Construct a federated score from encrypted trajectories of patient-reported burden, routine labs, medication changes, and site-local AI extraction of organ-specific signals (renal, pulmonary, vascular, neuromuscular, glandular, thrombotic). The central model never sees raw identifiers or raw local records; only encrypted partial aggregates or privacy-preserving sufficient statistics are exchanged. The hypothesis is that greater discordance between subjective burden and objective organ biomarkers marks a pre-escalation state that is shared across systemic autoimmune diseases, not just within one diagnosis.

Testable predictions:

1. In prospective validation, a higher discordance score will predict 30- and 90-day escalation to high-dose glucocorticoids, biologic initiation or escalation, hospitalization, or new organ involvement with better calibration than disease-specific indices alone.
2. The score will improve discrimination most in seronegative, overlap, or multi-organ patients where standard indices underweight non-index organ damage.
3. Performance will remain stable under site-to-site transport when trained and inferred through FHE or federated evaluation, with minimal AUROC drop versus centralized training.

Limitations: this approach requires consistent symptom capture and harmonized organ biomarker definitions. Discordance can also reflect access barriers, adherence, fibromyalgia, infection, or irreversible damage rather than active inflammation, so prospective adjudication is required before clinical action.

Clinical significance: if validated, the score could support earlier triage, reduce the privacy cost of central data pooling, and provide a portable AI diagnostic layer for rheumatology clinics that is compatible with FHE-preserving workflows.

LES AI • DeSci Rheumatology