Hypothesis

Bayesian online change-point detection (BOCPD) applied to multivariate time series of serial immunoassay panels (anti-dsDNA, C3, C4, anti-C1q, IL-6, CXCL10, BLyS) identifies posterior change-point probabilities that signal subclinical disease state transitions in SLE 6–14 weeks before patients meet composite flare criteria (SELENA-SLEDAI flare index or BILAG-2004 flare).

Background and Rationale

Current flare prediction in SLE relies on threshold-based biomarker alerts (e.g., complement drop, anti-dsDNA rise) applied independently per analyte. This approach ignores:

1. Cross-analyte correlation structure — complement consumption, autoantibody production, and cytokine surges are mechanistically coupled but evaluated in isolation
2. Run-length dynamics — the duration of a stable immunological regime carries predictive information lost by snapshot thresholds
3. Non-stationarity — SLE immunological trajectories exhibit regime-switching behavior that violates stationary assumptions underlying standard monitoring

BOCPD (Adams & MacKay, 2007) maintains a posterior distribution over the current run length — the time since the last change-point — and naturally handles multivariate inputs via conjugate predictive models. By modeling the joint distribution of immunoassay panels under a multivariate normal-inverse-Wishart conjugate prior, the algorithm detects coordinated shifts in mean and covariance structure that precede clinical flare.

Testable Predictions

1. Primary: BOCPD change-point posterior probability exceeding 0.7 on ≥2 consecutive biweekly samples predicts composite flare within 6–14 weeks with AUROC >0.82, outperforming univariate anti-dsDNA or complement thresholds (expected AUROC 0.65–0.72)
2. Secondary: The detected change-points cluster into ≥3 distinct immunological regime types (complement-dominant, cytokine-dominant, mixed), each associated with different flare organ manifestations (renal vs. musculoskeletal vs. mucocutaneous) with Cramér V >0.35
3. Tertiary: Integration of BOCPD alerts into a treat-to-target protocol simulation reduces simulated time-to-intervention by >4 weeks compared to standard monitoring, with a number-needed-to-treat of ≤5 for flare prevention

Proposed Methodology

• Cohort: Prospective longitudinal SLE cohort (n≥200) with biweekly multiplex immunoassay panels and monthly clinical assessments over ≥18 months
• Model: Multivariate BOCPD with normal-inverse-Wishart conjugate predictive model, hazard function h(τ) = 1/λ with λ estimated from historical flare intervals
• Validation: 5-fold temporal cross-validation respecting chronological ordering; external validation on independent registry cohort
• Comparison: Univariate threshold alerts, CUSUM charts, and HMM-based approaches

Limitations

• Biweekly sampling may miss rapid transitions; higher-frequency sampling would strengthen detection but increases cost
• Normal-inverse-Wishart conjugate assumes multivariate normality — cytokine distributions are often right-skewed, requiring log-transformation or non-parametric extensions
• BOCPD assumes independent change-points across time; real immunological transitions may exhibit hysteresis or memory effects requiring extensions (e.g., hierarchical BOCPD)
• Prospective validation essential — retrospective application risks survivorship bias from patients lost to follow-up during severe flares

Clinical Significance

Early detection of subclinical immunological regime shifts could enable preemptive therapeutic adjustment (hydroxychloroquine dose optimization, early steroid bridging, or belimumab initiation) during a critical window where intervention has maximum efficacy. Unlike black-box ML approaches, BOCPD provides interpretable posterior probabilities with natural uncertainty quantification, facilitating clinical adoption and regulatory defensibility.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology