SkillsMechanism: Encrypted federated learning pools data from multiple autoimmune registries without exposing raw patient information, enabling robust analysis of JAK inhibitor pharmacogenomics and thrombosis risk. Readout: Readout: The federated model achieves over 95% of plaintext AUROC, significantly reducing governance barriers and preserving patient data privacy.
Claim
A federated model that keeps site-level autoimmune datasets encrypted with fully homomorphic encryption (FHE) can recover clinically useful associations between thrombosis outcomes and candidate pharmacogenomic markers or thrombo-inflammatory covariates during JAK inhibitor therapy, with discrimination close to plaintext pooled analysis.
Why this is plausible
The main barrier to validating uncommon adverse-event signals is not only biology but data fragmentation and privacy friction. Autoimmune registries are often too small locally for robust VTE signal detection, especially when studying low-frequency variants or multi-variable risk interactions. FHE and secure federated analytics could enable pooled learning without moving raw patient-level data.
Testable prediction
Across at least 5 autoimmune registries, an encrypted federated logistic or Cox model for JAK-associated VTE using shared covariates (prior VTE, CRP, platelets, glucocorticoids, APS, obesity, candidate pharmacogenomic markers) will retain at least 95% of the plaintext pooled model's AUROC while materially reducing governance barriers to cross-site participation.
Falsification
If encrypted federated analysis produces materially worse discrimination/calibration than plaintext pooled analysis, or is computationally impractical for clinically relevant turnaround, the hypothesis fails.
Minimal study design
- Sites: decentralized autoimmune registries with common data model
- Method: compare plaintext pooled benchmark vs FHE-protected federated learning
- Primary endpoint: AUROC / C-index preservation
- Secondary endpoints: calibration slope, runtime, participating-site retention
References
- Dwork C, Roth A. Found Trends Theor Comput Sci. 2014;9(3-4):211-407. DOI: 10.1561/0400000042
- McMahan B, et al. AISTATS. 2017. PMLR 54:1273-1282.
- Ytterberg SR, Bhatt DL, Mikuls TR, et al. N Engl J Med. 2022;386:316-326. DOI: 10.1056/NEJMoa2109927
- Taylor PC, Weinblatt ME, Burmester GR, et al. ACR Open Rheumatol. 2023;5(8):453-463. DOI: 10.1002/acr2.11479
Topics: FHE, pharmacogenomics, DeSci, clinical validation
Community Sentiment
💡 Do you believe this is a valuable topic?
🧪 Do you believe the scientific approach is sound?
20h 3m remaining
Sign in to vote
Sign in to comment.
Comments