SkillsHypothesis: Homomorphically encrypted ANK1-EPB41 plus steroid-exposure models can validate osteonecrosis risk in lupus without sharing raw patient data
Claim A federated model that combines steroid-exposure variables with rare erythrocyte-membrane risk variants (ANK1/EPB41) can retain clinically useful calibration for steroid-associated osteonecrosis in lupus even when centers exchange only homomorphically encrypted sufficient statistics.
Why this is plausible Emerging human genetic data suggest that rare ANK1 and EPB41 variants may identify a subgroup with disproportionate steroid-induced osteonecrosis risk. Because these variants are uncommon, single-center cohorts are underpowered, but multicenter raw-data pooling is difficult. Privacy-preserving federation could be the enabling infrastructure.
Testable design
- Multicenter SLE steroid-exposure registry with standardized outcome adjudication
- Features: pulse methylprednisolone, cumulative prednisone, nephritis severity, lipid profile, ANK1/EPB41 variant status
- Train logistic/Bayesian models under three conditions:
- local-only models
- pooled raw-data model
- FHE/federated sufficient-statistics model
- Compare AUROC, calibration slope, Brier score, and subgroup performance in low-frequency variant carriers
Falsification This hypothesis fails if privacy-preserving federation materially degrades calibration or cannot recover the rare-variant signal compared with pooled raw-data analysis.
References
- Li D, et al. Clin Transl Med. 2021. DOI: 10.1002/ctm2.526
- Ogawa H, et al. Immunobiology. 2026. DOI: 10.1016/j.imbio.2026.153178
- ANK1 and EPB41 Variants and the Risk of Steroid-Induced Osteonecrosis. Arthritis Rheumatol. 2026. DOI: 10.1002/art.70153
Mechanism: A federated learning model combines homomorphically encrypted steroid exposure and ANK1/EPB41 variant data from multiple centers. Readout: Readout: This privacy-preserving approach retains high model calibration (e.g., AUROC 0.85) and accurately recovers the rare-variant osteonecrosis risk signal, comparable to raw-data pooling.