Claim

In systemic autoimmune disease, a privacy-preserving AI triage score that combines model uncertainty, cross-modal calibration error, and clinician-facing severity scoring will detect imminent organ-threatening disease earlier than conventional disease activity indices alone, especially in patients with lupus, rheumatoid arthritis, vasculitis, scleroderma, myositis, Sjogren's, and antiphospholipid syndrome.

Rationale

Conventional indices often lag behind clinically important transitions because they compress heterogeneous phenotypes into coarse activity scores. A multimodal AI system that tracks calibration error against structured clinical scoring, laboratory trends, and targeted narrative features may surface discordance when the model is least confident but the patient is still being categorized as low risk. Preserving these analytics with federated learning or homomorphic encryption can allow cross-site validation without centralizing sensitive patient data.

Testable predictions

• Higher short-term calibration error will precede organ-threatening escalation within 30-90 days more often than absolute disease activity score thresholds.
• The signal will be strongest in phenotypes with mixed inflammatory and vascular involvement, including lupus nephritis, ANCA-associated vasculitis, systemic sclerosis renal crisis risk, inflammatory myopathy with dysphagia, Sjogren's with extraglandular disease, and APS with thrombotic events.
• A privacy-preserving federated model will retain discrimination and calibration within a prespecified noninferiority margin relative to centralized training.
• Adding calibration-error trajectories to standard clinical scores will improve time-to-escalation prediction more than adding raw model probability alone.

Study design

• Multicenter retrospective-prospective cohort across autoimmune clinics
• Inputs: structured disease activity scores, lab trends, medication changes, clinician assessments, and AI model outputs
• Primary outcome: organ-threatening flare or escalation of care within 90 days
• Primary analysis: time-dependent survival model with calibration error as a dynamic predictor
• Validation: external transportability across sites using federated or encrypted aggregation

Limitations

• Calibration error may capture documentation quality as well as biology.
• Site-specific workflows could confound the apparent lead time of the score.
• Rare organ-threatening events may limit power for individual disease subgroups.
• Privacy-preserving computation can reduce leakage but does not remove all governance and bias risks.

Clinical significance

If confirmed, this would give rheumatology teams a deployable early-warning layer that is both clinically actionable and privacy-aware, enabling earlier escalation while reducing dependence on raw data pooling.

References

1. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25:44-56. DOI: 10.1038/s41591-018-0300-7
2. Esteva A, Robicquet A, Ramsundar B, et al. A guide to deep learning in healthcare. Nat Med. 2019;25:24-29. DOI: 10.1038/s41591-018-0316-z
3. Kaissis GA, Makowski MR, Rückert D, Braren RF. Secure, privacy-preserving and federated machine learning in medical imaging. Nat Mach Intell. 2020;2:305-311. DOI: 10.1038/s42256-020-0186-1
4. Froelicher D, Müller 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

LES AI • DeSci Rheumatology