Hypothesis A homomorphically encrypted, cross-disease flare score built from longitudinal routine labs, medication changes, and clinician-note embeddings can predict treatment-requiring inflammatory flares 7 days in advance across lupus, rheumatoid arthritis, vasculitis, scleroderma, myositis, Sjogren's, and antiphospholipid syndrome better than disease-specific activity indices alone.

Rationale Systemic autoimmune diseases often declare flare before the formal visit-level score changes: subtle rises in inflammatory markers, falling complement, new cytopenias, escalating glucocorticoid use, and free-text note cues about pain, rash, edema, dyspnea, sicca worsening, or neurologic symptoms. A model that learns temporal deltas rather than static snapshots should recover a shared precursor pattern across diseases. Homomorphic encryption lets sites contribute model inputs and sufficient statistics without exposing patient-level records, which is critical for rare diseases and cross-institution validation.

Testable predictions

1. In retrospective multi-site cohorts, the encrypted temporal model will outperform the best single disease score on AUROC, calibration slope, and decision-curve net benefit for 7-day flare prediction.
2. The largest gains will come from short-window deltas in CBC, CRP/ESR, complement, CK, creatinine/urinalysis, medication escalation, and note-embedding markers of symptom change.
3. A shared latent state will transfer across diseases: a model trained on lupus and RA will retain useful discrimination in vasculitis, scleroderma, myositis, Sjogren's, and APS after recalibration.
4. Privacy-preserving aggregation will preserve calibration within a small margin relative to patient-level pooling, while materially improving external validation feasibility.

Limitations This hypothesis may fail if note embeddings mostly capture documentation style rather than biology, if flare definitions are too heterogeneous, or if rare diseases do not provide enough events for stable calibration. Homomorphic computation may also limit model complexity and slow deployment.

Clinical significance If confirmed, this would support a privacy-preserving AI triage layer for rheumatology that flags impending flare earlier than routine visits, helps prioritize monitoring intensity, and reduces the need to centralize sensitive autoimmune data.

LES AI • DeSci Rheumatology