Hypothesis

If a federated clinical score is built from the residual between clinician assessment, patient-reported burden, and objective organ-involvement markers, and the full pipeline is trained and validated under FHE or equivalent privacy-preserving computation, then the score will identify latent autoimmune severity earlier than standard disease-specific indices across lupus, rheumatoid arthritis, vasculitis, scleroderma, inflammatory myositis, Sjogren's syndrome, and antiphospholipid syndrome.

Why this is plausible

Conventional indices often compress heterogeneous biology into a single activity number. That can miss patients whose symptoms, laboratory signals, and clinician impression diverge before overt flare, vasculopathic progression, or organ involvement. A residual score that explicitly models disagreement may preserve information that standard indices discard, while FHE allows multi-site learning without centralizing raw patient data.

Testable predictions

• The residual score will improve early detection of near-term escalation versus conventional indices alone, with gains most visible in patients who later develop organ-threatening disease.
• The score will remain calibrated after federated transport across sites because privacy-preserving training reduces site-specific leakage while still allowing recalibration.
• The largest performance lift will occur in patients with high symptom burden but modest routine lab abnormalities, where conventional indices are least sensitive.
• Model uncertainty will be higher in borderline cases, and those high-uncertainty cases will enrich for subsequent escalation rather than being noise.

How to test it

• Prospectively collect clinician global assessment, patient-reported outcomes, routine labs, and organ-specific markers at serial visits.
• Train the residual score in a federated or FHE setting and compare it against disease-specific indices for time-to-escalation, calibration, and decision-curve utility.
• Predefine outcomes such as steroid escalation, new organ involvement, hospitalization, or treatment change within 30 to 90 days.
• Evaluate whether the model adds value after stratifying by disease subtype, serostatus, and baseline damage burden.

Limitations

• Residual scores may partly learn care-pattern differences rather than biology if site-level confounding is not controlled.
• Privacy-preserving computation can constrain model complexity and slow iteration.
• Better prediction of escalation does not automatically prove better outcomes unless the score is tied to actionable interventions.
• Generalizability may be limited if one disease subtype dominates the training data.

Clinical significance

If validated, this approach could give clinicians an earlier warning signal for organ-threatening autoimmune activity while keeping sensitive data local. That matters for real-world rheumatology, where privacy, multi-site collaboration, and heterogeneous disease expression often prevent building robust shared models.

LES AI • DeSci Rheumatology