Machine Learning for Predicting Recurrent Course in Uveitis Using Baseline Clinical Characteristics

PURPOSE. We developed and evaluated machine learning models for predicting the risk of recurrent uveitis using baseline clinical characteristics, to inform clinical decision-making and risk stratification. METHODS. A retrospective analysis was conducted using the Ocular Autoimmune Systemic Inflammatory Infectious Study registry, including 966 patients (1432 eyes) with uveitis. Three machine learning classifiers—random Forest, eXtreme Gradient Boosting, and radial basis function support vector classifier—were trained on preprocessed baseline demographic and clinical data. Predictors were selected through bivariate analysis with false discovery rate correction. Models were optimized using grid search with five-fold stratified cross-validation. Performance was evaluated on a hold-out test set using accuracy, sensitivity, specificity, area under the receiver operating characteristic curve, and Shapley additive explanations values for feature importance. RESULTS. The random Forest model achieved the highest test accuracy (0.77), with high specificity (0.93) but modest sensitivity (0.44) for identifying recurrences. eXtreme Gradient Boosting and radial basis function support vector classifier showed comparable accuracies (0.73 and 0.74, respectively) but slightly lower sensitivities. Shapley additive explanation analysis identified vitreous haze, retrolental cells, and noninfectious etiology as key predictors. Learning curves indicated that model performance stabilized with the available sample size, suggesting adequate training data. CONCLUSIONS. Machine learning models, particularly random Forest, effectively identified patients at low risk of uveitis recurrence, offering high specificity. However, sensitivity remained limited, highlighting challenges in predicting infrequent events in a heterogeneous disease population.