Predictive Rule Extraction for Ocular Involvement in Behcet’s Disease

A Data Mining Approach Using Hospital-Based Data

Abstract

Introduction: Ocular involvement is among the most severe and vision-threatening complications of Behçet’s disease (BD), making early risk stratification vital. This study aims to develop and validate an interpretable predictive data mining model to identify patients at high risk of ocular manifestations using a comprehensive hospital-based cohort.

Material and Methods: A retrospective analysis was conducted on clinical and demographic data from 1,613 patients diagnosed with Behçet’s disease at the rheumatology research center, Shariati hospital, Tehran. The J48 decision tree algorithm was implemented to extract explicit predictive rules. To establish a rigorous benchmark, a traditional logistic regression (LR) model was also evaluated. Both models were rigorously assessed using a 10-fold cross-validation framework across core metrics including accuracy, sensitivity, specificity, F1-score, and ROC-AUC.

Results: The J48 decision tree model significantly outperformed the traditional logistic regression baseline, achieving a mean overall accuracy of 91.73±1.25, sensitivity of 74.78±1.84, specificity of 73.01±2.15 and AUC of 0.86±0.02 compared to an accuracy of 78.45 and an AUC of 74 for logistic regression. The root and primary nodes of the decision tree identified key risk factors, specifically genital ulceration, HLA-B51 status, sex, and ESR levels, generating transparent clinical rules for early identification.

Conclusion: The J48 decision tree algorithm provides superior predictive accuracy and clinical interpretability over conventional logistic regression for assessing ocular involvement in Behçet’s disease. These transparent rules offer a robust foundation for integration into Clinical Decision Support Systems (CDSS) to assist rheumatologists in proactive, individualized patient management.