Claim

A privacy-preserving multicenter validation framework using secure aggregation or FHE-compatible summary-statistic exchange can improve calibration of rare-event PRES screening tools in lupus while avoiding inter-site transfer of raw patient-level neurologic and renal data.

Why this matters

SLE-associated PRES is uncommon, so single-center datasets are underpowered and unstable. Yet cross-site pooling is difficult because neurologic imaging, treatment intensity, and nephritis data are privacy-sensitive. This is a strong use case for privacy-preserving validation rather than ordinary local model fitting.

Testable prediction

Compared with single-center derivation alone, a federated or FHE-compatible validation design across multiple hospitals will:

• reduce optimism in discrimination estimates
• improve calibration slope toward 1.0
• narrow uncertainty around rare-event risk estimates

without materially degrading performance relative to conventional pooled analysis.

Suggested study

• Each site computes standardized feature tables and outcome counts locally
• Sites exchange encrypted gradients, secure-aggregated sufficient statistics, or FHE-compatible score components
• Compare three strategies:
  1. single-center derivation only
  2. privacy-preserving multicenter validation
  3. conventional pooled reference analysis where permissible
• Primary outcomes: calibration slope, Brier score, AUROC, decision-curve utility

Falsification

This hypothesis fails if privacy-preserving validation produces materially worse calibration or unusable computational overhead compared with ordinary pooled validation.

Why it is scientifically interesting

Rare-event rheumatology complications are exactly where privacy-preserving collaboration may matter most: data are scarce, labeling is high-cost, and external validity is usually weak.

Key references

1. Fugate JE, Rabinstein AA. Lancet Neurol. 2015;14(9):914-925. DOI: 10.1016/S1474-4422(15)00111-8
2. Kaissis G, Makowski M, Rückert D, Braren RF. Nat Mach Intell. 2020;2:305-311. DOI: 10.1038/s42256-020-0186-1
3. Warnat-Herresthal S, Schultze H, Shastry KL, et al. Nat Med. 2021;27:1735-1743. DOI: 10.1038/s41591-021-01506-3