Hypothesis: Homomorphically encrypted pooled sufficient statistics can externally validate anti-drug antibody risk models across decentralized autoimmune cohorts without sharing raw patient-level data

Claim
A validation network that exchanges homomorphically encrypted pooled sufficient statistics rather than raw patient-level records can externally validate anti-drug antibody risk models across decentralized autoimmune cohorts with minimal loss of discrimination or calibration.
Why this matters
Rare-event immunogenicity models are hard to validate because no single center sees enough events, but cross-site raw data transfer is often blocked by governance, privacy, and cross-border constraints. If encrypted aggregation preserves model performance, it would make DeSci-style multicenter validation practical without giving up data sovereignty.
Mechanistic rationale
- Anti-drug antibody outcomes are sufficiently structured to support pooled summary-statistic modeling.
- Homomorphic encryption can permit computation on encrypted sufficient statistics while preserving patient privacy.
- Federated validation is especially attractive for pharmacogenomic and immunogenicity models where cross-site phenotype harmonization is possible but raw data sharing is not.
Testable prediction
Across a multicenter autoimmune biologics consortium:
- Encrypted pooled validation will recover AUROC and calibration slope within a small tolerance of conventional pooled validation.
- The approach will retain enough precision to evaluate HLA-DQA1*05 and methotrexate terms in the ADA model.
- Turnaround time will remain operationally acceptable for iterative model validation once the encryption workflow is automated.
Proposed study
- Population: multicenter RA, SpA, and psoriatic disease cohorts on biologic therapy
- Data elements: HLA-DQA1*05, methotrexate exposure, trough levels, antibody status, smoking, CRP, disease duration, prior biologic failure
- Analysis: compare encrypted pooled estimation with conventional pooled analysis for AUROC, calibration slope, Brier score, and decision-curve net benefit
- Design: privacy-preserving aggregation with no raw patient-level export
Falsifiability
This hypothesis fails if encrypted aggregation materially worsens calibration, discrimination, or feasibility relative to conventional pooling.
References
- Kaissis GA, Makowski MR, Rückert D, Braren RF. Secure, privacy-preserving and federated machine learning in medical imaging. Nat Mach Intell. 2020;2:305-311. DOI: 10.1038/s42256-020-0186-1
- Froelicher D, Troncoso-Pastoriza JR, Raisaro JL, et al. Truly privacy-preserving federated analytics for precision medicine with multiparty homomorphic encryption. Nat Commun. 2021;12:5910. DOI: 10.1038/s41467-021-25972-y
- Thapa C, Camtepe S, Sun H. Secure multi-party computation and fully homomorphic encryption for privacy-preserving data analytics: a survey. BMC Med Inform Decis Mak. 2022;22:260. DOI: 10.1186/s12911-022-01969-9