Claim

A federated validation framework based on homomorphically encrypted pooled time-to-event statistics can preserve clinically useful discrimination and calibration for steroid psychiatric toxicity models while avoiding transfer of raw patient-level psychiatric data, steroid timelines, or chart-level outcome details between institutions.

Rationale

Steroid psychiatric toxicity is important but relatively uncommon at single-center scale, so external validation needs multicenter pooling. Yet the relevant data are sensitive. Homomorphic encryption offers a path to aggregate sufficient statistics or prediction summaries without directly exposing underlying records.

Testable design

• Sites fit the same prespecified steroid psychiatric toxicity model using a shared variable dictionary
• Each site encrypts score vectors or survival sufficient statistics under a common homomorphic scheme
• Central aggregation derives pooled hazard ratios, calibration slope, integrated Brier score, and time-dependent C-index
• Compare encrypted pooled validation with conventional trusted-analyst pooling as the reference standard

Falsification criterion

The hypothesis is weakened if encrypted aggregation materially degrades calibration or discrimination relative to conventional pooling, or if the communication and computation costs make routine validation impractical.

Why this matters

If true, this would create a DeSci-ready path for multicenter validation of a clinically important but privacy-sensitive steroid safety model.

References

1. Froelicher D, Muller P, De Mestral C, et al. Truly privacy-preserving federated analytics for precision medicine with multiparty homomorphic encryption. Nat Commun. 2023;14:4540. DOI: 10.1038/s41467-023-40353-1
2. Warrington TP, Bostwick JM. Psychiatric adverse effects of corticosteroids. Mayo Clin Proc. 2006;81(10):1361-1367. DOI: 10.4065/81.10.1361
3. Patten SB, Neutel CI. Corticosteroid-induced adverse psychiatric effects: incidence, diagnosis and management. Drug Saf. 2000;22(2):111-122. DOI: 10.2165/00002018-200022020-00004