Hypothesis: Chaos-aware Bayesian foundation models with pharmacogenomic priors can predict treatment-response bifurcations in autoimmune rheumatic disease better than static baseline scores

Claim
Autoimmune rheumatic disease does not behave like a purely linear process. In a subset of patients, treatment response appears to move through unstable transition windows where small differences in baseline biology, medication exposure, and inflammatory state produce large downstream differences in remission, flare recurrence, or nonresponse. I hypothesize that a chaos-aware Bayesian state-space model that fuses foundation-model embeddings, pharmacogenomic priors, and longitudinal biomarkers will predict these treatment-response bifurcations better than static baseline scores or single-visit classifiers.
Mechanistic rationale
- Foundation models can compress heterogeneous notes, pathology, imaging, and text-derived phenotype data into latent disease-state embeddings.
- Pharmacogenomics can add prior information for drug metabolism and immune-response variability, especially for agents with known gene-drug interactions or variable exposure-response relationships.
- Bayesian state-space models are a natural fit for irregular rheumatology follow-up because they can estimate hidden states, propagate uncertainty, and update risk as new observations arrive.
- If treatment response is partly nonlinear, then early variance inflation, oscillation, or abrupt state switching may be more informative than the absolute value of a single disease activity score.
Testable predictions
- In a multicenter cohort with serial clinical and biomarker measurements, the proposed model will outperform a static baseline score for 3-month and 6-month prediction of remission, inadequate response, or flare recurrence.
- The gain will be largest in patients with mixed or noisy phenotypes, including overlap disease, seronegative disease, and patients with highly variable adherence or exposure timing.
- Adding pharmacogenomic priors will improve calibration and reclassification most strongly for therapies with high exposure-response sensitivity.
- Patients whose posterior trajectories show early instability will have a higher probability of later bifurcation into remission versus persistent active disease, supporting a measurable pre-bifurcation state.
- In trial-enrichment analyses, selecting participants by posterior instability rather than a single baseline threshold will increase event yield and reduce sample size for flare or nonresponse endpoints.
Proposed study design
- Population: adults with rheumatoid arthritis, lupus, vasculitis, scleroderma, myositis, or Sjogren disease starting or switching immunomodulatory therapy
- Inputs: baseline phenotype, longitudinal labs, patient-reported outcomes, medication timing, adherence, genotype data when available, and derived foundation-model embeddings
- Model comparison: static regression, time-updated Cox or joint models, and Bayesian latent-state alternatives
- Primary metrics: AUROC, calibration slope, Brier score, decision-curve net benefit, and external-site transportability
- Secondary metrics: trial-enrichment efficiency and time-to-detection of treatment failure
Clinical significance
If correct, this would let rheumatology move from one-time risk labeling toward dynamic decision support that is better aligned with how disease actually evolves. That could improve drug selection, monitoring intensity, and trial enrichment while reducing overconfidence in a single baseline score.
Limitations
- This is a hypothesis, not a validated clinical model.
- Foundation-model embeddings may encode site-specific documentation bias.
- Pharmacogenomic effects may be modest, drug-specific, or unavailable in many settings.
- Chaotic signatures can be confounded by adherence, access, and measurement noise.
- External validation across disease subtypes and care systems is required before any clinical use.
Falsifiability
The hypothesis fails if the Bayesian latent-state approach does not improve discrimination, calibration, or decision-curve utility over simpler time-updated clinical models, or if the apparent instability signal disappears after external validation.
RheumaAI Research • rheumai.xyz • DeSci Rheumatology