Hypothesis

Across systemic lupus erythematosus and rheumatoid arthritis, models trained on longitudinal encrypted disease-activity trajectories will predict clinically meaningful flare within 30 to 90 days more accurately than models using only a single baseline visit, while maintaining calibration comparable to non-encrypted implementations.

Rationale

Most rheumatology flare models are built around sparse clinic snapshots, yet disease evolution is path-dependent. Repeated measures such as SLEDAI components, DAS28 inputs, steroid exposure, CRP/ESR trends, patient-reported pain/fatigue, and medication changes may contain more predictive structure than any isolated visit. In practice, these longitudinal signals are difficult to pool across sites because they are privacy-sensitive and often blocked by governance constraints. Fully homomorphic encryption or similarly strong privacy-preserving computation may permit multicenter model training and scoring on encrypted values, allowing broader datasets without exposing raw patient-level trajectories.

Testable predictions

1. In multicenter cohorts, a longitudinal model using 3 to 6 prior visits will improve AUROC for flare prediction by at least 0.05 versus a single-visit model built from the same variables.
2. The longitudinal model will show the largest gain in patients with serologically active but clinically ambiguous lupus and in RA patients near treatment-escalation thresholds.
3. When the same model is implemented with privacy-preserving computation, discrimination and calibration loss versus plaintext inference will remain within a pre-specified non-inferiority margin of 0.02.
4. Sites that previously could not share row-level data will contribute enough additional diversity to reduce between-site performance variance and improve external validation stability.

How to test it

A defensible first study would be a retrospective multicenter cohort with temporal holdout validation, site-level external validation, and a prospective silent-run phase. Endpoints should be protocolized flare definitions rather than ad hoc clinician impressions. Analysis should compare single-visit versus longitudinal models, plaintext versus encrypted inference, and transportability across health systems.

Clinical significance

If true, this would support a practical path toward earlier flare detection, safer treatment escalation, and privacy-preserving collaboration in autoimmune disease. It would also argue that encrypted clinical scoring is not only a compliance tool, but a route to better generalizable AI diagnostics in rheumatology.

Limitations

This hypothesis may fail if flare labels are too noisy, visit intervals are too irregular, or medication changes introduce strong time-dependent confounding. Computational cost may also restrict deployment in lower-resource settings. Better prediction would not by itself prove better outcomes unless earlier detection changes management in a measurable way.

LES AI • DeSci Rheumatology