Background

Rare adverse events (AEs) in rheumatology trials — serious infections, malignancy, cardiovascular events under biologics — are typically modeled as homogeneous Poisson processes, assuming constant hazard rates across the trial period. This assumption systematically underestimates clustering risk: AEs tend to aggregate in temporal windows driven by latent immunological states (e.g., transient lymphopenia, complement consumption troughs, or corticosteroid taper-induced rebound).

Hypothesis

Modeling rare AEs in rheumatology trials as doubly stochastic Poisson processes (Cox processes), where the intensity function λ(t) is itself a stochastic process driven by latent immunological covariates, will:

1. Detect AE clustering that standard Poisson models miss, identifying 2–4 temporal vulnerability windows per trial arm
2. Improve AE rate estimation by 20–40% in mean squared error versus homogeneous Poisson baselines
3. Reveal shared latent intensity drivers across different AE categories (infection, cardiovascular, malignancy) via shared frailty components in the Cox process specification
4. Enable real-time adaptive safety monitoring through online Bayesian updating of the latent intensity surface

Proposed Model

Let N(t) be the counting process for AEs. Under the Cox process:

• λ(t) = λ₀ · exp(X(t)β + Z(t)), where X(t) are time-varying immunological covariates (lymphocyte counts, CRP, complement C3/C4), β are regression coefficients, and Z(t) is a Gaussian process with Matérn covariance capturing residual latent variation
• Prior on Z(t): GP(0, k_ν(s,t)) with ν = 3/2 Matérn kernel, length-scale estimated via empirical Bayes
• Posterior inference via particle MCMC (PMCMC) with sequential Monte Carlo for the latent path

Testable Predictions

1. Retrospective analysis of ≥3 Phase III rheumatology trials (anti-TNF, JAKi, anti-IL-6) will show Cox process models achieve lower DIC/WAIC than homogeneous Poisson by >10 units
2. Identified vulnerability windows will correlate with measurable immunological troughs (lymphocyte nadir, complement consumption) with Spearman ρ > 0.5
3. Cross-AE-category shared frailty components will explain >30% of the excess clustering variance
4. Prospective application as an adaptive safety monitoring tool will flag clustering events 2–6 weeks earlier than traditional DSMB frequency-based approaches

Limitations

• Rare events inherently limit statistical power; pooling across trials introduces heterogeneity
• Latent immunological covariates (lymphocyte subsets, complement dynamics) are not uniformly collected across legacy trials
• PMCMC is computationally expensive — real-time implementation requires approximate inference (variational or Laplace)
• Shared frailty interpretation assumes common latent drivers, which may not hold across mechanistically distinct AE categories
• Regulatory acceptance of Cox process-based safety signals requires validation against established pharmacovigilance methods

Clinical Significance

If validated, Cox process safety monitoring could transform DSMB operations in rheumatology trials by providing continuous, probabilistic AE clustering detection rather than periodic frequency counts. This enables earlier safety signals, more precise risk–benefit assessment during interim analyses, and patient-specific vulnerability profiling through the latent intensity surface. For DeSci infrastructure, the Bayesian updating framework naturally supports federated safety monitoring across decentralized trial sites without sharing individual patient data.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology