Hypothesis

Functional principal component analysis (FPCA) applied to biweekly tender joint count (TJC28) and swollen joint count (SJC28) trajectories during the first 8 weeks of methotrexate-based csDMARD therapy extracts latent functional modes whose scores, combined with baseline CYP2C19/MTHFR pharmacogenomic variants and serum folate levels, predict CDAI remission (≤2.8) at week 16 with AUROC >0.82 — outperforming single-timepoint DAS28 or CDAI change scores.

Background and Rationale

Current treat-to-target protocols in rheumatoid arthritis (RA) rely on disease activity assessments at discrete 3–6 month intervals, discarding the rich temporal information embedded in how joint counts evolve over time. Two patients may share identical DAS28 values at week 12 yet arrive there via fundamentally different trajectories — one through monotonic decline, the other through oscillatory partial response — with divergent long-term outcomes.

Functional data analysis (FDA) treats each patient's longitudinal joint count series not as a vector of discrete measurements but as a realization of a smooth stochastic process. FPCA decomposes the covariance structure of these functional observations into orthogonal modes of variation (eigenfunctions), each capturing a distinct pattern of temporal change: overall level, rate of early decline, mid-course inflection, etc. The individual scores on these components form a low-dimensional, information-dense representation of the entire trajectory.

Pharmacogenomic modifiers — particularly MTHFR C677T/A1298C (affecting methotrexate polyglutamation efficiency) and CYP2C19 metabolizer status (relevant to leflunomide co-therapy) — introduce systematic trajectory heterogeneity that single-timepoint models cannot capture but FPCA modes naturally absorb.

Proposed Methodology

1. Data: Prospective cohort of ≥300 early RA patients (2010 ACR/EULAR criteria, disease duration <2 years) initiating MTX ± leflunomide, with biweekly TJC28/SJC28 assessments for 16 weeks
2. Functional smoothing: B-spline basis expansion (order 4, 8 knots) with roughness penalty (GCV-selected λ) applied separately to TJC and SJC curves
3. FPCA: Extract first 3–4 FPCs explaining ≥90% variance from each joint count type
4. Prediction model: Bayesian logistic regression with horseshoe prior on FPC scores + pharmacogenomic indicators (MTHFR genotype, CYP2C19 metabolizer status, baseline folate) + baseline covariates (RF/ACPA status, erosion count)
5. Validation: 10-fold cross-validation with calibration assessment (Brier score, calibration slope)
6. Comparator: EULAR response criteria at week 12, ΔDAS28 at week 8, and machine learning on raw discrete timepoints

Testable Predictions

• The first FPC (overall trajectory level) combined with the second FPC (rate of early change) will jointly explain >70% of variance in joint count evolution
• MTHFR 677TT homozygotes will show a distinct third FPC mode characterized by delayed response onset (inflection point shifted >2 weeks later)
• The FDA-based model will achieve calibration slope 0.85–1.15, superior to discrete-timepoint random forest (expected calibration slope <0.75)
• At a clinically useful sensitivity threshold of 0.80 for remission prediction, the specificity will exceed 0.70, enabling confident early escalation decisions by week 8

Limitations

• Biweekly assessment frequency may underestimate true trajectory complexity; weekly assessments would improve FPCA resolution but increase patient burden
• B-spline smoothing assumes continuous underlying process, which may not hold during intercurrent infections or NSAID dose changes
• Pharmacogenomic associations with MTX response remain inconsistent across populations; effect sizes may be insufficient in non-European cohorts
• FPCA assumes Gaussian functional observations; joint counts are bounded non-negative integers requiring careful transformation (square-root or compositional log-ratio)
• Single-center design limits generalizability; federated FDA across DeSci-linked registries could address this in future work

Clinical Significance

If validated, FPCA-based trajectory modeling would enable week 8 go/no-go decisions for csDMARD therapy — 4–8 weeks earlier than current EULAR recommendations — by leveraging the shape of the response curve rather than its snapshot value. This directly reduces cumulative inflammatory burden during the critical early RA window of opportunity. Integration into DeSci infrastructure would enable privacy-preserving federated FPCA across institutions without centralizing raw longitudinal data.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology