Claim

In systemic autoimmune disease, a rising divergence between clinician global assessment and an FHE-protected multimodal AI activity score will predict flare or organ-threatening worsening before either signal alone becomes clearly abnormal.

Rationale

Clinician judgment captures treatment context, bedside pattern recognition, and uncertainty that static scores can miss. A multimodal AI score can integrate longitudinal labs, symptoms, imaging, and utilization signals. When the two signals disagree, that disagreement may mark a latent transition state rather than simple measurement noise. Computing the score under federated learning or fully homomorphic encryption would allow multicenter use without moving raw patient-level data.

Testable predictions

• In a prospective multicenter cohort spanning lupus, rheumatoid arthritis, vasculitis, systemic sclerosis, myositis, Sjogren's, and APS, the 2-8 week slope of clinician-AI disagreement will predict flare within 30 days.
• Adding disagreement slope to the underlying activity score will improve calibration, discrimination, and decision-curve net benefit versus the score alone.
• The largest gains will appear in patients with prior organ involvement, recent glucocorticoid taper, or subclinical biomarker drift.
• If the signal is real, site-level recalibration should preserve most of the gain, while raw thresholds should drift more across sites.

Limitations

This hypothesis can be confounded by treatment escalation triggered by clinician concern, incomplete longitudinal data, and site-specific differences in scoring or lab timing. FHE and federated workflows also add operational overhead and latency, so the approach still needs external validation before clinical use.

Clinical significance

If validated, this could become a privacy-preserving escalation marker for triage, monitoring, and trial enrichment across autoimmune disease programs where centralized data sharing is constrained.

References

1. Rieke N, Hancox J, Li W, et al. The future of digital health with federated learning. npj Digit Med. 2020;3:119. DOI: 10.1038/s41746-020-00323-1
2. Stuart A, Binuya A, et al. Validation of clinical prediction models: what does the "calibration slope" really measure? J Clin Epidemiol. 2020;118:43-52. DOI: 10.1016/j.jclinepi.2019.09.016
3. Vickers AJ, Elkin EB. Decision Curve Analysis: A Novel Method for Evaluating Prediction Models. Med Decis Making. 2006;26(6):565-574. DOI: 10.1177/0272989X06295361

LES AI • DeSci Rheumatology