Background

Systemic sclerosis (SSc) skin fibrosis progression measured by modified Rodnan Skin Score (mRSS) follows non-linear trajectories with abrupt accelerations that current linear monitoring fails to anticipate. Standard biomarker thresholds collapse high-dimensional immune information into univariate summaries, discarding distributional structure that may encode early fibrotic commitment signals.

Hypothesis

We hypothesize that Maximum Mean Discrepancy (MMD) computed in a reproducing kernel Hilbert space (RKHS) over serial multiparameter flow cytometry panels (≥20 markers: CD4/CD8/Treg/Th17/monocyte/B-cell/NK subsets with activation markers) will detect distributional regime shifts — defined as statistically significant increases in MMD between consecutive 3-month sampling windows — that predict mRSS acceleration (≥5-point increase over 12 months) 6–14 months before clinical detection, with AUROC >0.82.

The kernel choice is critical: we propose a characteristic kernel (Gaussian RBF with median heuristic bandwidth selection) ensuring that MMD = 0 if and only if distributions are identical, providing a proper metric on the space of probability distributions over immunophenotypic features.

Mechanistic Rationale

Fibrotic commitment in SSc involves coordinated shifts across multiple immune compartments — Th2/Th17 polarization, monocyte-to-fibrocyte transition, regulatory T-cell functional exhaustion, and plasmablast expansion — that individually may remain within reference ranges while their joint distributional geometry undergoes detectable deformation. MMD in RKHS captures these joint distributional changes without requiring explicit density estimation, which is intractable in ≥20 dimensions with clinical sample sizes.

Testable Predictions

1. Primary: MMD between consecutive 3-month immunophenotypic distributions will exceed the permutation-based significance threshold (α = 0.01, 10,000 permutations) ≥6 months before mRSS acceleration in >75% of progressors versus <15% of stable patients.
2. Secondary: The MMD trajectory slope (linear regression over 4+ consecutive windows) will independently predict progression in multivariable models adjusted for baseline mRSS, disease duration, anti-Scl-70 status, and lung involvement.
3. Kernel specificity: Gaussian RBF kernel will outperform linear and polynomial kernels (AUROC difference >0.08), confirming that non-linear distributional features carry predictive information.
4. Dimensionality: MMD computed on the full panel will outperform MMD on any single immune compartment subset, demonstrating that cross-compartment distributional coupling is necessary for optimal prediction.

Study Design

Prospective observational cohort: 150 early diffuse cutaneous SSc patients (<3 years from first non-Raynaud symptom), multiparameter flow cytometry every 3 months for 24 months, mRSS every 3 months. Sample size powered for 30% expected progression rate (45 progressors), achieving >80% power for AUROC >0.82 vs null 0.65.

Limitations

• Computational: MMD permutation testing scales as O(n²) per comparison; with panels of ~10⁴ cells per sample, each window comparison requires ~10⁸ kernel evaluations × 10⁴ permutations — feasible but requiring GPU acceleration.
• Bandwidth sensitivity: Median heuristic provides a reasonable default but may not be optimal; adaptive kernel selection or multiple kernel learning could improve performance but adds complexity.
• Confounding: Intercurrent infections or vaccinations could transiently shift immune distributions, generating false positives; clinical event logging and sensitivity analyses excluding ±4-week windows around acute events are necessary.
• Generalizability: Flow cytometry panel heterogeneity across centers limits multi-site applicability without harmonization protocols (e.g., EuroFlow).
• Interpretability: MMD detects that a distributional shift occurred but does not directly identify which immune populations drove it; post-hoc kernel SHAP or witness function analysis would be needed for mechanistic interpretation.

Clinical Significance

Early detection of fibrotic acceleration in SSc could enable preemptive therapeutic intensification (e.g., mycophenolate escalation, nintedanib initiation, or autologous HSCT evaluation) during a window when fibrosis may still be reversible. Current clinical practice detects progression only after mRSS increase is clinically apparent, by which time fibrotic remodeling may be established. A distributional monitoring framework based on MMD would represent a paradigm shift from threshold-based to geometry-based immune surveillance in autoimmune fibrotic disease.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology