Systemic lupus erythematosus (SLE) may not progress toward flare as a smooth linear process. A competing view is that flare emergence reflects entry into a transient high-risk latent state, or pre-flare attractor, driven by coupled immune, molecular, and symptom dynamics. I hypothesize that a Bayesian latent-state model combining longitudinal interferon-related transcriptomics, complement trajectory, anti-dsDNA trajectory, proteinuria where relevant, patient-reported symptoms, and treatment exposure can identify this pre-flare attractor 2 to 6 weeks before conventional activity indices or routine clinician assessment classify flare.

The core scientific claim is that SLE flare risk is better represented as probabilistic switching between hidden states than as threshold crossing on isolated biomarkers. If this is correct, then within-patient temporal structure, including variance, autocorrelation, and cross-lagged biomarker interactions, should carry predictive information that static snapshots miss. Foundation models for multimodal clinical time series may improve feature extraction, but the main inferential layer should remain Bayesian and interpretable so that uncertainty, state transitions, and treatment effects can be audited.

Testable predictions:

1. A Bayesian hidden Markov or state-space model trained on dense longitudinal SLE data will outperform static or last-observation models for predicting physician-adjudicated flare within 14, 28, and 42 days.
2. The model will identify a reproducible latent high-risk state characterized by increasing interferon-pathway instability, complement decline velocity, anti-dsDNA acceleration, and rising symptom volatility, even when absolute values remain within ranges not yet labeled as flare.
3. Transition probability into the high-risk state will increase before renal and non-renal flares, but the dominant leading indicators will differ by organ phenotype.
4. Trial enrichment based on posterior probability of the pre-flare state will reduce required sample size in flare-prevention studies relative to enrollment strategies based only on baseline disease activity.

How this could be tested: A prospective multicenter cohort could collect weekly or biweekly clinical variables, patient-reported outcomes, and blood transcriptomic panels for 6 to 12 months, with blinded flare adjudication using prespecified criteria. The primary analysis would compare discrimination, calibration, and decision utility of the Bayesian latent-state model against conventional predictors such as SLEDAI components, complement alone, anti-dsDNA alone, and clinician global assessment. A secondary translational analysis could evaluate whether foundation-model embeddings from longitudinal clinical plus omics data improve early warning performance while preserving calibration once passed through the Bayesian state model. A methodological extension would simulate adaptive trial enrollment using posterior state probability as an inclusion criterion and estimate gains in power and efficiency.

Limitations: This hypothesis depends on access to dense longitudinal data that many real-world clinics do not collect. Transcriptomic measurements may be too expensive for routine deployment and may reduce generalizability across settings. Hidden-state models can overfit if sampling frequency is inconsistent or if treatment changes are not explicitly modeled. Causal interpretation would remain limited because latent states may reflect correlated disease activity rather than mechanistic flare drivers. External validation across ancestry groups, organ phenotypes, and treatment regimens would be essential.

Clinical significance: If validated, this framework could shift SLE monitoring from reactive flare detection to probabilistic preemption. That would matter for patient safety, trial design, and precision immunomodulation. Clinically, it could support earlier follow-up, targeted lab reassessment, or short-term preventive adjustments before irreversible organ injury occurs. Methodologically, it would support the broader idea that rheumatic disease activity may contain stochastic and weakly chaotic structure that is more visible in trajectories than in single visits.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology