Can a brain function model predict Alzheimer's-type tau across disorders?

Predicting tau from metabolism alone

The paper asks whether a single, widely used imaging input — FDG-PET, which measures glucose metabolism — can stand in for a tau tracer that is harder to obtain. The target is Alzheimer's-type tau, measured with flortaucipir-PET (FTP-PET), across Alzheimer's disease and related disorders (ADRD), a spectrum that includes typical and atypical AD, dementia with Lewy bodies, and frontotemporal lobar degeneration syndromes.

The model is a population-level regression framework. It does not simulate protein spreading region by region. Instead, it learns, across a large cohort, how patterns of whole-brain metabolism co-vary with regional tau uptake, then applies those learned weights to a new participant's FDG-PET scan. The authors describe the output as an estimate of participant-level spatial distributions of FTP-PET burden, not a direct measurement of tau.

That framing matters for how the results should be read. A high correlation means the model reproduces the spatial pattern of tau that the tracer would have shown; it does not mean the model detected tau protein. The paper's own interpretation is that predictions are driven by specific patterns of whole-brain metabolism rather than by translating regional FDG-PET signal into spatially equivalent FTP-PET values.

Eigenbrains compress a scan into 150 numbers

The input features come from StateViewer, a clinical decision support system derived from a principal component decomposition of 1000 randomly sampled FDG-PET scans. That decomposition yields 150 components, called eigenbrains, which act as a low-dimensional latent space describing dominant patterns of brain metabolism. Each participant's scan is projected onto these components, producing a set of eigenvalues — one number per eigenbrain — that summarize where that person's metabolism sits relative to the population.

The team then fitted linear regression models relating these eigenvalues to FTP-PET uptake in each region of the Schaefer-400 atlas, adding eigenbrains iteratively and selecting the model with the best median leave-one-out cross-validation R² in the training set. The best model used 50 eigenbrains. Models for the CenTauR atlas, used for tau staging, used the same number of eigenbrains.

The authors also benchmarked this linear approach against five non-linear and regularized machine learning models, and against a base model using age, sex, APOE4 carriership, clinical diagnosis, and CDR-SB instead of FDG-PET eigenvalues. Reporting that comparison matters because it separates the value of the metabolic features from the value of readily available clinical variables.

Strong fit inside the 1088-person training cohort

The training set was 1088 Mayo Clinic research participants spanning normal aging, mild cognitive impairment, AD, dementia with Lewy bodies, and frontotemporal dementia spectra, all with FDG-PET, FTP-PET, and amyloid-PET. In this cohort the 50-eigenbrain model reached a median R² of 0.71 (IQR 0.61–0.77) across the Schaefer-400 atlas, with a cross-validated R²_pred of 0.65 (IQR 0.54–0.73), mean absolute error of 0.09, and mean squared error of 0.01.

Two different R² values appear here and should not be conflated. The 0.71 figure describes correspondence across regions and participants pooled at the atlas level; the participant-level spatial R², which compares observed and predicted values across the 400 regions within each person, had a median of 0.79 (IQR 0.73–0.84). The first asks whether the model gets regional tau levels right across the cohort; the second asks whether it reproduces each person's spatial pattern.

Performance was uneven by group. Spatial R² was markedly higher in groups with a wider FTP-PET dynamic range, such as dysexecutive AD and posterior cortical atrophy, than in groups with minimal FTP-PET signal, such as cognitively unimpaired groups, semantic dementia, and progressive supranuclear palsy. Bland-Altman plots showed that high-tau groups (amnestic AD, dysexecutive AD, logopenic variant primary progressive aphasia) had slight proportional bias, with overestimation at lower mean values and underestimation at higher ones.

Performance drops but holds in independent cohorts

Validation used three out-of-sample settings. In a longitudinal subset of the training cohort (215 participants, 478 time points), in the Alzheimer's Disease Neuroimaging Initiative cohort (295 participants), and in a clinically heterogeneous Mayo Alzheimer's Disease Treatment Clinic cohort (264 patients, of whom 28 had FTP-PET), the Mayo-trained model was applied using each participant's own StateViewer eigenvalues.

The abstract summarizes participant-level prediction across these settings as R² of roughly 0.34 to 0.71. The cohort-level numbers show the spread. In ADNI, mixed-effect models across the Schaefer-400 atlas reached R² = 0.33 (IQR 0.21–0.41), MAE = 0.11, and MSE = 0.02, while individual spatial R² had a median of 0.72. In the 28 ADTC patients with FTP-PET, the corresponding values were R² = 0.41 (IQR 0.27–0.52), MAE = 0.12, MSE = 0.02, and a median individual spatial R² of 0.80.

The pattern is consistent: the model reproduces the spatial layout of tau within a person better than it reproduces absolute tau levels across a cohort. The ADNI and ADTC cohort-level R² values sit well below the training cohort's 0.71, which is the expected cost of moving to independent data and different scanners and protocols. The ADTC estimate also rests on only 28 people with the reference tracer.

Staging accuracy depends on who is included

The clinically actionable claim is tau staging under the revised AD diagnostic criteria, which combine amyloid status with CenTauR values in mesial temporal and temporoparietal regions to define stages A through D. The model regressed FDG-PET eigenvalues onto CenTauR values, and predicted stages were compared with observed stages using a confusion matrix.

In ADNI, collapsing stages into A/B versus C/D and excluding amyloid-negative individuals, agreement was 79%: 82 cells correctly labeled A/B, 41 correctly labeled C/D, 12 incorrectly labeled A/B, and 22 incorrectly labeled C/D. Including amyloid-negative individuals raised accuracy to 89%. The abstract reports the overall range as 79–95% across cross-sectional and longitudinal analyses in training and validation cohorts.

The gap between 79% and 89% is not a minor detail. Adding amyloid-negative participants makes the classification task easier, because those individuals are largely low-tau by definition, so the higher figure should not be read as better discrimination among the ambiguous cases that matter clinically. The paper also states that stage agreement was not assessed in the ADTC cohort because too few participants had FTP-PET and class imbalance was significant.

Group differences appear even without tau PET

The ADTC cohort is where the model's practical use case is most visible, because most of those patients had no FTP-PET. Predicted CenTauR values differed significantly across clinical groups in the mesial temporal area, F(4, 259) = 14.7, P < 0.001, the temporoparietal area, F(4, 259) = 23.4, P < 0.001, and the temporal meta-ROI, F(4, 259) = 20.1, P < 0.001.

Post-hoc comparisons showed the atypical AD amyloid-positive group had higher predicted values than all other groups across all regions, and the AD amyloid-positive group had higher values than the non-AD amyloid-negative group in the mesial temporal lobe and temporal meta-ROI. In ADNI, predicted values were higher in Dementia A+T+ than in CU A-T-, CU A+T-, Dementia A-T-, and all MCI groups, and higher in MCI A+T+ than in MCI A-T- and MCI A+T-.

These are group-level comparisons of predicted values, not validated individual diagnoses. The authors interpret them as suggesting the model could help flag symptomatic individuals whose symptoms may be driven by co-pathology, and could identify rare AD phenotypes traditionally attributed to non-AD etiologies. That is a hypothesis about future use, supported by group separation rather than by individual-level clinical validation.

Low predictions in primary tauopathies support specificity

The model predicted low FTP-PET in non-AD primary tauopathies such as progressive supranuclear palsy and corticobasal syndrome. The authors attribute this to the lower binding sensitivity of existing FTP-PET tracers to non-AD tau, and read it as evidence that the model is specific to AD-type tau rather than a generic neurodegeneration detector.

This interpretation is plausible within the paper's logic but rests on the tracer's known behavior rather than on an independent tau measure. It also means the model's low predictions in these syndromes cannot be used to rule out non-AD tau pathology; they reflect what the reference tracer would show.

The broader mechanistic framing is the principle of shared vulnerability across large-scale brain systems: functional gradients that support specific cognitive domains are thought to shape where neurodegeneration appears across dementia syndromes. The paper presents the model as consistent with that framework, not as a test that establishes it.

A scalable surrogate, not a tau measurement

The paper's practical claim is that FDG-PET, already used clinically to evaluate degenerative cognitive disorders, can be repurposed to approximate AD-type tau burden when FTP-PET is unavailable or burdensome. The strongest support is the within-person spatial agreement, with median spatial R² of 0.72 in ADNI and 0.80 in the 28 ADTC patients with FTP-PET, plus the 79–95% staging accuracy range.

The boundaries are equally specific. Cohort-level R² in independent data was 0.33 in ADNI and 0.41 in ADTC, well below the training cohort's 0.71, so absolute tau levels transfer less reliably than spatial patterns. Stage agreement was never tested in the ADTC cohort. High-tau groups showed proportional bias, with underestimation at the highest values — precisely the range where clinical decisions are hardest. And the model was trained on Mayo research participants, so its behavior in other populations and scanner protocols is inferred from ADNI and ADTC rather than established broadly.

The most useful way to read this paper is as a comparison between two kinds of accuracy. Within a person, the model recovers the spatial distribution of AD-type tau well enough to support staging and group-level stratification. Across people and cohorts, its absolute predictions are looser, and the 79% versus 89% staging gap shows how much the reported accuracy depends on whether easy amyloid-negative cases are included. The model's value is highest where FTP-PET is missing and the question is whether tau burden is low or high, and lowest where an exact tau value at the upper end would change management.

Sources used in this answer