Mapping Rise-and-Fall Brain Network Changes onto a Continuous Alzheimer's Disease Clock

Alzheimer's disease unfolds over decades, yet studies of brain network function still sort patients into coarse diagnostic bins (healthy, mild cognitive impairment, dementia) and average connectivity over whole networks. Both habits hide the same signal. In early disease, some parts of a brain network become more synchronized, which is thought to be compensation, while neighboring parts decouple. Averaged over the network, these opposing changes cancel out. Averaged over a diagnostic bin, the brief compensatory phase blurs into its neighbors.

We propose to place every person on one continuous disease clock and to learn, for each functional subnetwork, how its connectivity changes along that clock. The curves are allowed to rise and then fall. The clock is anchored by markers known to worsen steadily, such as amyloid load, hippocampal volume and cognition. Functional subnetworks are then allowed to help locate people on the clock, especially people whose structural scans still look normal.

The idea in two sentences: steady markers fix the direction of the clock, and rise-and-fall functional markers, read at subnetwork resolution, add timing information exactly where steady markers are flat. The test is whether this places structurally normal but amyloid-positive people on the clock more accurately, judged by their later cognitive decline.

Core Claim

Continuous progression models in Alzheimer's disease have used only steadily worsening features. Monotonic Gaussian Process analysis factorizes MRI and PET into spatial maps with strictly monotone time courses plus per-subject time shifts. The regularized MLP of Wang et al. penalizes any decrease in predicted risk between visits. SuStaIn orders biomarker abnormality events from regional volumes. BrainODE integrates a learned rate of change of hippocampal and ventricle shape, using a neural ordinary differential equation (a network that outputs a derivative which is then integrated over time).

None of these uses functional connectivity, which is the correlation of blood-oxygen signals between brain regions at rest. A monotone constraint would also be wrong for it. Cross-sectional studies report early hyperconnectivity followed by breakdown, and they report it at the level of sub-parts of networks. Xue et al. find that the anterior default-mode network gains frontal connectivity in subjective cognitive decline while the posterior precuneus decouples, with only 67 subjects across four bins. The cognitive-control dynamic connectivity study finds in OASIS-3 that time spent in a hippocampus-coupled state falls with dementia, using k-means states and diagnostic groups.

Our claim has two parts. First, a joint latent-time model can recover subnetwork-specific trajectories that are non-monotone and of opposite sign. These trajectories become nearly flat when averaged at the network level. Second, these trajectories carry information about disease time that steady markers lack. In amyloid-positive people with no detectable atrophy, the analogue of the 58% "Subtype Zero" in Shand et al., functional subnetworks should predict who declines next.

The second part is not obvious. A rise-and-fall curve alone is ambiguous, because one value can occur at two times. Our bet is that the steady anchors resolve which side of the peak a person is on, and the functional marker then resolves where on that side.

Research Program

Each subject $i$ has visits at ages $t_{ij}$. Each visit is mapped to disease time $s_{ij} = \gamma_i (t_{ij} - \tau_i)$, where $\tau_i$ is a subject-specific onset shift and $\gamma_i > 0$ is a progression rate. This is the same kind of time reparameterization MGPA uses. Each marker $k$ is modeled as

$$y_{ijk} = f_k(s_{ij}) + b_{ik} + \varepsilon_{ijk},$$

where $f_k$ is the population trajectory of marker $k$, $b_{ik}$ is a subject offset, and $\varepsilon_{ijk}$ is noise. The noise variance is estimated per marker from repeat scans so that unreliable connectivity features are down-weighted.

For anchor markers we force $f_k$ to be monotone. These are amyloid PET load, hippocampal volume and a cognitive composite. For functional markers we place an unconstrained Gaussian process prior on $f_k$, which is a smooth random function with no imposed shape. Inference is variational over $\tau_i$, $\gamma_i$ and the curves.

The functional markers are subnetwork connectivity summaries extracted in two ways, so that no single parcellation drives the result. The first is high-order independent component analysis, the decomposition used by Sui et al.. The second is Schaefer-atlas subnetworks, as in SLMIL-GCTrans. The summaries cover:

  • within- and between-subnetwork connectivity for the anterior and posterior default-mode network, cognitive-control sub-components, limbic and dorsal-attention coupling, and default-mode to control coupling;
  • dynamic-state occupancy rates computed as in the cognitive-control dynamic connectivity study.

Three analyses follow. The first estimates each subnetwork trajectory and asks where its peak falls relative to the amyloid and atrophy curves. The second compares held-out fit and forecasting for three versions of the model: the full model, the same model with monotone functional curves, and the same model with network-averaged rather than subnetwork features. The third examines structurally normal amyloid-positive subjects. We ask whether their inferred disease times, now informed by connectivity, spread out in a way that predicts their later decline. We will run scanner and site harmonization before fitting and report motion-matched sensitivity analyses. Connectivity test-retest noise is the most likely way this project fails.

Baselines

The naive reference is diagnosis-only staging. Each subject's disease time is the ordinal diagnosis (normal, early MCI, late MCI, dementia) plus age. Decline is forecast from group-mean slopes. Any method that cannot beat this inside the structurally normal amyloid-positive subgroup has learned nothing about early timing. All diagnostic labels are identical within that group, so this baseline predicts no within-group spread at all.

The strongest existing continuous progression method is MGPA, refit on the same ADNI and OASIS-3 subjects using only monotone structural, amyloid and cognitive features, with the same visit splits. It reports correlations of time shift with ADAS11, FAQ and MMSE, but the file contains no single headline number. For context, SuStaIn reports PACC decline of -0.24 per year in its cortical subtype (Shand et al.). We will run SuStaIn on our cohort as a second structural staging baseline.

For conversion prediction, the strongest reported number is HiMAL's AUROC of 0.923 and AUPRC of 0.623 for 6-month MCI-to-AD conversion (HiMAL). That model is volume- and cognition-based, and that task is not ours. We will retrain HiMAL on our subject set with matched inputs plus connectivity features rather than compare against its published figure. On the functional side, SLMIL-GCTrans reports 73.1% accuracy and 72.9% AUC for early MCI versus controls on ADNI. We will retrain it on our splits as the best cross-sectional connectivity classifier. For progression prediction specifically, no reported result was found on Lacuna for continuous functional-connectivity trajectory forecasting.

Two ablations isolate the key ingredients:

  • Monotone constraint ablation: the full model with monotone functional curves isolates the value of allowing rise-and-fall shapes.
  • Resolution ablation: the full model with network-averaged connectivity isolates the value of subnetwork resolution.

All models get the same features, the same subject-level train and test splits, and an equal hyperparameter search budget. Beating the field means two things. In the structurally normal amyloid-positive subgroup, forecasts of 2-year PACC or ADAS13 decline must have lower held-out error than both MGPA and HiMAL, with paired bootstrap confidence intervals excluding zero. Overall forecasting must be no worse than MGPA.

Why not just…?

Why not just run SuStaIn or an event-based model on connectivity features? These models treat each feature as passing once from normal to abnormal. A subnetwork that rises and then returns toward normal would be scored as "abnormal" early and "normal" late, which scrambles the event order. We will still run this as a baseline on the same features. If it matches our forecasts, the non-monotonicity claim is weak for practical staging.

Why not just regress connectivity on an existing disease time, such as MGPA time shifts or centiloid amyloid, with splines? This is the real threat for the descriptive half of the claim. A two-stage spline fit could recover the same rise-and-fall curves. We concede that point and include the two-stage approach as a baseline for curve estimation. What it cannot do is let connectivity move a person's estimated disease time, because in the two-stage approach the time is fixed before connectivity is seen. The decisive test is therefore the structurally normal subgroup. If two-stage times forecast decline as well as jointly inferred times, connectivity adds description but not staging, and we will report it that way.

Why not just train a deep connectivity classifier for conversion? Classifiers such as SLMIL-GCTrans or HOPE (72.0% accuracy on ADNI, 67.6% on NACC) answer a binned question. They inherit the within-label variation documented by Kang et al. and do not yield trajectories or peak timing. We compare against them on conversion AUROC. Our claim does not rest on beating them there. It rests on producing interpretable, testable timing that they cannot produce.

Proposed Experiments

We will use ADNI participants with resting fMRI (ADNI GO, 2 and 3), keeping the MCI subjects that BrainODE excluded. OASIS-3 serves as an independent cohort; it provides the 1,385 resting-fMRI subjects analyzed by the cognitive-control dynamic connectivity study, and we will use its amyloid PET where available. We train on one cohort and test on the other in both directions.

The first decisive check is qualitative but falsifiable. The model should recover, in continuous time, the opposite-sign anterior and posterior default-mode trajectories reported by Xue et al. The anterior curve should peak before atrophy onset, and the network-averaged curve should be flat within its credible band. If the subnetwork curves are flat as well, the heterogeneity premise fails at this resolution.

The second check is quantitative. In amyloid-positive subjects at structural stage zero, jointly inferred disease time should correlate with subsequent PACC or ADAS13 slope more strongly than MGPA time, two-stage time and SuStaIn stage. The difference must hold under a paired bootstrap. This motivates by analogy to the subthreshold-amyloid decline reported by Kim et al..

The third check covers the ablations. The model should achieve lower held-out marker likelihood than its monotone version, and better forecasts than its network-averaged version. Both comparisons are repeated under the ICA and Schaefer parcellations.

The fourth check is a robustness test. The estimated peak times must agree across ADNI and OASIS-3 within their credible intervals. Agreement would show that the rise-and-fall shape reflects biology rather than scanner or site effects.