Coverage-Aware Federated Learning of Disease Progression Curves

Hospitals that track the same biomarker over time usually see different parts of the disease. A primary-care network sees patients whose plasma p-Tau217 is just starting to rise. A tertiary memory clinic sees patients whose hippocampus is already shrinking fast. If the hospitals jointly train one model of how the biomarker changes, without sharing records (federated learning), each hospital's local update reshapes the model across the whole biomarker range. That includes stages the hospital never observes, where it has no data to keep the model accurate. We propose to treat this "stage-window skew" as the continuous counterpart of the missing-class problem in federated classification, and to correct it with weights that track which part of the disease range each hospital actually covers.

The core idea fits in two sentences. Represent the shared progression model as a sum of local bumps along the biomarker axis, so that each bump answers the question "how fast does the biomarker move when it is near this value, for this kind of patient?" Then let each hospital train and vote only on the bumps its patients actually pass through. Hospitals report noisy counts of how many visits fall under each bump, and these counts set both their local loss weights and their weight when the server combines updates. The approach targets a regime the literature leaves open: multi-site longitudinal tracking with nonlinear covariate effects, where pooling is forbidden and the sites' observation windows barely overlap.

Where the gap is

Distribution-aware loss weighting is a mature tool for federated classification. FedSLD has clients share class counts and reweight each local batch toward a global label prior (FedSLD). FedIR does the same with importance ratios (Federated Visual Classification with Real-World Data Distribution). FedDisco replaces sample-count aggregation with a weight that penalizes a client's scalar label-distribution discrepancy (FedDisco). The most relevant to us is WSM, which diagnoses client drift as forgetting. Its argument is that the softmax denominator pushes down classes a client never sees, and masking those classes removes the damage, raising CIFAR-100 accuracy under strong skew from 49.4% to 64.0% (Re-Weighted Softmax Cross-Entropy). WSM therefore supplies our central mechanism: do not let a client push on parts of the output it has no evidence about. Every one of these methods, however, is defined over a discrete class histogram, and none says what "absent class" means for a continuous trajectory.

Longitudinal tracking supplies the motivation. Several studies show that trajectories carry prognostic signal that snapshots miss:

  • Albumin change predicts post-infarction mortality better than admission albumin (albumin changes).
  • hsCRP paths predict metabolic syndrome better than baseline values (hsCRP trajectories).
  • In a spline-based joint model, the roughness of blood-pressure trajectories carries a hazard ratio of 1.14 per SD beyond mean level (Weighted biomarker variability).

These methods all assume centrally pooled data. That includes the trajectory generators, such as multi-marginal flow matching on ADNI (Longitudinal Flow Matching), and the globally aligned enrichment model, whose "global/local" split is a centralized ADMM solve rather than a federated protocol (Globally Aligned Longitudinal Enrichments).

The closest bridging work is federated survival analysis.

  • FairFSA turns censored outcomes into pseudo-value regression targets and aggregates by validation concordance (Fair Federated Survival Analysis). It uses static entry covariates only.
  • FedSurf federates survival models across six Indian dialysis zones. It reports that DeepSurv degrades under regional heterogeneity, but it also collapses an inherently longitudinal treatment into 17 static variables (FedSurf).
  • Dirichlet survival splitting shows that label skew over time bins, unlike quantity skew, produces genuinely different client survival curves (Heterogeneous Datasets for Federated Survival Analysis Simulation). We reuse its splitting method, applied to biomarker state rather than event time.

None of these federates a trajectory model, and none weights anything by where along the disease course a client's data lies.

Core Claim

When federated sites cover different windows of a biomarker's range, ordinary averaging corrupts the shared progression model in each site's uncovered region. We call this range forgetting. The fix is to make both the loss and the aggregation coverage-aware at the level of local regions of biomarker state, rather than through a single scalar per client. We expect the main gains in exactly those regions, and in the downstream conversion prediction that depends on them. We expect little gain where coverage overlaps.

Research Program

We model progression as a velocity field: the expected rate of change of biomarker $m$ given its current value and patient covariates $x$ (age, APOE status, site-agnostic demographics). This "time-free" form matters because sites enter patients at different, unknown disease times. A velocity indexed by state, rather than by calendar time, lets fragments from different sites stitch into one curve. The model is

$$\frac{dm}{dt} = \sum_{j=1}^{J} B_j(m)\, g_j(x;\theta,\phi_j),$$

where:

  • $B_j$ are $J$ fixed B-spline basis functions with local support along the biomarker axis;
  • $g_j$ is a small neural head with parameters $\phi_j$ on top of a shared covariate encoder with parameters $\theta$.

Training pairs are consecutive visits; the model integrates the ODE between visit times. Each patient's visit pairs share one unit of total weight, so frequently seen (often sicker) patients do not dominate. This is a partial answer to informative visit frequency, and we will report its effect separately.

Each client $k$ computes $c_{kj}$, the number of its visit-pairs whose states fall in the support of $B_j$. It releases these counts with Laplace noise calibrated to a fixed privacy budget. This is the continuous analogue of FedSLD's class counts, and it inherits the same re-identification concern for sparse bins. We will therefore report results at several noise levels.

The counts are used in three places:

  1. Local loss (FedSLD-style). Visit-pairs are reweighted by the ratio of global to local coverage of their region.
  2. Function-space anchor (WSM-style). At covariate values drawn from the client's own patients, paired with states outside the client's coverage, the local model is penalized for moving its predicted velocity away from the current global model. This plays the role WSM's mask plays for absent classes: the client keeps its hands off regions it cannot see. Because the anchor needs no extra data, it adds no communication.
  3. Server aggregation, per basis rather than per client. The shared encoder is aggregated with FedDisco-style weights, using a coverage-histogram discrepancy as the scalar $d_k$. The basis heads are combined region by region:

$$\phi_j^{\text{global}} = \frac{\sum_k c_{kj}\,\phi_{kj}}{\sum_k c_{kj}}.$$

Here $\phi_{kj}$ is client $k$'s updated head for basis $j$, so a hospital's vote on a stage is proportional to how much of that stage it has seen.

A survival head then attaches to the learned trajectory. It is a discrete-time hazard on current predicted level, velocity, and accumulated roughness $\int (m'')^2$, following the joint-model covariates of Weighted biomarker variability. Its loss is separable across patients, which sidesteps the joint EM that does not federate. This addresses the tension between FairFSA's need for separable losses and the joint model's coupling. The cost is that the trajectory is not jointly re-estimated from event information, and we will measure how much that costs against a centralized joint fit.

Baselines

The naive reference points are local-only training at each site and FedAvg weighted by sample count on the identical architecture. FedAvg is the method whose range forgetting we claim exists. We beat it only if error in each client's uncovered region drops by more than seed-to-seed variability while covered-region error does not rise.

The upper reference is the same model trained on pooled data. We report the fraction of the FedAvg-to-pooled gap that is closed.

The strongest federated competitor on this question is FedDisco aggregation plus FedProx's proximal term, given the same coverage-histogram discrepancy we use, so it receives exactly the same metadata (FedDisco). It is the strongest because it already down-weights skewed clients and adds generic drift control. What it lacks is region resolution. On CIFAR-10, FedDisco reaches 55% accuracy in 65.52% fewer communication rounds than FedAvg; no reported result found on Lacuna for any federated method on longitudinal trajectory forecasting.

For downstream discrimination, the reference is the static snapshot. Baseline p-Tau217 alone reaches an AUROC of 0.787 for 4-year MCI-to-AD conversion in 201 ADNI participants (Predicting Progression from MCI to AD). Beating it means a federated trajectory model exceeds that AUROC on a matched ADNI subset. Because our cohort definition will differ, we refit the static model on our exact split and compare against that refit, not against the published figure.

All federated methods share the same architecture, rounds, local steps, client sampling, and a 20-configuration hyperparameter search on held-out validation patients.

Ablations isolate the key ingredient:

  • per-basis aggregation replaced by the scalar client weight;
  • anchor removed;
  • coverage counts replaced by uniform counts;
  • counts at increasing noise levels.

Why not just…?

Why not just pool sufficient statistics for a spline mixed model? If the velocity field were linear in the spline coefficients with no covariate nonlinearity, clients could send their summed design cross-products. The server would then recover the pooled solution exactly in one round, and coverage weighting would be unnecessary. This is a real threat, so we include it as a baseline. Our claim survives only where it loses: when progression rate depends nonlinearly on covariates, or when the biomarker is high-dimensional (FreeSurfer regional volumes), so that the cross-product matrices become large and identifying. If the sufficient-statistic spline matches our method on ADNI, we will report that the method is unnecessary for low-dimensional biomarkers and restrict the claim accordingly.

Why not just use FedAvg with FedProx or FedDisco unchanged? Their weighting is one number per client, so a site that covers half the disease range either moves everything or moves nothing. The per-basis ablation tests this directly. The prior evidence from classification (WSM's 14.6-point gain over FedAvg on CIFAR-100, and its additive gain on top of SCAFFOLD) suggests that drift control without output masking leaves forgetting in place. Whether the same holds for regression curves is exactly what the uncovered-region metric measures.

Why not just extract per-patient slopes locally and federate a static survival model like FairFSA? It is simple and fully federable. However, two-stage estimation of slopes shows bias and poor coverage relative to joint modeling in the simulations of Weighted biomarker variability. It also never produces a shared curve that a site can use to forecast stages it has not seen. We run it as the two-stage baseline with FairFSA-style concordance-weighted aggregation, and compare conversion AUROC, concordance, and integrated Brier score.

Why not just personalize, fine-tuning the global model at each site? Personalization helps each site on its own window, but it cannot fix the shared model in regions no single site covers well. That shared model is the thing a primary-care site needs when its patients later progress. We report both global and fine-tuned variants, and evaluate on future visits that leave the site's training range.

Proposed Experiments

The first testbed is synthetic. We generate trajectories from a known nonlinear velocity field with covariate interactions and informative visit frequency. We partition patients into sites by Dirichlet allocation over biomarker-state bins, adapting the label-skew splitter of Heterogeneous Datasets for Federated Survival Analysis Simulation, and sweep concentration from near-uniform to near-disjoint. Ground truth lets us measure velocity-field error per region and confirm that range forgetting exists under FedAvg before any fix is credited.

The primary real cohort is ADNI. We build stage-window sites: clinics enriched for cognitively normal and early-MCI participants, clinics enriched for late MCI and AD, and a few mixed sites. Biomarkers are hippocampal volume, ventricular volume, and MMSE. Metrics are:

  • forecast RMSE at held-out future visits, stratified by whether the target state lies inside the training site's coverage;
  • worst-site RMSE;
  • 4-year conversion AUROC and concordance from the survival head.

The MRC elderly trial's systolic blood pressure series is a secondary cohort if access permits (Weighted biomarker variability). There we additionally check whether the federated roughness covariate recovers a hazard ratio close to that of the centralized joint model.

The decisive check is the uncovered-region error under near-disjoint partitions. If FedDisco+FedProx with identical metadata comes within seed noise of our method there, or if the one-shot sufficient-statistic spline matches it on ADNI, the central claim is wrong. If our method closes most of the gap to pooled training in uncovered regions while matching baselines in covered ones, and the gain shrinks as partitions become IID, then stage-window skew is a real and fixable failure mode of federated longitudinal tracking.