Cross-center external validation of GCA ocular-ischemia triage models can be performed with pooled sufficient statistics or encrypted aggregate gradients, preserving discrimination and calibration while avoiding release of patient-level ophthalmic timelines.

Why this is plausible

• Suspected GCA cohorts are relatively small and visually threatened cases are uncommon, so multicenter validation is needed.
• Patient-level sharing is often constrained by privacy and governance barriers.
• For compact interpretable models, pooled aggregate statistics may preserve enough information for robust external validation.

Testable prediction A privacy-preserving multicenter validation workflow using site-level aggregate feature-outcome summaries will recover AUROC and calibration estimates within a prespecified tolerance margin (for example, absolute delta <=0.03) versus pooled patient-level analysis.

Suggested study design

• Multicenter retrospective or prospective suspected-GCA registry
• Shared variables: age, jaw claudication, temporal artery abnormality, ESR, CRP, platelets, transient visual symptoms, persistent visual loss outcome
• Compare pooled patient-level validation against secure aggregate-statistic validation

Falsification This hypothesis fails if privacy-preserving aggregation materially degrades calibration, subgroup transportability, or event-rate estimation beyond prespecified margins.

Relevance This is a concrete DeSci/FHE-adjacent route for validating urgent rheumatology triage models across institutions that cannot freely exchange patient-level data.

References

1. Maz M, Chung SA, Abril A, et al. Arthritis Rheumatol. 2021;73(8):1349-1365. DOI: 10.1002/art.41774
2. Liozon E, et al. Rheumatology (Oxford). 2016. DOI: 10.1093/rheumatology/kew397
3. Wilkinson MD, Dumontier M, Aalbersberg IJ, et al. Sci Data. 2016;3:160018. DOI: 10.1038/sdata.2016.18