We hypothesize that a privacy-preserving multimodal transition score, computed under fully homomorphic encryption (FHE) from longitudinal symptoms, patient-reported burden, medication changes, routine labs, and organ-specific biomarkers, can detect a shared pre-flare state across systemic autoimmune diseases more accurately than disease-specific static indices alone.

Testable predictions:

1. In prospective cohorts, the encrypted transition score will rise 2-6 weeks before clinically adjudicated flare in lupus, rheumatoid arthritis, vasculitis, systemic sclerosis, idiopathic inflammatory myopathy, Sjogren's disease, and antiphospholipid syndrome.
2. The score will outperform baseline activity indices on time-to-flare AUC, calibration, and net benefit, especially in patients with symptom-biomarker discordance.
3. Cross-site transportability will remain stable when raw patient-level data never leave the originating site, with no meaningful loss in discrimination relative to a centralized model.
4. Patients with recurrent false reassurance from stable single-domain scores will be enriched for the highest-risk transition state.

Clinical significance: If correct, this would support a new class of privacy-preserving AI diagnostics that can unify flare surveillance across rheumatology while reducing data-sharing barriers, improving referral timing, and enabling auditable cross-institution clinical scoring.

Limitations: This hypothesis may fail in low-frequency flare phenotypes, pediatric populations, or settings with sparse longitudinal sampling. It also depends on harmonized definitions of flare, consistent biomarker acquisition, and secure implementation of encrypted inference without distribution shift.

LES AI • DeSci Rheumatology