Background

Drug-free remission (DFR) in rheumatoid arthritis remains poorly predicted by conventional models. Standard Cox regression treats all patients as eventually relapsing, ignoring the cured fraction — patients who achieve sustained DFR indefinitely. Meanwhile, pharmacogenomic markers (HLA-DRB1 shared epitope, CYP3A4/CYP2C19 metabolizer status, NAT2 acetylator phenotype) interact with treatment history in time-dependent ways that static models cannot capture.

Hypothesis

A cure-mixture survival model (Berkson-Gage framework) with time-varying pharmacogenomic covariates will significantly outperform standard Cox and random survival forest models in predicting which RA patients can successfully discontinue biologics without relapse.

Specifically:

1. The cure fraction π(x) is modeled via logistic regression on baseline HLA-DRB1 shared epitope allele dosage, ACPA status, and cumulative biologic exposure duration
2. The susceptible subgroup survival function S(t|uncured) follows a flexible Royston-Parmar spline model with time-varying coefficients for CYP metabolizer status × biologic type interaction
3. A Bayesian estimation framework with weakly informative priors (half-Cauchy on variance components) provides posterior distributions over cure probability per patient

Testable Predictions

• P1: The cure-mixture model achieves C-statistic ≥ 0.78 for 2-year DFR prediction vs. ≤ 0.68 for standard Cox (validated on held-out cohort)
• P2: HLA-DRB1*04:01 homozygosity reduces estimated cure fraction by ≥ 40% compared to non-shared-epitope carriers (posterior 95% credible interval excludes null)
• P3: CYP2C19 poor metabolizers on conventional DMARDs show time-varying hazard ratio that crosses 1.0 between months 6-12 post-discontinuation, creating a critical surveillance window not detected by time-constant models
• P4: Model calibration (Hosmer-Lemeshow across deciles) shows p > 0.20, indicating no systematic miscalibration

Statistical Framework

• Likelihood: L = ∏[π + (1-π)·S(t)]^δ · [(1-π)·f(t)]^(1-δ) where δ = censoring indicator
• MCMC: No-U-Turn Sampler (NUTS) via Stan, 4 chains × 2000 iterations post-warmup
• Model comparison: WAIC and LOO-CV with Pareto-k diagnostics
• Time-varying coefficients: B-spline basis expansion with penalized complexity priors

Limitations

• Cure-mixture identifiability requires sufficient follow-up (≥5 years) with plateau in Kaplan-Meier curve
• HLA genotyping not universally available in clinical practice — limits immediate generalizability
• Biologic switching sequences create complex treatment-confounder feedback that marginal structural models may handle better for causal claims
• Sample size requirements for interaction terms (HLA × CYP × biologic type) may exceed single-center capacity — multi-site federated estimation preferred

Clinical Significance

Accurate identification of the curable subgroup could safely guide biologic tapering in 15-30% of stable RA patients, reducing unnecessary immunosuppression exposure and annual treatment costs ($15,000-$40,000 USD per patient). The time-varying pharmacogenomic interactions would define patient-specific surveillance schedules during discontinuation, replacing the current one-size-fits-all monitoring approach.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology