Autoimmune rheumatic disease may not flare as a smooth linear process; instead, some patients may move through a noisy, low-dimensional transition regime in which small perturbations in inflammation, medication exposure, or adherence produce disproportionate clinical change. I hypothesize that a Bayesian neural state-space model with explicit change-point and censoring components can recover this latent dynamics better than conventional activity scores. The model would fuse longitudinal biomarkers, symptom trajectories, medication timing, and unstructured clinical text embeddings from a foundation model, while using pharmacogenomic priors from HLA, CYP, TPMT/NUDT15, and related loci to modulate transition probability and toxicity risk.

Testable predictions: (1) in repeated-measures cohorts, the posterior transition probability, predictive entropy, and a proxy for local dynamical instability will rise days to weeks before flare or treatment failure, preceding threshold-based disease indices; (2) models that include pharmacogenomic priors will improve calibration and net benefit for drug-specific response and toxicity prediction, especially for methotrexate, azathioprine, JAK inhibitors, and biologics; (3) trial enrichment based on posterior flare risk will increase event rates and reduce required sample size relative to conventional inclusion criteria; and (4) external-site performance will remain stable only if the model explicitly accounts for informative visit timing and treatment censoring. The hypothesis is falsifiable: if the Bayesian state-space model does not outperform standard mixed-effects or Cox-based baselines on temporally held-out and site-held-out validation, the chaotic-transition claim is weak.

Limitations: apparent chaos may partly reflect unmeasured treatment changes, sparse sampling, regression to the mean, or site-level documentation artifacts. Foundation-model embeddings can inherit corpus bias and may obscure mechanistic interpretability. Pharmacogenomic effects are often drug-specific and ancestry-dependent, so any gain may not generalize across populations without careful stratification and calibration.

Clinical significance: if confirmed, this framework could support earlier flare interception, better drug matching, more efficient adaptive trials, and a reusable DeSci-ready analytics layer for federated rheumatology cohorts without forcing a single deterministic disease trajectory onto biologically heterogeneous patients.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology