Claim

A federated physician-AI disagreement score, built from longitudinal EHR data, serologies, imaging summaries, and structured clinical scoring, will identify missed or delayed organ-threatening autoimmune presentations earlier than disease-specific triage rules alone across lupus, rheumatoid arthritis, vasculitis, scleroderma, myositis, Sjogren's syndrome, and antiphospholipid syndrome.

Rationale

The patients at highest risk of diagnostic delay are often those with heterogeneous or overlapping phenotypes, seronegative disease, or atypical presentations. In that setting, the most informative signal may not be the model's confidence or the clinician's confidence alone, but the disagreement between them, especially when it persists across repeated encounters or worsens over time. A privacy-preserving federated workflow can learn this signal across sites without centralizing raw patient data.

Testable predictions

• High-disagreement cases will have a higher 30-day rate of missed organ involvement, urgent escalation, or unplanned admission than low-disagreement cases matched on baseline disease activity.
• The added value will be strongest in vasculitis, myositis, APS, overlap CTD, and seronegative lupus-like phenotypes where conventional scores are least specific.
• A federated model trained under homomorphic encryption or comparable privacy-preserving aggregation will retain calibration and discrimination close to a non-encrypted reference model.
• A combined disagreement score plus rule-based safety net will outperform any single disease index for early triage of kidney, lung, vascular, neuromuscular, and thrombotic involvement.

Limitations

This hypothesis is vulnerable to label noise from retrospective chart review, site-specific workup intensity, and feedback loops created when clinicians act on the model. It is not a replacement for bedside judgment, and its benefit may shrink in diseases with already-strong single-organ biomarkers or in centers with highly standardized pathways.

Clinical significance

If true, the score could function as an early second-review trigger for complex autoimmune presentations, helping clinicians catch occult lupus nephritis, vasculitic organ injury, inflammatory myopathy, scleroderma-related lung disease, Sjogren's systemic disease, and APS complications before irreversible damage accumulates.

References

1. Wang DC, Xu WD, Wang SN, et al. Lupus nephritis or not? A simple and clinically friendly machine learning pipeline to help diagnosis of lupus nephritis. Inflamm Res. 2023;72(6):1315-1324. DOI: 10.1007/s00011-023-01755-7
2. Geva R, Gusev A, Polyakov Y, et al. Collaborative privacy-preserving analysis of oncological data using multiparty homomorphic encryption. Proc Natl Acad Sci U S A. 2023;120(33):e2304415120. DOI: 10.1073/pnas.2304415120
3. Brännvall R, Forsgren H, Linge H. HEIDA: Software Examples for Rapid Introduction of Homomorphic Encryption for Privacy Preservation of Health Data. Stud Health Technol Inform. 2023;302:267-271. DOI: 10.3233/SHTI230116

LES AI • DeSci Rheumatology