-4 35347127 Jones, D. et al. A computational model of neurodegeneration in Alzheimer’s disease. Nat. Commun.13, 1643 (2022).10.1038/s41467-022-29047-4 PMC8960876 35347127

26.Brown JA Functional network collapse in neurodegenerative disease Nat. Commun. 2025 16 10273 10.1038/s41467-025-65156-6 41271684 Brown, J. A. et al. Functional network collapse in neurodegenerative disease. Nat. Commun.16, 10273 (2025).10.1038/s41467-025-65156-6 PMC12639071 41271684

27.Ottoy J Tau follows principal axes of functional and structural brain organization in Alzheimer’s disease Nat. Commun. 2024 15 5031 10.1038/s41467-024-49300-2 38866759 Ottoy, J. et al. Tau follows principal axes of functional and structural brain organization in Alzheimer’s disease. Nat. Commun.15, 5031 (2024).10.1038/s41467-024-49300-2 PMC11169286 38866759

28.Brown, J. A. et al. Functional network collapse in neurodegenerative disease. bioRxiv 2023.12.01.569654 10.1101/2023.12.01.569654 .(2023).PMC12639071 41271684

29.Rittmo, J. et al. Age and Alzheimer’s disease affect functional connectivity along separate axes of functional brain organization. 2025.05.22.655469 Preprint at 10.1101/2025.05.22.655469 (2025).

30.Minoshima, S., Thientunyakit, T., Cross, D. J. & Mosci, K. FDG PET imaging evaluation of neurodegenerative dementias. In Molecular Imaging of Neurodegenerative Disorders (eds Cross, D. J., Mosci, K. & Minoshima, S.) 57–71 10.1007/978-3-031-35098-6_4 .(Springer, 2023).

31.Herholz K FDG PET and differential diagnosis of dementia Alzheimer Dis. Associated Disord. 1995 9 6 16 10.1097/00002093-199505000-00004 7605624 Herholz, K. FDG PET and differential diagnosis of dementia. Alzheimer Dis. Associated Disord.9, 6–16 (1995).10.1097/00002093-199505000-00004 7605624

32.Barnard L An FDG-PET–based machine learning framework to support neurologic decision-making in alzheimer disease and related disorders Neurology 2025 105 e213831 40577677 10.1212/WNL.0000000000213831 Barnard, L. et al. An FDG-PET–based machine learning framework to support neurologic decision-making in alzheimer disease and related disorders. Neurology105, e213831 (2025).10.1212/WNL.0000000000213831 PMC12207676 40577677

33.Petersen RC Alzheimer’s disease neuroimaging initiative (ADNI): clinical characterization Neurology 2010 74 201 209 20042704 10.1212/WNL.0b013e3181cb3e25 PMC2809036 Petersen, R. C. et al. Alzheimer’s disease neuroimaging initiative (ADNI): clinical characterization. Neurology74, 201–209 (2010).10.1212/WNL.0b013e3181cb3e25 PMC2809036 20042704

34.Neth, B. J. et al. Establishing an Alzheimer disease therapeutics clinic: experience from a year of evaluations. Mayo Clin. Proc.101, 1497–1511 (2026).10.1016/j.mayocp.2025.09.019 41575410

35.Jack CR Revised criteria for diagnosis and staging of Alzheimer’s disease: Alzheimer’s Association workgroup Alzheimer’s. Dement. 2024 20 5143 5169 10.1002/alz.13859 PMC11350039 38934362 Jack, C. R. et al. Revised criteria for diagnosis and staging of Alzheimer’s disease: Alzheimer’s Association workgroup. Alzheimer’s. Dement.20, 5143–5169 (2024).10.1002/alz.13859 PMC11350039 38934362

36.Wolk DA Clinical criteria for limbic-predominant age-related TDP-43 encephalopathy Alzheimer’s. Dement. 2025 21 e14202 10.1002/alz.14202 PMC11772733 39807681 Wolk, D. A. et al. Clinical criteria for limbic-predominant age-related TDP-43 encephalopathy. Alzheimer’s. Dement.21, e14202 (2025).10.1002/alz.14202 PMC11772733 39807681

37.Corriveau-Lecavalier N Clinical criteria for a limbic-predominant amnestic neurodegenerative syndrome Brain Commun. 2024 6 fcae183 39021510 10.1093/braincomms/fcae183 Corriveau-Lecavalier, N. et al. Clinical criteria for a limbic-predominant amnestic neurodegenerative syndrome. Brain Commun.6, fcae183 (2024).10.1093/braincomms/fcae183 PMC11251771 39021510

38.Nelson PT Limbic-predominant age-related TDP-43 encephalopathy (LATE): consensus working group report Brain 2019 142 1503 1527 31039256 10.1093/brain/awz099 Nelson, P. T. et al. Limbic-predominant age-related TDP-43 encephalopathy (LATE): consensus working group report. Brain142, 1503–1527 (2019).10.1093/brain/awz099 PMC6536849 31039256

39.Armstrong MJ Criteria for the diagnosis of corticobasal degeneration Neurology 2013 80 496 503 23359374 10.1212/WNL.0b013e31827f0fd1 Armstrong, M. J. et al. Criteria for the diagnosis of corticobasal degeneration. Neurology80, 496–503 (2013).10.1212/WNL.0b013e31827f0fd1 PMC3590050 23359374

40.Ghirelli A Relationships between hypometabolism and both β-amyloid and tau PET in corticobasal syndrome Alzheimer’s Dement. 2025 21 e70018 10.1002/alz.70018 PMC11923383 40108988 Ghirelli, A. et al. Relationships between hypometabolism and both β-amyloid and tau PET in corticobasal syndrome. Alzheimer’s Dement.21, e70018 (2025).10.1002/alz.70018 PMC11923383 40108988

41.Corriveau-Lecavalier N Uncovering the distinct macro-scale anatomy of dysexecutive and behavioural degenerative diseases Brain 2024 147 1483 1496 37831661 10.1093/brain/awad356 Corriveau-Lecavalier, N. et al. Uncovering the distinct macro-scale anatomy of dysexecutive and behavioural degenerative diseases. Brain147, 1483–1496 (2024).10.1093/brain/awad356 PMC10994526 37831661

42.Ossenkoppele R Research criteria for the behavioral variant of Alzheimer disease: a systematic review and meta-analysis JAMA Neurol. 2022 79 48 60 34870696 10.1001/jamaneurol.2021.4417 Ossenkoppele, R. et al. Research criteria for the behavioral variant of Alzheimer disease: a systematic review and meta-analysis. JAMA Neurol.79, 48–60 (2022).10.1001/jamaneurol.2021.4417 PMC8649917 34870696

43.Thomas Yeo BT The organization of the human cerebral cortex estimated by intrinsic functional connectivity J. Neurophysiol. 2011 106 1125 1165 10.1152/jn.00338.2011 PMC3174820 21653723 Thomas Yeo, B. T. et al. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J. Neurophysiol.106, 1125–1165 (2011).10.1152/jn.00338.2011 PMC3174820 21653723

44.Leuzy A Tau PET imaging in neurodegenerative tauopathies—still a challenge Mol. Psychiatry 2019 24 1112 1134 10.1038/s41380-018-0342-8 30635637 Leuzy, A. et al. Tau PET imaging in neurodegenerative tauopathies—still a challenge. Mol. Psychiatry24, 1112–1134 (2019).10.1038/s41380-018-0342-8 PMC6756230 30635637

45.Groot C Villeneuve S Smith R Hansson O Ossenkoppele R Tau PET imaging in neurodegenerative disorders J. Nucl. Med. 2022 63 20S 26S 35649647 10.2967/jnumed.121.263196 Groot, C., Villeneuve, S., Smith, R., Hansson, O. & Ossenkoppele, R. Tau PET imaging in neurodegenerative disorders. J. Nucl. Med.63, 20S–26S (2022).10.2967/jnumed.121.263196 35649647

46.Lee J Synthesizing images of tau pathology from cross-modal neuroimaging using deep learning Brain 2024 147 980 995 37804318 10.1093/brain/awad346 Lee, J. et al. Synthesizing images of tau pathology from cross-modal neuroimaging using deep learning. Brain147, 980–995 (2024).10.1093/brain/awad346 PMC10907092 37804318

47.Bischof GN Impact of tau and amyloid burden on glucose metabolism in Alzheimer’s disease Ann. Clin. Transl. Neurol. 2016 3 934 939 28097205 10.1002/acn3.339 Bischof, G. N. et al. Impact of tau and amyloid burden on glucose metabolism in Alzheimer’s disease. Ann. Clin. Transl. Neurol.3, 934–939 (2016).10.1002/acn3.339 PMC5224823 28097205

48.Lowe VJ Tau-positron emission tomography correlates with neuropathology findings Alzheimer’s. Dement. 2020 16 561 571 10.1016/j.jalz.2019.09.079 PMC7067654 31784374 Lowe, V. J. et al. Tau-positron emission tomography correlates with neuropathology findings. Alzheimer’s. Dement.16, 561–571 (2020).10.1016/j.jalz.2019.09.079 PMC7067654 31784374

49.Ossenkoppele R Tau PET patterns mirror clinical and neuroanatomical variability in Alzheimer’s disease Brain 2016 139 1551 1567 26962052 10.1093/brain/aww027 Ossenkoppele, R. et al. Tau PET patterns mirror clinical and neuroanatomical variability in Alzheimer’s disease. Brain139, 1551–1567 (2016).10.1093/brain/aww027 PMC5006248 26962052

50.Sintini I Longitudinal rates of atrophy and tau accumulation differ between the visual and language variants of atypical Alzheimer’s disease Alzheimer’s. Dement. 2023 19 4396 4406 10.1002/alz.13396 PMC10592409 37485642 Sintini, I. et al. Longitudinal rates of atrophy and tau accumulation differ between the visual and language variants of atypical Alzheimer’s disease. Alzheimer’s. Dement.19, 4396–4406 (2023).10.1002/alz.13396 PMC10592409 37485642

51.Fu H Hardy J Duff KE Selective vulnerability in neurodegenerative diseases Nat. Neurosci. 2018 21 1350 1358 10.1038/s41593-018-0221-2 30250262 Fu, H., Hardy, J. & Duff, K. E. Selective vulnerability in neurodegenerative diseases. Nat. Neurosci.21, 1350–1358 (2018).10.1038/s41593-018-0221-2 PMC6360529 30250262

52.Arnatkeviciute A Genetic influences on hub connectivity of the human connectome Nat. Commun. 2021 12 4237 10.1038/s41467-021-24306-2 34244483 Arnatkeviciute, A. et al. Genetic influences on hub connectivity of the human connectome. Nat. Commun.12, 4237 (2021).10.1038/s41467-021-24306-2 PMC8271018 34244483

53.Arnatkeviciute A Fulcher BD Bellgrove MA Fornito A Imaging transcriptomics of brain disorders Biol. Psychiatry Glob. Open Sci. 2022 2 319 331 36324650 10.1016/j.bpsgos.2021.10.002 PMC9616271 Arnatkeviciute, A., Fulcher, B. D., Bellgrove, M. A. & Fornito, A. Imaging transcriptomics of brain disorders. Biol. Psychiatry Glob. Open Sci.2, 319–331 (2022).10.1016/j.bpsgos.2021.10.002 PMC9616271 36324650

54.Cirrito JR Synaptic activity regulates interstitial fluid amyloid-β levels in vivo Neuron 2005 48 913 922 16364896 10.1016/j.neuron.2005.10.028 Cirrito, J. R. et al. Synaptic activity regulates interstitial fluid amyloid-β levels in vivo. Neuron48, 913–922 (2005).10.1016/j.neuron.2005.10.028 16364896

55.Warren JD Molecular nexopathies: a new paradigm of neurodegenerative disease Trends Neurosci. 2013 36 561 569 23876425 10.1016/j.tins.2013.06.007 Warren, J. D. et al. Molecular nexopathies: a new paradigm of neurodegenerative disease. Trends Neurosci.36, 561–569 (2013).10.1016/j.tins.2013.06.007 PMC3794159 23876425

56.Jones DT Cascading network failure across the Alzheimer’s disease spectrum Brain 2016 139 547 562 26586695 10.1093/brain/awv338 PMC4805086 Jones, D. T. et al. Cascading network failure across the Alzheimer’s disease spectrum. Brain139, 547–562 (2016).10.1093/brain/awv338 PMC4805086 26586695

57.Jones DT Tau, amyloid, and cascading network failure across the Alzheimer’s disease spectrum Cortex 2017 97 143 159 29102243 10.1016/j.cortex.2017.09.018 Jones, D. T. et al. Tau, amyloid, and cascading network failure across the Alzheimer’s disease spectrum. Cortex97, 143–159 (2017).10.1016/j.cortex.2017.09.018 PMC5773067 29102243

58.Coughlan GT Sex differences in longitudinal Tau-PET in preclinical Alzheimer disease: a meta-analysis JAMA Neurol. 2025 82 364 375 40029638 10.1001/jamaneurol.2025.0013 Coughlan, G. T. et al. Sex differences in longitudinal Tau-PET in preclinical Alzheimer disease: a meta-analysis. JAMA Neurol.82, 364–375 (2025).10.1001/jamaneurol.2025.0013 PMC11877413 40029638

59.Johnson DR Staging Alzheimer’s disease through amyloid and tau PET Brain 2025 149 awaf346 10.1093/brain/awaf346 PMC13016809 40974012 Johnson, D. R. et al. Staging Alzheimer’s disease through amyloid and tau PET. Brain149, awaf346 (2025).10.1093/brain/awaf346 PMC13016809 40974012

60.Brendel M Assessment of 18F-PI-2620 as a biomarker in progressive supranuclear palsy JAMA Neurol. 2020 77 1408 1419 33165511 10.1001/jamaneurol.2020.2526 Brendel, M. et al. Assessment of 18F-PI-2620 as a biomarker in progressive supranuclear palsy. JAMA Neurol.77, 1408–1419 (2020).10.1001/jamaneurol.2020.2526 PMC7341407 33165511

61.Brum WS A blood-based biomarker workflow for optimal tau-PET referral in memory clinic settings Nat. Commun. 2024 15 2311 10.1038/s41467-024-46603-2 38486040 Brum, W. S. et al. A blood-based biomarker workflow for optimal tau-PET referral in memory clinic settings. Nat. Commun.15, 2311 (2024).10.1038/s41467-024-46603-2 PMC10940585 38486040

62.Petersen RC Current concepts in mild cognitive impairment Arch. Neurol. 2001 58 1985 11735772 10.1001/archneur.58.12.1985 Petersen, R. C. et al. Current concepts in mild cognitive impairment. Arch. Neurol.58, 1985 (2001).10.1001/archneur.58.12.1985 11735772

63.McKhann GM The diagnosis of dementia due to Alzheimer’s disease: recommendations from the National Institute on Aging-Alzheimer’s Association workgroups on diagnostic guidelines for Alzheimer’s disease Alzheimer’s. Dement. 2011 7 263 269 10.1016/j.jalz.2011.03.005 PMC3312024 21514250 McKhann, G. M. et al. The diagnosis of dementia due to Alzheimer’s disease: recommendations from the National Institute on Aging-Alzheimer’s Association workgroups on diagnostic guidelines for Alzheimer’s disease. Alzheimer’s. Dement.7, 263–269 (2011).10.1016/j.jalz.2011.03.005 PMC3312024 21514250

64.Townley RA Progressive dysexecutive syndrome due to Alzheimer’s disease: a description of 55 cases and comparison to other phenotypes Brain Commun. 2020 2 fcaa068 32671341 10.1093/braincomms/fcaa068 Townley, R. A. et al. Progressive dysexecutive syndrome due to Alzheimer’s disease: a description of 55 cases and comparison to other phenotypes. Brain Commun.2, fcaa068 (2020).10.1093/braincomms/fcaa068 PMC7325839 32671341

65.Gorno-Tempini ML Classification of primary progressive aphasia and its variants Neurology 2011 76 1006 1014 21325651 10.1212/WNL.0b013e31821103e6 Gorno-Tempini, M. L. et al. Classification of primary progressive aphasia and its variants. Neurology76, 1006–1014 (2011).10.1212/WNL.0b013e31821103e6 PMC3059138 21325651

66.Crutch SJ Consensus classification of posterior cortical atrophy Alzheimer’s Dement. 2017 13 870 884 10.1016/j.jalz.2017.01.014 PMC5788455 28259709 Crutch, S. J. et al. Consensus classification of posterior cortical atrophy. Alzheimer’s Dement.13, 870–884 (2017).10.1016/j.jalz.2017.01.014 PMC5788455 28259709

67.Rascovsky K Sensitivity of revised diagnostic criteria for the behavioural variant of frontotemporal dementia Brain 2011 134 2456 2477 21810890 10.1093/brain/awr179 Rascovsky, K. Sensitivity of revised diagnostic criteria for the behavioural variant of frontotemporal dementia. Brain134, 2456–2477 (2011).10.1093/brain/awr179 PMC3170532 21810890

68.Höglinger GU Clinical diagnosis of progressive supranuclear palsy: the movement disorder society criteria Mov. Disord. 2017 32 853 864 28467028 10.1002/mds.26987 Höglinger, G. U. et al. Clinical diagnosis of progressive supranuclear palsy: the movement disorder society criteria. Mov. Disord.32, 853–864 (2017).10.1002/mds.26987 PMC5516529 28467028

69.Jack CR NIA-AA research framework: toward a biological definition of Alzheimer’s disease Alzheimer’s. Dement. 2018 14 535 562 10.1016/j.jalz.2018.02.018 PMC5958625 29653606 Jack, C. R. et al. NIA-AA research framework: toward a biological definition of Alzheimer’s disease. Alzheimer’s. Dement.14, 535–562 (2018).10.1016/j.jalz.2018.02.018 PMC5958625 29653606

70.Roberts RO The Mayo Clinic study of aging: design and sampling, participation, baseline measures and sample characteristics Neuroepidemiology 2008 30 58 69 18259084 10.1159/000115751 Roberts, R. O. et al. The Mayo Clinic study of aging: design and sampling, participation, baseline measures and sample characteristics. Neuroepidemiology30, 58–69 (2008).10.1159/000115751 PMC2821441 18259084

71.Jack CR An operational approach to National Institute on Aging-Alzheimer’s Association criteria for preclinical Alzheimer disease Ann. Neurol. 2012 71 765 775 22488240 10.1002/ana.22628 Jack, C. R. et al. An operational approach to National Institute on Aging-Alzheimer’s Association criteria for preclinical Alzheimer disease. Ann. Neurol.71, 765–775 (2012).10.1002/ana.22628 PMC3586223 22488240

72.Jack CR Defining imaging biomarker cut points for brain aging and Alzheimer’s disease Alzheimer’s. Dement. 2017 13 205 216 10.1016/j.jalz.2016.08.005 PMC5344738 27697430 Jack, C. R. et al. Defining imaging biomarker cut points for brain aging and Alzheimer’s disease. Alzheimer’s. Dement.13, 205–216 (2017).10.1016/j.jalz.2016.08.005 PMC5344738 27697430

73.Klunk WE The centiloid project: standardizing quantitative amyloid plaque estimation by PET Alzheimer’s. Dement. 2015 11 1 15.e4 10.1016/j.jalz.2014.07.003 PMC4300247 25443857 Klunk, W. E. et al. The centiloid project: standardizing quantitative amyloid plaque estimation by PET. Alzheimer’s. Dement.11, 1–15.e4 (2015).10.1016/j.jalz.2014.07.003 PMC4300247 25443857

74.Schwarz CG Considerations for performing Level-2 centiloid transformations for amyloid PET SUVR values Sci. Rep. 2018 8 7421 10.1038/s41598-018-25459-9 29743601 Schwarz, C. G. et al. Considerations for performing Level-2 centiloid transformations for amyloid PET SUVR values. Sci. Rep.8, 7421 (2018).10.1038/s41598-018-25459-9 PMC5943521 29743601

75.Joie R Multisite study of the relationships between antemortem [11C]PIB-PET centiloid values and postmortem measures of Alzheimer’s disease neuropathology Alzheimer’s Dement. 2019 15 205 216 10.1016/j.jalz.2018.09.001 PMC6368897 30347188 Joie, R. et al. Multisite study of the relationships between antemortem [11C]PIB-PET centiloid values and postmortem measures of Alzheimer’s disease neuropathology. Alzheimer’s Dement.15, 205–216 (2019).10.1016/j.jalz.2018.09.001 PMC6368897 30347188

76.Iaccarino L rPOP: Robust PET-only processing of community acquired heterogeneous amyloid-PET data NeuroImage 2022 246 118775 34890793 10.1016/j.neuroimage.2021.118775 Iaccarino, L. et al. rPOP: Robust PET-only processing of community acquired heterogeneous amyloid-PET data. NeuroImage246, 118775 (2022).10.1016/j.neuroimage.2021.118775 34890793

77.Lowe VJ Neuroimaging correlates with neuropathologic schemes in neurodegenerative disease Alzheimer’s Dement. 2019 15 927 939 10.1016/j.jalz.2019.03.016 PMC6662599 31175025 Lowe, V. J. et al. Neuroimaging correlates with neuropathologic schemes in neurodegenerative disease. Alzheimer’s Dement.15, 927–939 (2019).10.1016/j.jalz.2019.03.016 PMC6662599 31175025

78.Villemagne VL CenTauR: toward a universal scale and masks for standardizing tau imaging studies Alzheimer’s Dement. Diagnosis Assess. Dis. Monit. 2023 15 e12454 10.1002/dad2.12454 PMC10326476 37424964 Villemagne, V. L. et al. CenTauR: toward a universal scale and masks for standardizing tau imaging studies. Alzheimer’s Dement. Diagnosis Assess. Dis. Monit.15, e12454 (2023).10.1002/dad2.12454 PMC10326476 37424964

79.Leuzy A Harmonizing tau positron emission tomography in Alzheimer’s disease: the CenTauR scale and the joint propagation model Alzheimer’s Dement. 2024 20 5833 5848 10.1002/alz.13908 PMC11497758 39041435 Leuzy, A. et al. Harmonizing tau positron emission tomography in Alzheimer’s disease: the CenTauR scale and the joint propagation model. Alzheimer’s Dement.20, 5833–5848 (2024).10.1002/alz.13908 PMC11497758 39041435

Come citare questo studio

Nick Corriveau-Lecavalier, Gemeng Zhang, Hugo Botha, Leland R Barnard, Heather J Wiste, Ellen Dicks, Venkatsampath R Gogineni, Jeyeon Lee, Irene Sintini, Matthew L Senjem, Christopher G Schwarz, Petrice M Cogswell, Derek R Johnson, Jeffrey L Gunter, Michael G Kamykowski, Jonathan Graff-Radford, Vijay K Ramanan, Julie A Fields, Mary M Machulda, Nikki H Stricker, Bradley F Boeve, Val J Lowe, Kejal Kantarci, David S Knopman, Jennifer L Whitwell, Keith A Josephs, Ronald C Petersen, Clifford R Jack, David T Jones, or the Alzheimer’s Disease Neuroimaging Initiative. (2026). A model of brain function to predict tau across Alzheimer's disease and related disorders.. Nature communications. DOI: 10.1038/s41467-026-78300-7

Consulta la pubblicazione originale ↗