Background

Current rheumatoid arthritis (RA) disease activity assessment relies on discrete clinic visits where composite indices (DAS28, CDAI, SDAI) are calculated from point-in-time measurements. Between visits, disease trajectories are unobserved, creating blind spots where flares initiate and escalate before clinical detection. Wearable sensors now generate continuous streams of inflammation-correlated signals (joint temperature, grip strength variability, accelerometry-derived morning stiffness duration, heart rate variability), while home-based laboratory platforms enable frequent CRP and ESR self-measurement.

Hypothesis

We hypothesize that a sequential Monte Carlo (SMC) particle filter operating on fused wearable biometric streams and intermittent laboratory values can maintain a real-time posterior distribution over a latent disease activity state variable, and that the particle filter's effective sample size (ESS) collapse — indicating rapid posterior concentration — serves as an early warning signal for flare onset 2–6 weeks before DAS28-CRP crosses the moderate activity threshold (>3.2).

Proposed Model

The state-space model defines:

• State equation: Latent disease activity d(t) evolves as an Ornstein-Uhlenbeck process with patient-specific mean-reversion rate (θ), equilibrium level (μ), and volatility (σ), estimated online via sufficient statistics.
• Observation equation: Wearable signals and laboratory values are conditionally independent given d(t), with sensor-specific likelihood functions calibrated during a 4-week run-in period.
• Particle filter: N=5000 particles with systematic resampling. Adaptive tempering handles multimodal posteriors during treatment transitions. Liu-West kernel smoothing maintains parameter diversity.

The key innovation is using ESS dynamics as a diagnostic: when ESS drops below N/3 over consecutive time steps, this indicates the posterior is concentrating rapidly — the system is transitioning between disease states. This ESS collapse precedes clinical composite index changes because the particle filter integrates high-frequency wearable data that captures early pathophysiological changes invisible to intermittent clinical assessment.

Testable Predictions

1. The particle filter's median posterior d(t) will correlate with clinic-measured DAS28-CRP at r > 0.80 (Spearman) across validation visits.
2. ESS collapse events (ESS < N/3 sustained ≥48 hours) will precede DAS28-CRP flare threshold crossing by 2–6 weeks with sensitivity >75% and positive predictive value >60%.
3. The continuous posterior will identify ≥30% of subclinical flares (DAS28-CRP increase >0.6 units without crossing 3.2) missed by standard visit-based monitoring.
4. Patient-specific OU parameters (θ, μ, σ) estimated online will stabilize within 8 weeks and discriminate treatment responders from non-responders with AUROC >0.75.

Proposed Validation

Prospective cohort study: 150 RA patients, 6-month follow-up. Continuous wrist-worn sensor + monthly home CRP. Clinic visits every 4 weeks for ground-truth DAS28-CRP. Primary endpoint: time-dependent AUROC for flare prediction at 2, 4, and 6-week horizons. Comparison against: (a) last-observation-carried-forward, (b) linear interpolation, (c) Kalman filter (Gaussian assumption).

Limitations

• Wearable signal quality degrades with poor adherence; missing data patterns may be informative (not missing at random).
• OU process assumes continuous mean-reversion, which may not capture abrupt medication changes — jump-diffusion extensions may be needed.
• Particle degeneracy risk in high-dimensional extensions (adding multiple joint-specific channels); Rao-Blackwellization required.
• Calibration period (4 weeks) delays clinical utility.
• Sensor heterogeneity across devices requires harmonization layer.

Clinical Significance

If validated, SMC-based continuous disease monitoring would transform RA management from reactive (treat after flare detection at clinic visit) to preemptive (intervene at first posterior evidence of trajectory change). The probabilistic framework naturally quantifies uncertainty, enabling risk-stratified clinical decision support. Integration with pharmacogenomic priors (CYP metabolism rates, HLA-mediated drug response probability) could further personalize the state-space model, creating a patient-specific digital twin for real-time therapeutic optimization.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology