Hypothesis: A Bayesian chaotic latent-state model that fuses longitudinal biomarkers, foundation-model embeddings, and pharmacogenomic priors will outperform conventional disease-activity scores in forecasting autoimmune rheumatic flares and in enriching clinical trials for treatment response heterogeneity.

Rationale: Autoimmune rheumatic disease trajectories often look irregular, but irregularity does not imply randomness. A plausible mechanistic explanation is that patients move through a low-dimensional stochastic dynamical system with regime shifts triggered by immune feedback loops, treatment perturbation, infection, adherence gaps, and genotype-dependent drug response. Standard linear regression and static classifiers compress that behavior into overly smooth averages. A hierarchical Bayesian state-space model can instead represent latent inflammatory states, transition intensities, and patient-specific random effects while explicitly quantifying uncertainty. Foundation models can supply dense representations of chart text, pathology, imaging reports, and symptom narratives; genomics and pharmacogenomics can act as priors on state transitions and treatment response; and sparse laboratory time series can anchor the posterior to measurable biology.

Testable predictions:

1. In prospective cohorts, a Bayesian switching state-space model with foundation-model embeddings will achieve better calibrated flare prediction than conventional indices such as DAS28, SLEDAI, or physician global assessment alone, especially 2 to 8 weeks before flare.
2. The latent-state process will exhibit evidence of low-dimensional chaotic dynamics or sensitive dependence on initial conditions in a subset of patients, detectable by improved short-horizon forecasts and non-linear dynamical diagnostics relative to linear baselines.
3. Genotype-stratified transition matrices will show clinically meaningful pharmacogenomic effects, for example differential probabilities of remission maintenance, toxicity, or flare under TNF inhibitors, JAK inhibitors, methotrexate, or glucocorticoids.
4. Trial enrichment using posterior probability of an imminent flare or nonresponse will increase event rates, reduce sample size, and improve treatment-effect precision compared with unselected enrollment.
5. Calibration will degrade if the model is trained on one health system and deployed in another without hierarchical site-level updating, which makes transportability a key biostatistical endpoint rather than an afterthought.

Clinical significance: If true, this framework would turn flare prediction from a static risk score into a dynamic decision-support problem. That could support earlier escalation, tighter monitoring, smarter tapering, and more efficient clinical trial design. It also provides a concrete path for DeSci infrastructure: federated or privacy-preserving learning across sites, reproducible Bayesian workflows, and open model governance for auditability.

Limitations: This hypothesis may fail if observed nonlinearity is mostly measurement noise, if foundation-model embeddings add little beyond structured data, or if genotype effects are too small for routine use. Short follow-up, treatment confounding, and missingness could also produce misleading apparent chaos. Any real-world implementation must pre-specify calibration, external validation, subgroup analysis, and harm monitoring before clinical use.

Methods to falsify it: Compare Bayesian switching state-space models against elastic net, random forest, and gradient-boosted baselines using time-split validation, calibration curves, decision-curve analysis, and prospective trial-enrichment simulations. If the Bayesian model does not improve both discrimination and calibration, or if the supposed chaotic structure disappears under external validation, the hypothesis should be rejected.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology