Hypothesis: Privacy-preserving pooled sufficient statistics can externally validate denosumab hypocalcemia models across decentralized CKD and autoimmune cohorts without sharing raw laboratory data

Claim External validation of denosumab-associated hypocalcemia risk models can be achieved across decentralized autoimmune and CKD cohorts using privacy-preserving pooled sufficient statistics or FHE-compatible summary exchange, with minimal loss of calibration compared with centralized raw-data pooling.

Why this matters Severe denosumab hypocalcemia is uncommon but clinically important, which means single-center cohorts are often underpowered. Multi-site validation is needed, yet calcium, PTH, vitamin D, dialysis, and medication records are sensitive. A decentralized validation strategy could improve evidence quality without moving patient-level data.

Testable design

• Sites keep raw data locally.
• Each site computes agreed sufficient statistics or encrypted gradient updates for a transparent prespecified model using variables such as CKD stage, dialysis, baseline calcium, PTH, vitamin D status, supplementation, loop diuretics, and symptoms.
• A coordinating node aggregates only privacy-preserving outputs to estimate coefficients or to validate a fixed heuristic/biostatistical model.
• Compare decentralized versus centralized validation on AUROC, calibration intercept/slope, observed-to-expected ratio, and subgroup stability in dialysis vs non-dialysis patients.

Falsification The hypothesis fails if decentralized aggregation materially degrades calibration or subgroup performance beyond a prespecified margin versus centralized analysis.

Why it is plausible Many clinically useful validation tasks depend on low-dimensional sufficient statistics rather than unrestricted sharing of row-level data. This is especially attractive for DeSci consortia trying to validate rare but serious drug-safety events ethically.

Limitations Heterogeneous lab assays, missingness patterns, and incompatible supplementation protocols may dominate performance loss rather than the privacy method itself.

References

1. Hardy M, Heneghan C, Mahtani KR, et al. BMC Med Res Methodol. 2022;22:315. DOI: 10.1186/s12874-022-01757-3
2. Brisimi TS, Chen R, Mela T, Olshevsky A, Paschalidis IC, Shi W. IEEE J Biomed Health Inform. 2018;22(3):578-585. DOI: 10.1109/JBHI.2017.2706378
3. Kairouz P, McMahan HB, Avent B, et al. Found Trends Mach Learn. 2021;14(1-2):1-210. DOI: 10.1561/2200000083