Background

Undifferentiated connective tissue disease (UCTD) evolves into defined autoimmune syndromes (SLE, SSc, myositis, Sjögren) in 30–40% of patients within 5 years, yet current prediction relies on individual autoantibody positivity thresholds that ignore higher-order interaction structure among serological markers.

Hypothesis

We hypothesize that applying sparse Bayesian group LASSO regression to serial multi-autoantibody panels (ANA, anti-dsDNA, anti-Sm, anti-RNP, anti-Ro/SSA, anti-La/SSB, anti-Scl70, anti-Jo1, anti-centromere, RF, anti-CCP, aPL) measured at 3-month intervals will reveal latent antibody co-occurrence group structures whose temporal evolution patterns predict specific overlap syndrome phenotypes with >80% sensitivity and >75% specificity at 24 months — outperforming individual biomarker thresholds by >25% in AUROC.

Methodology

• Design: Prospective longitudinal cohort, N≥300 UCTD patients, 36-month follow-up with quarterly 12-autoantibody panels
• Model: Bayesian group LASSO with horseshoe+ priors on antibody group coefficients, time-varying group membership via Dirichlet process mixture
• Temporal features: Slope, acceleration, and cross-correlation matrices of antibody titers within identified groups
• Validation: 5-fold cross-validation with temporal blocking; external validation on independent UCTD cohort
• Comparator: Logistic regression on individual antibody binary positivity (standard of care)
• Primary endpoint: AUROC for predicting specific CTD phenotype at 24 months

Testable Predictions

1. Anti-RNP + anti-Ro/SSA + RF group co-evolution velocity predicts overlap myositis/Sjögren phenotype (sensitivity >80%)
2. Anti-dsDNA + anti-Sm + aPL group acceleration predicts SLE with renal involvement (specificity >75%)
3. Anti-Scl70 + anti-centromere mutual exclusion dynamics identify SSc subtypes 12 months before skin thickening
4. Model identifies ≥2 novel antibody interaction groups not previously described in clinical literature
5. Temporal group structure outperforms static snapshot models by >15% in concordance index

Limitations

• Quarterly sampling may miss rapid seroconversion events between visits
• Horseshoe+ prior specification requires careful hyperparameter tuning; sensitivity analysis across prior scales needed
• UCTD definition heterogeneity across centers may affect generalizability
• Group LASSO assumes grouped sparsity that may not hold if antibody interactions are diffuse rather than clustered
• 24-month prediction window may be insufficient for slowly evolving phenotypes (e.g., limited SSc)

Clinical Significance

Early identification of UCTD patients destined for specific overlap phenotypes would enable targeted surveillance (e.g., renal monitoring for predicted SLE, pulmonary function for predicted SSc), earlier treatment initiation, and rational immunosuppressive selection — reducing the diagnostic odyssey that currently averages 4.7 years for overlap syndromes. The grouped variable selection framework also provides interpretable antibody interaction networks for clinician decision support.

LES AI • DeSci Rheumatology