Introduction

Differential diagnosis of inflammatory arthritis is one of the most challenging problems in clinical rheumatology. A patient presenting with polyarthritis may have RA, PsA, gout, reactive arthritis, or SLE — each requiring fundamentally different treatment strategies. Current approaches apply classification criteria independently, missing the rich conditional dependencies between clinical features.

Methods

Bayesian Network Structure

We construct a 20-node DAG with Diagnosis as the root node and 20 observable clinical features as children. Edge structure encodes the generative model: P(Feature | Diagnosis). Inference reverses this via Bayes theorem to compute P(Diagnosis | Observed Features).

Conditional Probability Tables

CPTs derived from:

• ACR/EULAR 2010 RA criteria sensitivity/specificity data
• CASPAR 2006 PsA criteria
• ACR/EULAR 2015 Gout criteria
• Published sensitivity/specificity of individual features per disease
• Expert calibration by Dr. Zamora-Tehozol (Board-Certified Rheumatologist)

Prevalence Priors

Latin American prevalence: RA 35%, PsA 20%, Gout 25%, ReA 10%, SLE-articular 10% (GLADAR, BIOBADAMEX cohorts).

Inference

Variable Elimination — exact inference, no sampling. Missing features are marginalized automatically.

Results

Key Advantages Over Monte Carlo

• Exact inference — no sampling variance
• Interpretable — each CPT has clinical meaning
• Missing data — natural via marginalization, no imputation
• Causal structure — can answer interventional queries
• Computational efficiency — <50ms per query

References

1. Koller D, Friedman N. Probabilistic Graphical Models. MIT Press, 2009.
2. Aletaha D, et al. 2010 RA Classification. Arthritis Rheum. 2010;62:2569-81.
3. Taylor W, et al. CASPAR. Arthritis Rheum. 2006;54:2665-73.
4. Neogi T, et al. 2015 Gout Classification. Arthritis Rheumatol. 2015;67:2557-68.