Organizations: Wuhan University · China Academy of Information and Communications Technology · The University of Sydney · Tsinghua University · The Chinese University of Hong Kong, Shenzhen
World models learn state evolution from trajectories, making access to temporal context a central training requirement. Federated learning can use distributed records, while ownership boundaries within a trajectory restrict the examples each client can construct. Our study benchmarks this cross-time setting through hourly action-conditioned clinical prediction on eight MIMIC-IV disease cohorts, comprising 40.87 million transition memberships. We specify severity-based client ownership, patient-separated construction, local history and future-window rules, and paired rollout evaluation from one to 32 hours. A matrix of ten federated algorithms covers 32 disease--partition configurations under five rounds of ten-percent participation. Three findings emerge from existing results and training logs. First, client ownership and participation jointly restrict long-window coverage: only 7.55%--21.36% of pooled-available 32-step windows have a locally complete anchor visited during training, averaged across diseases. Second, finer severity partitions accompany higher FedAvg error in 15 of 16 paired comparisons, while algorithm gains are small and horizon-dependent: FedProx reduces mean error by 0.56%, with no consistent improvement at 32 steps. Third, algorithm labels conceal distinct update behavior, including inactive extrapolation and orders-of-magnitude differences in update scale. Cached-update performance also varies strongly across trajectory partitions under the same benchmark protocol. These results establish temporal access, participation coverage, optimization behavior, and horizon-resolved prediction as complementary dimensions for evaluating federated clinical world models.
Figures & tables
Figure 1: World models motivate federated learning from distributed trajectories. History and actions condition future-state prediction; learning these dynamics draws on trajectory breadth, diversity, and temporal context. FL connects local learners through model updates while records remain local.
Learning setting
Predictive task
Trajectory access
Study
Federated training
Clinical tasks
Multi-step prediction
Action conditioning
Fragmented ownership
Temporal access audit
Vista [ 21 ]
–
–
✓
✓
–
–
EHRWorld [ 26 ]
–
✓
✓
✓
–
–
EHR-MPC [ 28 ]
–
✓
✓
✓
–
–
LEAF [ 6 ]
✓
–
–
–
–
–
FedScale [ 7 ]
✓
–
–
–
–
–
Table 1: Coverage of representative world models and federated studies. Our study jointly evaluates clinical dynamics, federated training, and temporal access within trajectories.
Figure 2: Cross-time ownership and local training windows. Clients A, B, and C hold different segments of one stay. History retains same-client positions within the lookback, and future supervision stops at an ownership boundary. Sampled clients update the shared model; evaluation follows patient-disjoint trajectories across stage boundaries.
Figure 3: SOFA and OASIS partitions in the shared latent space of our model, using the same 4,000 Test20 anchors. Upper panels show points, centroids, and 1.5-standard-deviation covariance ellipses; arrows connect centroids in severity order. Lower panels show smoothed PC1 distributions normalized to their own peaks. PC1 and PC2 explain 10.5% and 6.4% of variance. K counts severity groups before patient sharding.
Figure 5: Paired scores for the best structured model and the semantic model.
Figure 6: Retention of complete local training windows relative to pooled access. Curves average per-disease retention ratios, using the same eligible anchors within each comparison. H1 is fully retained, while longer horizons increasingly encounter ownership changes.
Figure 7: Temporal continuity changes the severity composition of training examples. Bars compare each group’s share of eligible anchors ( pg ) with its share of complete local H32 windows ( qg,32 ) in five-band partitions. Shares are calculated within each disease before averaging across eight diseases. Groups aggregate ten patient shards; window counts precede participation sampling.
Figure 8: Only 7.55%–21.36% of pooled-available H32 windows have a locally complete anchor visited during training. The stacked categories share one denominator and exhaust the pooled-available windows. Ownership changes and nonparticipation reduce complete-window coverage through separate paths.
Figure 9: Disease-wise changes from three to five severity bands. Left: complete H32-window retention change in percentage points. Right: relative change in the six-horizon FedAvg nMAE. Colors identify SOFA and OASIS.
Figure 10: nMAE ( ↓ ) across eight disease cohorts (columns) and four trajectory partitions (rows). Curves show Central and ten federated methods at H1–H32. Vertical limits are shared within each disease column.
Figure 11: nMSE ( ↓ ) for the same diseases, partitions, methods, and horizons as Figure 10 . Errors are normalized by training feature variance; vertical limits are shared within each disease column.
Figure 12: Consistency across 32 settings. Better/worse denote changes beyond ±1% relative to FedAvg; similar denotes changes within this range.
Figure 13: Horizon profiles averaged equally over eight diseases and four configurations. The left panel shows mean paired relative changes for methods close to FedAvg (negative is better); the right shows absolute macro nMAE for the remaining methods and FedAvg. Both use all 32 settings and six horizons.
Figure 14: Central and FedAvg across all eight diseases and both score families. Points average nMAE over six horizons. The reference and federated configurations follow Section 4.1 .
Figure 15: Configured update rules produce different relative scales across rounds. Each point is the median of 32 within-configuration method/FedAvg update-norm ratios. Both panels use the same logarithmic scale and include FedAvg at one.
Figure 16: FedExP activation under the benchmark protocol. Left: rounds at the factor-one floor versus active extrapolation. Right: all 40 round factors per partition; horizontal offsets separate observations for display.
Figure 17: Partition-dependent behavior of cached updates. FedVARP and FedAvg are compared within each disease under the same protocol. Lines connect six-horizon mean nMAE values across disease cohorts; solid and dashed lines denote three and five severity bands.
Figure 18: Evidence chain for the fixed benchmark: temporal ownership determines visible context, participation determines its coverage, update behavior characterizes algorithm responses, and horizon-specific prediction measures outcomes.
Recent reviews find that the vast majority of published healthcare federated learning (FL) studies never reach real-world deployment. We developed an embedding-based FL pipeline for iron deficiency prediction from routine full blood count (FBC) data and deployed it across real institutional environments at Amsterdam University Medical Centre (AUMC) and NHS Blood and Transplant (NHSBT), two clinical environments that differ markedly in iron deficiency prevalence, ferritin distribution, and subject populations. A frozen domain-specific haematology foundation model, DeepCBC, performs site-local representation extraction, restricting federated training to a compact downstream classifier and substantially reducing recurrent communication relative to full-encoder federation. The two clinical datasets are structurally not independent and identically distributed (non-IID), with heterogeneity arising from distinct population differences rather than sampling artefacts. Runtime governance is enforced by FLA3, a healthcare-oriented FL platform providing study-scoped execution, policy-based authorisation, and signed audit logging. Standard sample-size-weighted aggregation (FedAvg) reduced the area under the receiver operating characteristic curve (ROC-AUC) at both sites relative to local-only training, as the global update was biased towards the larger AUMC distribution. FedMAP, a personalised aggregation method, raised ROC-AUC from 0.9470 to 0.9594 at AUMC and from 0.8558 to 0.8671 at NHSBT relative to local-only training, achieving the highest macro ROC-AUC of 0.9133 and the best macro balanced accuracy overall. These results support personalised aggregation in clinical federations where client sample size and task relevance diverge substantially.
Fan Zhang, Simon Deltadahl, Majid Lotfian Delouee +10
Dept. of Applied Mathematics and Theoretical Physics University of Cambridge Cambridge, UK · Translational AI Laboratory, Dept. of Laboratory Medicine Amsterdam UMC Amsterdam, The Netherlands · Precision Health University Research Institute Queen Mary Univ. of London London, UK +3
The federated learning (FL) paradigm fosters distributed pervasive computing combined with artificial intelligence techniques, allowing for optimized data usage and improved mitigation of privacy concerns. Indeed, model training occurs on the client's local devices, and model parameters are subsequently shared with a centralized server. However, there is a need to find a tradeoff between models' personalization and generalization capabilities. In this paper, we design and implement several testing scenarios devoted to evaluating and comparing the centralized, local, and federated paradigm performances. We also design and implement a scenario that emulates a change in clients' data. We then present experimental results of the FedAvg algorithm applied to the Human Activity Recognition (HAR) domain to understand the trade-off between personalized and generalized accuracy. Results show that, although FedAvg confirms a higher degree of personalization capabilities while keeping a high degree of generalization with respect to the traditional centralized learning, this result is not so obvious under stressful conditions, such as when varying class distribution over clients.
Andrea De Luna, Susanna Peretti, Chiara Contoli +1
Federated learning usually aggregates client updates using dataset size or gradient-level criteria, while overlooking internal signals about how each client model is organizing its input space during training. We introduce FedOUI, a simple aggregation rule based on the Overfitting-Underfitting Indicator (OUI), an activation-based and label-free metric. Each participating client sends its local update together with a OUI value computed on a fixed probe batch, and the server estimates the round-wise OUI distribution to assign lower weights to structurally atypical clients through a smooth reweighting rule. We evaluate FedOUI on CIFAR-10 under strong non-IID partitioning and noisy-client conditions, comparing it with FedAvg, FedProx, and a gradient-alignment baseline. The clearest gains appear under strong heterogeneity, where OUI-based weighting improves aggregation quality while remaining lightweight and interpretable. These results show that internal activation structure can provide useful information for federated aggregation beyond client size and gradient geometry.
Alberto Fernández-Hernández, Jose I. Mestre, Cristian Pérez-Corral +3
Universitat Politècnica de València, Valencia, Spain · Universitat Jaume I, Castelló de la Plana, Spain · Openchip & Software Technologies S.L., Spain