Background

Current thrombotic risk stratification in antiphospholipid syndrome (APS) relies on categorical aPL positivity (lupus anticoagulant, anti-cardiolipin, anti-β2GPI) and clinical history. The Global APS Score (GAPSS) improves prediction but treats variables as independent linear contributors, ignoring nonlinear interactions between immunological pathways that converge on thrombosis.

Hypothesis

A Bayesian network model integrating (1) quantitative aPL isotype profiles (IgG/IgA/IgM anti-cardiolipin and anti-β2GPI titers as continuous variables), (2) serial complement consumption trajectories (C3, C4, CH50 slopes over 6-month windows), and (3) circulating NETosis biomarkers (cell-free DNA, citrullinated histone H3, myeloperoxidase-DNA complexes) will outperform GAPSS in predicting first and recurrent thrombotic events at 12 months, with an expected AUC improvement of ≥0.12.

Mechanistic Rationale

aPL antibodies activate complement via the classical pathway and simultaneously prime neutrophils for NETosis. NETs provide a phospholipid-rich scaffold that amplifies thrombin generation. These three pathways — humoral autoimmunity, complement activation, and innate immune dysregulation — are mechanistically coupled but measured independently in clinical practice. A directed acyclic graph (DAG) structure can capture conditional dependencies: aPL titer → complement consumption → NET release → thrombotic probability, with feedback from NET-bound complement amplification.

Testable Predictions

1. The Bayesian network model achieves AUC ≥0.85 for 12-month thrombotic events vs. GAPSS AUC ~0.70-0.75 in validation cohorts
2. Complement trajectory slope (not single-point values) is a stronger conditional predictor than static C3/C4 levels (likelihood ratio ≥3.0)
3. CitH3 levels above the 75th percentile combined with triple aPL positivity identify a subgroup with >40% annual thrombosis rate vs. <10% in the low-risk stratum
4. The model identifies a "complement-driven" phenotype (~20% of APS patients) where thrombosis risk is primarily mediated by complement consumption independent of aPL titer magnitude

Study Design

Prospective multicenter cohort, n≥400 APS patients (ACR/EULAR 2023 classification criteria), 24-month follow-up. Bayesian network structure learned via hill-climbing with BIC scoring, validated with 10-fold cross-validation and external cohort. Primary endpoint: first/recurrent arterial or venous thrombosis. Calibration assessed via Brier score and calibration plots.

Limitations

• NETosis biomarkers lack standardized assays; inter-laboratory variability may reduce reproducibility
• Bayesian network structure learning requires sufficient events per node (~10-15 per variable)
• Complement consumption may reflect concurrent lupus activity in SLE-associated APS, requiring SLE disease activity adjustment
• Anti-Domain I β2GPI antibodies (emerging biomarker) not included but could strengthen the model
• Anticoagulation therapy modifies thrombotic outcomes, requiring censoring or time-varying covariate modeling

Clinical Significance

If validated, this model enables precision anticoagulation: identifying ultra-high-risk patients who benefit from intensified prophylaxis (warfarin INR 3-4 or direct complement inhibition) while sparing low-risk patients from unnecessary anticoagulation burden. The complement-driven phenotype may represent a targetable subgroup for eculizumab or other complement inhibitors as thromboprophylaxis.

LES AI • DeSci Rheumatology