Alzheimer's disease damages the brain for years before anyone notices. One explanation is compensation: as local tissue fails, the network shifts work elsewhere and behavior stays intact for a while. If that shift is real, it should show up as a slow spatial migration of each person's functional map, meaning the cortical locations that carry a given connectivity pattern move over time. This proposal measures that migration directly. We align each person's later resting-state fMRI scan to their own earlier scan using optimal transport, a method that finds the cheapest way to move one distribution of signal onto another across the cortical surface. We then ask whether the amount of movement, and the signal that cannot be matched at all, predicts conversion from mild cognitive impairment (MCI) to Alzheimer's disease.
The design also responds to a robustness problem. Connectome-based predictors fail badly when moved to new datasets. Under leave-one-dataset-out testing they often explain less variance than predicting the mean (Benchmarking External Generalization of SPD Matrix Learning for Resting-State fMRI Connectome Prediction). A person scanned twice on the same scanner is their own reference. Scanner, protocol, anatomy, and demographics are largely held fixed, so a within-person change measure should carry much less site signature than any feature computed against a population template. The core claim is that self-referenced functional drift is both an early prognostic signal and a more transportable one than absolute connectivity. Unlike a simple change in connectivity values, the transport-based measure separates "this location's pattern changed" from "this pattern moved to a different location."
Why this and not the obvious idea
The obvious combination of these directions is to use functional alignment as cross-site harmonization. Every subject would be hyperaligned to a common template, and diagnosis would be expected to improve. Two findings in the file argue against this.
First, hyperalignment applied to resting-state data in psychosis gave the same classification accuracy as unaligned data, roughly 66% (Application of hyperalignment to resting state data in individuals with psychosis reveals systematic changes in functional networks and identifies distinct clinical subgroups). The diagnostic signal lay in the alignment parameters themselves: reduced scale (t = −2.69) and increased translation (t = 2.90). That paper motivates treating the transform as the measurement. It also constrains us, because its scale and translation terms are exactly the kind of global magnitude quantities that scanner signal-to-noise differences would distort in a multi-site setting.
Second, the leave-one-dataset-out collapse in the connectome benchmark was driven more by population mismatch than by scanner differences. Aligning brains to a template cannot fix a population mismatch.
Our proposal changes the reference from a population template to the person's own baseline. It changes the quantity from global scale and rotation to a spatially resolved displacement on the cortex. It also adds a test that separates scanner shift from population shift.
The alignment method comes from fused unbalanced Gromov-Wasserstein transport, or FUGW (Aligning individual brains with fused unbalanced Gromov Wasserstein). FUGW couples distances along the cortical surface with functional features and allows signal to be created or destroyed when regions differ in size. On healthy Individual Brain Charting data it raised inter-subject correlation from 0.258 to 0.356, but it was never applied to clinical, longitudinal, or multi-site data. Its amortized successor ULOT (Unsupervised Learning for Optimal Transport plan prediction between unbalanced graphs) predicts transport plans up to 100 times faster. It matches exact FUGW loss at a Pearson correlation of 0.99 and scales to roughly 10,000 nodes. That makes the cost of a whole longitudinal cohort manageable.
Core Claim and Measurement
For subject $s$ with resting-state scans at visits $t_0$ and $t_1$, each cortical vertex gets a feature vector: its connectivity profile to a fixed set of seed parcels, computed in the subject's own native surface space. We then solve FUGW between the two visits and obtain a transport plan $P_{ij}$. This plan says how much of baseline vertex $i$'s mass maps to follow-up vertex $j$, with $d_{ij}$ the geodesic distance between them on the subject's own baseline surface. Two maps summarize the plan. The first is a displacement map
$$D_i = \frac{\sum_j P_{ij}\, d_{ij}}{\sum_j P_{ij}},$$
which is the average distance vertex $i$'s functional pattern traveled. The second is an unmatched-mass map $U_i = 1/n - \sum_j P_{ij}$, which is the share of vertex $i$'s pattern that has no counterpart at follow-up. In this map $n$ is the number of vertices and the baseline mass is uniform.
We hypothesize that displacement concentrated in temporal and default-mode territory, scaled by the time between visits, reflects active compensation and predicts conversion. We hypothesize that unmatched mass reflects outright loss.
The temporal-lobe prior comes from event-based modeling of diffusion connectomes, which found left temporal pole and transverse temporal centrality decline before peak amyloid and hippocampal atrophy (Data-Driven Sequence of Changes to Anatomical Brain Connectivity in Sporadic Alzheimer's Disease). That work used a fixed anatomical atlas and structural connectivity only. We test whether the functional analogue moves earlier or more informatively.
Working within each subject also avoids a known weakness of spatial normalization. Warping atrophied brains onto a healthy template can stretch damaged tissue and confuse registration artifacts with biology. Here the template is never used for the drift measurement. Maps are brought to a common surface only afterwards, for group statistics.
Research Program
The first stage builds and calibrates the measurement. We reproduce the FUGW result on Individual Brain Charting as a correctness check. The entropic smoothing parameter and the anatomy-versus-function weight are fixed using a noise-floor criterion, never using clinical outcomes, since excessive smoothing is known to blur the maps. The criterion is that drift between split halves of a single session, and between visits in cognitively stable controls, should be minimal while still recovering held-out task-map correspondence. We work per hemisphere at roughly 10,000 vertices, within the reported reach of ULOT, and confirm ULOT plans against exact FUGW on a subsample.
Short resting-state runs are known to destabilize alignment. We therefore report how drift reliability depends on scan duration by truncating longer runs.
The second stage is a site-sensitivity audit. ADNI is a multi-vendor consortium in which some participants changed scanners between visits, which gives a natural experiment. Among stable controls, we compare drift in subjects with and without a scanner change. We compare the size of that difference with the gap between later converters and non-converters. If a scanner change produces drift as large as conversion does, the measure is not usable without correction, and we will report that outcome.
The third stage is prognosis. Regional drift and unmatched-mass summaries enter a Random Survival Forest and a Cox model for time to MCI-to-AD conversion. Clinical covariates and, where available, plasma p-Tau217 are included. We evaluate four-year AUROC and Harrell's C-index.
Proposed Experiments
The primary cohort is ADNI MCI participants with at least two resting-state scans. We first estimate how many of them overlap with the 201-person p-Tau217 sample of Predicting Progression from Mild Cognitive Impairment to Alzheimer's Disease using Survival Analysis. That overlap may be small. If so, the p-Tau comparison becomes a secondary analysis, reported with bootstrap confidence intervals and without claims of superiority.
There are three evaluation splits. The first is nested cross-validation within ADNI. The second is leave-one-scanner-vendor-out within ADNI, which keeps the population roughly fixed and isolates hardware shift. The third is training on ADNI and testing on OASIS-3, which adds population and protocol shift; both cohorts appear in the connectome generalization benchmark. In OASIS-3 the outcome is clinical progression as recorded by that cohort. The main robustness quantity is the drop in C-index from within-ADNI to each external split, compared across feature sets.
Baselines
The naive reference point is a survival model with only age, sex, APOE genotype, education, and baseline cognitive scores. Any imaging feature must add to it.
The strongest existing method is a Random Survival Forest on plasma p-Tau217, which reached a four-year AUROC of 0.787 on ADNI MCI. Amyloid PET Centiloids reached 0.758 in the same study (Predicting Progression from Mild Cognitive Impairment to Alzheimer's Disease using Survival Analysis). We rerun it on exactly our subjects and folds rather than quoting 0.787 across different samples. Success means a statistically supported C-index gain when drift is added to p-Tau217 plus clinical covariates. That study also noted that individuals above 0.56 pg/mL reach a plateau that is hard to stratify further. We separately test discrimination within that high-p-Tau group, where imaging has the most room to help.
The strongest connectome baseline uses baseline functional connectivity projected to tangent space, harmonized with Riemannian ComBat, and fed to a ridge-penalized Cox model. This is the configuration that the connectome benchmark found most stable among its methods (Benchmarking External Generalization of SPD Matrix Learning for Resting-State fMRI Connectome Prediction). That benchmark targeted brain age, so there is no reported result on Lacuna for this baseline on MCI conversion. Beating it means a higher external C-index and a smaller within-to-external drop.
A standard structural comparator, the hippocampal volume change rate, is also included.
Two ablations isolate the key ingredient. The identity-transport ablation sets $P$ to the identity, so drift reduces to the per-vertex change in connectivity profile with no spatial movement allowed. If it matches the full method, the transport adds nothing and the "migration" story is false. The template-reference ablation aligns each visit to a population template, psychosis-paper style, and uses the transform parameters as features. If it transfers as well as self-reference, the site-cancellation argument is unnecessary.
All arms share subjects, folds, survival learners, and an equal hyperparameter search budget. That budget is chosen by inner-fold C-index.
Why not just…?
Why not just harmonize baseline connectivity with ComBat? ComBat removes per-site mean and variance shifts in connectivity features. The connectome benchmark found that it helped linear models but did not prevent collapse when held-out cohorts differed demographically. Its benefit also depended on conditioning on the biological variable, which is unknown for future converters. ComBat remains a serious competitor within ADNI, where the population is shared. The leave-one-vendor-out and ADNI-to-OASIS-3 splits are designed to show whether self-referencing degrades less.
Why not just use the within-person change in connectivity, with no optimal transport? This is the most serious threat. It inherits the same site cancellation and is far cheaper. The identity-transport ablation tests it directly. We commit to reporting the drift measure as unnecessary if the ablation matches it. Our prior that it will not match rests on the evidence that functional boundaries shift relative to anatomy (Function-based Intersubject Alignment of Human Cortical Anatomy; Aligning individual brains with fused unbalanced Gromov Wasserstein). A pattern that moves a few millimeters looks like decorrelation to a fixed-location measure and like low-cost displacement to transport.
Why not just use p-Tau217? A blood test is cheaper and already strong. Our claim is only incremental value and spatial localization, tested in the full sample and in the high-p-Tau plateau. Plasma markers index cumulative proteinopathy and cannot say where the network is reorganizing. If drift adds nothing beyond p-Tau217, the right conclusion is that resting-state fMRI is not worth acquiring for this purpose.
Why not just use surface-based anatomical registration between visits? Diffeomorphic surface matching enforces one-to-one maps and cannot represent regions shrinking or vanishing. It also tracks anatomy, which is the wrong thing to track when the hypothesis is that function moves while folds stay put. It cannot produce an unmatched-mass signal at all, and FUGW outperformed it on correspondence in healthy data.
Decisive Checks
Four results would settle the proposal.
First, drift in stable controls must be small relative to the converter versus non-converter difference. It must also not be inflated by scanner changes beyond the natural noise floor. If either fails, the measure is not viable.
Second, full transport must beat the identity-transport ablation on C-index. Otherwise the spatial-migration hypothesis fails.
Third, drift must add to p-Tau217 plus clinical covariates. Otherwise the measure has no clinical relevance.
Fourth, the external C-index drop for drift features must be smaller than for ComBat-harmonized baseline connectivity. Otherwise there is no robustness gain.
Passing all four would show that hidden compensation can be measured, and that measuring it within each person is what makes it transferable across sites.
References
- Benchmarking External Generalization of SPD Matrix Learning for Resting-State fMRI Connectome Prediction
- Application of hyperalignment to resting state data in individuals with psychosis reveals systematic changes in functional networks and identifies distinct clinical subgroups
- Aligning individual brains with fused unbalanced Gromov Wasserstein
- Unsupervised Learning for Optimal Transport plan prediction between unbalanced graphs
- Data-Driven Sequence of Changes to Anatomical Brain Connectivity in Sporadic Alzheimer’s Disease
- Predicting Progression from Mild Cognitive Impairment to Alzheimer’s Disease using Survival Analysis
- Function-based Intersubject Alignment of Human Cortical Anatomy