Forecasting Individual Tau-PET Change on a Shared Amyloid-Anchored Disease Clock

Tau-PET scans are the most direct living measurement of where Alzheimer's disease is doing damage. PET (positron emission tomography) images a radioactive tracer that sticks to a target protein. Tau tangles, unlike the diffuse amyloid plaques that precede them, accumulate focally in the regions that later fail clinically. A clinician looking at one or two tau scans wants to know where tau will be in two to six years, and whether a change between scans is real progression or noise. Current tools answer this badly. Generic image-sequence forecasters often fail to beat simply copying the last scan forward. Population staging models describe the average patient but do not forecast individuals. Longitudinal tau-PET cohorts are also small: many subjects have one or two follow-ups, because each scan is expensive and adds radiation dose.

We propose to forecast each patient's regional tau trajectory by placing them on one shared, monotone disease clock, learned mostly from plentiful single-visit scans. The clock is anchored early by amyloid and later by tau. Only two numbers are fitted per patient: how far along the clock they are, and how fast they move along it. The bet is that, with scarce longitudinal data, nearly all forecastable individual change comes from position and speed on a common trajectory. A model with two free parameters per patient should therefore beat both the copy-the-last-scan baseline and high-capacity voxel forecasters, especially at long horizons and when only one prior scan exists.

Core Claim

Let $y_{ir}(a)$ be the tau standardized uptake value ratio (SUVR, regional tracer uptake divided by uptake in a cerebellar reference region) for patient $i$ in brain region $r$ at age $a$. We model

$$y_{ir}(a) = f_r\big(\gamma_i\,(a - \tau_i)\big) + b_{ir} + \varepsilon_{ir},$$

where $f_r$ is a learned nondecreasing curve for region $r$, shared across all patients and parameterized as a small monotone network. Here $\tau_i$ is the patient's time shift, $\gamma_i > 0$ is their speed, $b_{ir}$ is a static per-region offset capturing tracer off-target uptake and partial-volume bias, and $\varepsilon_{ir}$ is measurement noise. Amyloid SUVR gets its own curve $f_A$ on the same clock, which pins down early disease time before tau has risen.

The claim is that this structure captures most of the forecastable change. If it holds, it gives a practical recipe: learn the shapes $f_r$ from thousands of single-visit scans, and use the scarce longitudinal pairs only to calibrate speed.

The idea builds on three lines of work. CNODE supplies the method. It learns patient-specific onset shifts and speeds to align 161 Parkinson's patients with irregular visits on a common timeline, but it uses MRI cortical thickness and a neural ODE trained only on longitudinal data. We move this alignment to molecular PET, impose monotonicity, use a second tracer as a clock anchor, and train the shapes on cross-sectional data. The temporal event-based model on BioFINDER-2 motivates the shared clock. It shows amyloid and tau unfold over about 20 years, with about 8 years from amyloid positivity to entorhinal tau and 2–3 year steps through neocortex. Its output is a discrete population ordering over five coarse regions, though, and it cannot forecast an individual's future values. CRONOS motivates beating the copy baseline: its continuous-time flow-matching forecaster reached NRMSE 7.55 against 8.38 for copying the last scan on sparse glioma MRI. It also constrains our design, because it shows that even a well-built voxel generator only modestly beats persistence when data are sparse.

Research Program

We will build the shared-clock model on regional tau and amyloid SUVRs over a standard parcellation, with Braak-stage composite regions as the clinically read summary. Fitting alternates between two steps. First, the curves $f_r$ and $f_A$ are fitted on all scans, cross-sectional and longitudinal. Second, each patient's $(\tau_i, \gamma_i)$ is inferred with a population prior. For a patient with a single scan, $\gamma_i$ falls back to a prior predicted from covariates such as APOE ε4 status, age and amyloid level. With two or more scans it is updated from the observed change.

The model outputs a predictive distribution, not a point estimate. That lets us ask whether an observed change exceeds what measurement noise plus the patient's expected trajectory would produce, which is the "is this change real" question.

Two known weaknesses of PET quantification are handled explicitly rather than ignored. The first is off-target binding of first-generation tau tracers, which Park et al. show reflects MAO-B as well as tangles. The static offset $b_{ir}$ absorbs this, and we analyse the newer tracers MK6240 and RO948 separately. The second is that the reference region may acquire specific binding late in disease. We will rerun everything with an alternative reference region and report how much the late-stage forecasts move.

Baselines

The do-nothing reference is copying the last scan forward: predicted future regional SUVR equals the most recent observed value. The longitudinal literature identifies this as the bar most sequence models fail, and every result is reported as a skill score relative to it. No reported result found on Lacuna for copy-forward error on tau-PET regional forecasting. The only reported comparison against this baseline is CRONOS's NRMSE of 7.55 versus 8.38 on Lumiere glioma MRI (CRONOS), which is a structural, not molecular, setting.

The strongest simple competitor is a per-region linear mixed-effects model. It has random intercepts and slopes, with baseline tau SUVR, amyloid SUVR, APOE ε4 and age as fixed effects. It is fitted on exactly the same subjects, regions and preprocessing, and its covariates are selected by the same cross-validation used for our hyperparameters. It is the most dangerous baseline because, over two-year horizons, accelerating tau may be locally almost linear. Beating it means a lower mean absolute error of regional change at four- and six-year horizons and in the single-prior-scan regime, with paired bootstrap confidence intervals excluding zero.

The strongest existing deep method is CRONOS, retrained on co-registered 3D tau-PET volumes with its published architecture and time encoding. Its forecasts are reduced to the same regional SUVRs for scoring, so it is judged on the quantity clinicians read. It gets an equal GPU-hour budget for tuning. We also include a T-EBM-style stage-then-average baseline: assign the patient's stage at baseline, then add the population mean regional rate for that stage. This isolates whether individual speed matters beyond staging.

Three ablations isolate the key ingredients. The first removes the per-patient speed, setting $\gamma_i = 1$; the Neural Koopman Machine found that removing its patient-specific control term caused the largest drop in cognitive forecasting. The second removes the monotonicity constraint. The third trains the curves only on longitudinal subjects, discarding cross-sectional scans.

Why not just…?

Why not just fit linear mixed-effects models per region? This is a real threat. The TADPOLE results show simple methods and averaged ensembles winning most forecasting tasks, and cross-sectional data costing only 2 percentage points in diagnosis prediction. That pattern suggests most signal lies in where a patient currently is, which a mixed model with a baseline-SUVR covariate already captures. Our advantage should appear only where linearity breaks: long horizons and patients crossing from medial temporal to neocortical regions, where BioFINDER-2 shows acceleration. If the mixed model ties us there, we will report that the shared clock adds interpretability but no forecasting gain.

Why not just run CRONOS or another voxel flow model on tau-PET? Voxel generators spend capacity on anatomy that does not change. Here they would be trained on a few hundred longitudinal pairs, against the data-hungry regime they assume. TauPETGen (Jang et al.) shows that even 2D tau synthesis from 139 pairs needed an MRI anatomical prior to avoid incoherent brains. The CRONOS baseline run settles whether voxel detail buys anything once outputs are scored regionally.

Why not just stage patients with an event-based model and use stage-average rates? Staging discards within-stage speed differences. The stage-then-average baseline and the $\gamma_i = 1$ ablation measure exactly how much individual speed is worth.

Proposed Experiments

The primary cohort is BioFINDER-2: 960 individuals with amyloid- and tau-PET at 1, 2, 4 and 6 years. Its spread of horizons is what makes the long-horizon test possible. The model is trained on ADNI flortaucipir scans and tested on BioFINDER-2 and TRIAD (231 participants with MK6240 tau-PET), which measures transfer across tracer and site; we also report the reverse direction. Subjects, never scans, are held out.

The first decisive check is a learning curve over the number of longitudinal training subjects: 25, 50, 100, 200 and all. The constraint of Bria et al., that many medical applications have fewer than 100 subjects, sits on this curve. The prediction is that the shared-clock model's skill over copy-forward saturates by about 50 subjects, while CRONOS's skill is still climbing.

The second check is stratified error by horizon and by number of prior scans. The prediction is the largest margin over the mixed-effects model at six years with one prior scan.

The third check is calibration of the 80% prediction intervals for regional change. We also compute AUROC for predicting a Braak-stage advance by the next visit.

A final clinical check asks whether forecasted tau change explains later change in neuropsychiatric symptom scores in TRIAD beyond baseline Braak stage. Macedo et al. showed that baseline Braak stage alone predicts the symptom trajectory. Given TADPOLE's finding that cognitive forecasting was no better than chance, we treat a null here as plausible, and we will report it rather than tune toward it.