2324: Early Prediction of Complex Forms of Spondyloarthritis Using Artificial Intelligence: A Modeling Study Based on Routine Clinical Data

Background/Purpose: Spondyloarthritis (SpA) is a chronic inflammatory disease with heterogeneous clinical manifestations. While some patients exhibit mild forms, others progress to complex forms characterized by persistent inflammation, organ involvement (pulmonary, cardiovascular), or severe structural damage. Early identification of high-risk profiles is crucial for tailoring therapeutic strategies. Artificial intelligence (AI) represents a promising tool to predict disease progression based on simple, readily available clinical data.

The objective was to evaluate the performance of a machine learning model (Random Forest) in predicting complex clinical forms of SpA using routine clinical variables that are independent of the criteria defining complexity.

Methods: The study included 319 patients with SpA. A complex form was defined by the presence of at least two of the following criteria:

• Elevated CRP ( >10 mg/L)


• Uveitis


• Pulmonary involvement


• Cardiovascular complications


• Hip involvement (coxitis)

The model was trained using five clinical variables unrelated to the definition of complexity:

• Age


• Sex


• CRP


• Erythrocyte sedimentation rate (ESR)


• Glomerular filtration rate (GFR)

The dataset was split into a training set (80%) and a test set (20%). The performance of the Random Forest model was assessed using accuracy, AUC-ROC, sensitivity, precision, and confusion matrix.

Results: The Random Forest model, based on five simple clinical variables (age, sex, CRP, ESR, GFR), achieved an overall accuracy of 78% on the test set. Its discriminative capacity, as measured by the area under the ROC curve (AUC), was high at 0.88, indicating strong ability to distinguish between simple and complex forms of SpA. The model correctly identified 71% of patients with a complex form (sensitivity), while maintaining a specificity of 81% for simple forms. The precision for complex cases was 65%, suggesting useful performance for identifying high-risk patients. It also demonstrated a high negative predictive value, indicating good reliability in ruling out severe forms.

Among the variables used, CRP was the most predictive factor, followed by ESR and GFR. Age and sex provided complementary contributions to model stability but were not major determinants.

Conclusion: This study shows that an artificial intelligence model based on simple and accessible clinical data can reliably anticipate complex forms of spondyloarthritis. AI does not replace the clinician but serves as a strategic partner in personalizing care and improving patient outcomes in rheumatology.