Hospitals that want to train a shared model on brain scans without pooling their data typically use federated learning, in which each site trains locally and a server combines the resulting models. The better versions of this combination are personalized. Each site receives a weighted blend of its peers' models, and more similar peers get more weight. The difficulty in neurodegeneration is that sites can look dissimilar for two unrelated reasons. One is technical: they use different scanners. The other is biological: their patients sit at different points of the disease. The biological case is unusual. Early in Alzheimer's-spectrum decline, the brain often raises connectivity between regions to compensate for damage. Later, that compensation collapses into low connectivity. A site full of early-stage patients and a site full of late-stage patients therefore deviate from healthy controls in opposite directions along the same brain networks.
The proposal is to compute client similarity from a within-site case-versus-control contrast rather than from raw model weights or raw features. Each site summarizes how its patients differ from its own healthy controls, measured in the shared model's internal representation. An additive scanner offset hits patients and controls at a site alike, so it largely cancels out of this difference. What survives is the disease signature, including whether that site's patients are in the hyperconnected compensatory phase or the collapsed phase. Sites are then weighted by how similar their contrasts are, so early-stage sites learn mostly from other early-stage sites whatever scanner they use. The central empirical question is whether this similarity measure groups sites by disease stage while existing measures group them by scanner vendor, and whether that difference improves detection of early mild cognitive impairment (MCI).
Core Claim
Existing similarity-weighted aggregation measures affinity with statistics that are dominated by whatever varies most across sites. FedCAP (FedCAP) uses cosine similarity of model updates and treats divergent updates as suspect. FedBip (Bi-level Personalization for Federated Foundation Models) uses layer-wise cosine similarity of task vectors, meaning fine-tuned minus initial weights. FedAux (Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections) learns a one-dimensional projection that provably converges to the principal eigenvector of each client's embedding covariance. In multi-site resting-state functional MRI, the dominant axis of variation is plausibly the scanner. We claim that conditioning the similarity signature on diagnosis, by taking the difference between patient and control representations within each site, removes most of that axis and exposes stage-dependent compensation as the thing that determines who learns from whom.
Concretely, let $h_i(x)$ be the pooled embedding that the current model at site $i$ assigns to subject $x$. Let $\mu_i^{P}$ and $\mu_i^{C}$ be the mean embeddings of that site's patients and controls, and let $\sigma_i^{C}$ be the per-dimension standard deviation over its controls. The site's contrast signature and the aggregation weights are
$$s_i = \frac{\mu_i^{P} - \mu_i^{C}}{\sigma_i^{C}}, \qquad W_{ij} = \frac{\exp(\cos(s_i, s_j)/T)}{\sum_k \exp(\cos(s_i, s_k)/T)},$$
where $T$ is a temperature and the division is elementwise. Dividing by the controls' spread is meant to absorb multiplicative scanner gain as well as additive offset. The signature is recomputed every round, so it tracks the representation as the model learns. We apply $W$ only to the upper, task-specific layers. The lower connectivity-encoding layers are averaged across all sites. This follows FedBip-L's observation that lower layers carry shared structure and higher layers carry task-specific structure.
Because the signed cosine can be negative, sites whose patients deviate in opposite directions receive little weight. This is the behavior the inverted-U predicts should be useful, and it is also the behavior most likely to be wrong, so we test it directly.
Grounding in the Two Literatures
The neuroscience literature motivates the design. Han Yingmei et al. (brain network imaging biomarkers of subjective cognitive decline; Research progress on brain network imaging biomarkers of subjective cognitive decline) review evidence that networks move from hyperconnectivity to a breakdown in efficiency once pathological load exceeds resilience. Wentao Hu et al. (Functional Segregation‐Integration Preference) find that integrated networks support resilience in mild and moderate but not severe small-vessel disease. Peirong Wu et al. (enhanced prefrontal subnetwork in MCI) report recruitment of the right dorsolateral prefrontal cortex in MCI. Together these predict that the sign of the patient-control difference in specific networks depends on stage, and a contrast signature is designed to be sensitive to exactly that.
These studies also constrain the proposal. The research file records that early compensators can pass clinical screening. Diagnostic labels at early stages are therefore noisy, and the contrast at a site dominated by early patients will be small and variable. We address this by requiring a minimum number of subjects per class before a site's signature is trusted. Until that threshold is met, the site falls back to uniform weights.
On the federated side, DAFed (Domain Adversarial Federated learning for multi-site functional connectivity; A Privacy-Preserving Domain Adversarial Federated learning for multi-site brain functional connectivity analysis) is the closest application. It trains a graph convolutional network, a neural network that operates on the region-by-region connectivity graph, federated across ABIDE sites and across ADNI-3 scanner vendors. It removes site information with a gradient-reversal adversary, which pushes the model's features to be uninformative about which site a scan came from. It then aggregates into a single global model. The contrast approach differs in two ways. It keeps a model per site rather than one global model. And instead of erasing between-site variance in general, it removes scanner variance and deliberately preserves stage variance.
UG-FedDA (Uncertainty-Guided Federated Domain Adaptation) weights clients by a single uncertainty number each and works on structural MRI. It cannot express which peers belong together. ClusMFL (ClusMFL) weights by modality availability. FedMEPD (FedMEPD) personalizes individual parameters using a gradient-consistency mask. None of the three computes pairwise, biologically conditioned affinity.
BioBGT (Biologically Plausible Brain Graph Transformer) supplies the local model. It is a centralized graph transformer that encodes hub importance and functional-module structure, and its code is available as pcyyyy/BioBGT. We use it unchanged at each site, so any gain can be attributed to aggregation rather than to the architecture.
Research Program
We will build a federated wrapper around BioBGT with swappable similarity signatures. The candidates are uniform weights, update cosine in the style of FedCAP, layer-wise task-vector cosine in the style of FedBip-L, a FedAux-style projection, and the contrast signature. We will also run the contrast signature restricted to module-level summaries. That variant uses Louvain modules from BioBGT's extractor, so its signature has one entry per module pair instead of one per embedding dimension. It is cheaper to communicate and easier to interpret.
Brain graphs will be constructed with a fixed atlas and pipeline drawn from DataCentricBrainGraphs, so that no method benefits from preprocessing choices.
The analysis has three parts. First, we measure what each signature's weight matrix groups sites by: scanner vendor or stage composition. Second, we measure whether that grouping changes per-site accuracy, especially for the early-MCI-versus-control task, which is where masking makes detection hardest. Third, we inspect which modules drive the signed contrast. The prediction is that sites dominated by early-stage patients show positive prefrontal and hub deviations, and later-stage sites show negative ones. Finding no sign flip would falsify the biological premise, even if accuracy improved.
The privacy cost is modest but not zero. The server sees one normalized mean-difference vector per site per round, which is lower-dimensional than a task vector. We will report a membership-inference attack against the signatures, analogous to the check FedAux ran on its projection vectors.
Baselines
The naive reference point has two parts. One is purely local training, in which each site trains BioBGT alone. The other is FedAvg over the same model. DAFed reports FedAvg at 66.9% accuracy on four-site ABIDE, against its own 76.4% (DAFed). The file reports no ADNI-3 MCI-versus-control number for DAFed ("no reported result found on Lacuna for DAFed on ADNI-3"). For context on difficulty, the only related ADNI figures are BioBGT's centralized three-way accuracy of 52.08% and UG-FedDA's structural-MRI NC-versus-MCI accuracy of 76.87%. Neither is directly comparable, because they use a different task and a different modality respectively.
The strongest existing competitor is DAFed-style adversarial invariance, because it targets exactly the confound we target: scanner versus biology. We will reimplement its gradient-reversal domain head on the same BioBGT backbone. On ABIDE, the reimplementation must reach approximately its reported 76.4% before any comparison is trusted.
The strongest personalized competitors are FedBip-L's layer-wise task-vector weighting and FedAux's learned projection. Both are ported to the same backbone and use the same layer split as our method.
All methods get the same subjects, splits, number of communication rounds, and local epochs. Each gets the same hyperparameter search budget, for example twenty configurations over learning rate and each method's own knob, selected on site-held-out validation data.
The key ablation replaces the contrast signature with the site's unconditioned mean embedding $\mu_i$. Everything else stays the same: normalization, temperature, and layer split. That comparison isolates diagnosis conditioning as the ingredient responsible for any gain. A second ablation clips negative cosines to zero, which tests whether down-weighting opposite-sign sites matters.
Beating the field means one outcome in particular: a significant gain over both DAFed and FedBip-L in mean per-site balanced accuracy on early MCI versus control, without loss in worst-site accuracy, across five seeds.
Why not just…?
Why not just harmonize the connectivity matrices first and then run FedAvg or FedBip? This is the most serious threat. Standard statistical harmonization removes additive and multiplicative site effects while preserving specified biological covariates, and if it succeeds, any similarity measure might then group sites by stage. We found no reported result for harmonized federated connectome classification on Lacuna. We will run a harmonize-then-FedBip-L arm. If it matches the contrast method, the honest conclusion is that the contrast signature is simply harmonization built into the aggregation step. The remaining advantage would then be only practical: harmonization typically needs a joint fit or shared statistics across sites, and our signature does not.
Why not just use adversarial domain invariance, as DAFed does? Its adversary removes everything that distinguishes sites. When stage mix is confounded with site, that includes the compensation signal. UG-FedDA's authors describe related negative transfer when disparate distributions are forced into alignment. The crossed vendor-by-stage design below tests this directly. DAFed should lose most on early-MCI sites whose vendor peers are mostly late-stage.
Why not just cluster sites by known metadata, such as vendor or reported stage mix? When metadata is accurate, this is a strong oracle, and we include it as an upper-reference arm. But stage labels at early stages are exactly what the masking confound makes unreliable. A learned signature can also capture compensation that does not match nominal labels. If the metadata oracle matches our method, the method's value reduces to the case where metadata is missing or wrong. We test that case by corrupting a fraction of the stage metadata.
Proposed Experiments
The primary testbed is a semi-synthetic federation built from ADNI resting-state functional MRI. Following DAFed, we split ADNI-3 by scanner vendor (GE, Siemens, Philips). Within each vendor we form two clients, one enriched for early MCI and one enriched for late MCI and Alzheimer's disease, and both keep controls. This yields six clients in which vendor and stage are fully crossed.
The decisive check comes before any accuracy is measured. For each signature, we correlate the learned weight matrix with a same-vendor indicator matrix and with a same-stage indicator matrix. The prediction is that update cosine, task vectors, FedAux, and the unconditioned-mean ablation load on vendor, while the contrast signature loads on stage. If the contrast signature also loads on vendor, the premise fails and we stop.
A controlled simulation checks the mechanism. We inject a known multiplicative scanner gain into all subjects at chosen clients, and a known hub hyperconnectivity only into early-stage patients. The contrast signature should be invariant to the first and sensitive to the second.
Accuracy is then reported as per-site balanced accuracy and AUC for three tasks: early MCI versus control, late MCI versus control, and early versus late MCI. We report both the mean and the worst site. The module-level sign analysis described above supplies the interpretive check.
ABIDE serves as a negative control, since there is no expectation of an inverted-U there. The contrast method should roughly match DAFed's 76.4% rather than lose to it. If it helps on ADNI but hurts on ABIDE, that would show it is specific to stage heterogeneity, and we would report that as its scope rather than hide it.
References
- FedCAP: Robust Federated Learning via Customized Aggregation and Personalization
- Bi-level Personalization for Federated Foundation Models: A Task-vector Aggregation Approach
- Personalized Subgraph Federated Learning with Differentiable Auxiliary Projections
- Research progress on brain network imaging biomarkers of subjective cognitive decline
- Functional Segregation‐Integration Preference Configures the Cognitive Decline Against Cerebral Small Vessel Disease: An <scp>MRI</scp> Study
- Network-based statistics reveals an enhanced subnetwork in prefrontal cortex in mild cognitive impairment: a functional near-infrared spectroscopy study
- A Privacy-Preserving Domain Adversarial Federated learning for multi-site brain functional connectivity analysis
- UG-FedDA: Uncertainty-Guided Federated Domain Adaptation for Multi-Center Alzheimer's Disease Detection
- ClusMFL: A Cluster-Enhanced Framework for Modality-Incomplete Multimodal Federated Learning in Brain Imaging Analysis
- Federated modality-specific encoders and partially personalized fusion decoder for multimodal brain tumor segmentation
- Biologically Plausible Brain Graph Transformer