Background

Disease activity in rheumatoid arthritis (RA) is typically modeled as a Markovian process where future states depend only on the present. However, clinical experience suggests that disease trajectories exhibit memory — patients with prolonged subclinical inflammation respond differently to DMARDs than those with acute-onset disease, even at identical DAS28 values. This long-range dependence (LRD) has never been formally quantified in RA.

Hypothesis

Serial DAS28 measurements in RA follow fractional Brownian motion (fBM) rather than standard Brownian motion, with patient-specific Hurst exponents (H) that encode clinically actionable information:

1. H > 0.5 (persistent/trending trajectories) identifies patients in whom early aggressive treatment will yield disproportionate long-term benefit (window of opportunity)
2. H ≈ 0.5 (memoryless/random walk) characterizes treatment-resistant stochastic disease requiring frequent reassessment
3. H < 0.5 (anti-persistent/mean-reverting) identifies patients likely to achieve spontaneous improvement where watchful waiting may be appropriate

Proposed Methodology

• Data: ≥500 RA patients with ≥8 serial DAS28 measurements over ≥24 months from registries (CORRONA, RABBIT, NOR-DMARD)
• Hurst estimation: Detrended fluctuation analysis (DFA) and rescaled range (R/S) analysis on individual DAS28 time series, with wavelet-based estimators as robustness check
• Validation: Compare fBM model fit vs. standard AR(1), ARIMA, and Ornstein-Uhlenbeck processes via Bayesian model comparison (WAIC/LOO-CV)
• Clinical prediction: Cox proportional hazards with time-varying H as covariate for CDAI remission at 12 months
• Confounding: Adjust for treatment changes, age, RF/ACPA status, baseline DAS28, and disease duration via inverse probability of treatment weighting

Testable Predictions

1. 60% of RA DAS28 trajectories will reject the null H = 0.5 (Kolmogorov-Smirnov test, α = 0.05)

2. Patients with H > 0.65 in the first 6 months who receive early biologic escalation will achieve CDAI remission at rates ≥20% higher than H-matched controls on csDMARDs alone
3. The fBM model will achieve superior predictive accuracy (ΔWAIC > 10) compared to AR(1) in ≥70% of individual trajectories
4. Anti-persistent trajectories (H < 0.35) will show ≥50% spontaneous DAS28 improvement without treatment escalation

Limitations

• DAS28 measurement frequency in registries (typically every 3-6 months) may be insufficient for reliable Hurst estimation; minimum 8 timepoints required per patient
• fBM assumes Gaussian increments, which may not hold for composite scores with floor/ceiling effects
• Treatment changes during observation violate stationarity assumptions — piecewise estimation with change-point detection is required
• Self-similar processes may be confounded by regime-switching dynamics (addressed by comparing against hidden Markov models)

Clinical Significance

If validated, Hurst exponent estimation from routine DAS28 monitoring would provide a novel, zero-cost biomarker for treatment decision-making. Persistent trajectories (high H) would mandate early aggressive therapy, while anti-persistent patterns (low H) could justify conservative management — personalizing the "treat-to-target" paradigm with mathematical rigor derived from stochastic process theory.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology