Cross-site validation of rituximab safety models is limited by small cohorts and privacy barriers. I hypothesize that homomorphically encrypted longitudinal models using time-updated IgG, infection outcomes, and maintenance-rituximab exposure can preserve clinically useful calibration and ranking performance while enabling decentralized validation across autoimmune registries.

Why this is plausible

• The key predictors are structured numeric variables that fit privacy-preserving computation well.
• Safety validation depends on cross-site heterogeneity; single-center autoimmune cohorts are underpowered.
• DeSci infrastructure is strongest when external validation does not require raw patient-level data transfer.

Testable prediction A CKKS-based encrypted pooled analysis of longitudinal rituximab safety data will achieve calibration slope and AUROC within a pre-specified non-inferiority margin versus plaintext pooled analysis for predicting serious infection or severe hypogammaglobulinemia.

Suggested study design

• Multi-center retrospective or prospective autoimmune registry collaboration
• Inputs: baseline IgG, serial IgG, rituximab schedule, co-immunosuppression, serious infection outcomes
• Compare encrypted pooled model vs plaintext pooled model on discrimination, calibration, and net benefit
• Pre-specify acceptable degradation margin before claiming success

Falsification This hypothesis is weakened if encrypted modeling produces clinically important calibration drift or ranking loss relative to plaintext analysis.

References

1. Acar A et al. A survey on homomorphic encryption schemes: Theory and implementation. ACM Comput Surv 2018;51(4):79. DOI: 10.1145/3214303
2. Barmettler S et al. Front Immunol 2021;12:671503. DOI: 10.3389/fimmu.2021.671503
3. Md Yusof MY et al. Rheumatol Int 2021;41(11):1981-1993. DOI: 10.1007/s00296-021-04847-x