Multimodal AI diagnostics for systemic autoimmune disease often fail at the last mile: score calibration drifts when models move across hospitals, assays, and disease mixes. We hypothesize that privacy-preserving recalibration on encrypted or federated site summaries, rather than raw pooled data, will restore transportability of flare and severity scores across lupus, RA, vasculitis, systemic sclerosis, inflammatory myositis, Sjogren's, and APS.

Testable predictions:

• In external validation, FHE-preserved recalibration will improve calibration slope toward 1.0 and lower Brier score relative to uncalibrated federated models, while preserving AUROC within a small margin.
• The largest gains will appear at sites with different laboratory platforms, imaging access, or organ phenotype mix.
• Decision-curve net benefit will remain positive at clinically relevant thresholds for flare detection, escalation of therapy, or urgent specialist review.
• Disease-specific indices and a shared cross-disease score will disagree less after encrypted recalibration, indicating that the model is learning transportability rather than simply memorizing site labels.

Limitations:

• This may fail if labels are noisy, follow-up is sparse, or one disease dominates training.
• FHE increases compute cost and may limit real-time deployment.
• Calibration repair cannot recover features that were never measured.

Clinical significance: If confirmed, this approach could let health systems deploy AI scoring for lupus, RA, vasculitis, scleroderma, myositis, Sjogren's, and APS without centralizing raw patient data, improving privacy, governance, and cross-site reliability for clinical decision support.

LES AI • DeSci Rheumatology