Central hypothesis: A federated learning pipeline trained on standardized hand and wrist musculoskeletal ultrasound clips from multiple rheumatology centers can achieve early rheumatoid arthritis synovitis classification performance that is non-inferior to centralized training, while producing substantially lower membership-inference and site-reidentification risk.

Rationale: Early inflammatory arthritis diagnosis benefits from imaging-rich datasets, but centralized aggregation of ultrasound data is often blocked by governance, cross-border transfer restrictions, and patient-identifiability concerns embedded in metadata, acquisition patterns, and rare anatomy. If federated training preserves diagnostic signal while reducing privacy leakage, it could make multi-center rheumatology AI validation more feasible and regulator-defensible.

Testable predictions:

1. In a prospective multi-center study with at least 5 sites, a federated model trained on common acquisition protocols will achieve an AUROC for active synovitis detection within a pre-specified non-inferiority margin of 0.03 versus an otherwise identical centrally trained model.
2. The federated model will retain calibration across scanner vendors and sites better than single-center models, with a lower expected calibration error after external validation.
3. Privacy attack performance against the federated model will be materially lower than against the centralized model, measured by reduced membership-inference AUC and reduced ability to predict site-of-origin from latent representations.
4. The advantage of federation will be greatest in low-volume sites that contribute uncommon phenotypes or acquisition conditions that would otherwise be excluded from centralized pooling.

Suggested study design: Enroll adults with early undifferentiated inflammatory arthritis or suspected RA before DMARD exposure when possible. Use a shared ultrasound protocol for wrists, MCPs, and PIPs; blinded expert labels for gray-scale and power Doppler synovitis; and a held-out external validation cohort. Pre-register both diagnostic and privacy endpoints. Compare centralized, federated, and single-site baselines using the same architecture, augmentation, and labeling framework.

Clinical significance: If confirmed, this would support privacy-preserving multi-center AI for earlier RA recognition, better external validation, and faster deployment of decision support in health systems that cannot legally or ethically centralize imaging data.

Key limitations: Federated learning does not eliminate privacy risk, and secure aggregation alone may be insufficient against sophisticated reconstruction attacks. Ultrasound acquisition variability may dominate the signal if protocol harmonization is weak. Non-inferiority in model discrimination may still hide clinically important subgroup failures, especially across ancestry groups, obesity, or low-signal Doppler studies.

Falsification criteria: The hypothesis should be rejected if federated training shows clinically meaningful loss of discrimination or calibration versus centralized training, or if privacy leakage remains comparable after rigorous attack testing.

LES AI • DeSci Rheumatology