This study by Baloun et al. developed machine learning models to identify clinical features that predict progression to difficult-to-treat rheumatoid arthritis (D2T RA). Researchers analyzed 8,543 RA patients in the Czech ATTRA registry (2002–2023), comparing 641 D2T RA cases to 1,825 patients in sustained remission (characterized by Simple Disease Activity Index<3.3 and ≤1 swollen joint). Five machine learning algorithms - lasso, ridge regression, support vector machines, random forest, and XGBoost - were trained on 25 routinely collected clinical and demographic variables. Models achieved accuracy 0.606–0.747 and AUC 0.656–0.832, with performance improving as the D2T RA stage approached. Explainable AI using SHapley Additive exPlanations (SHAP) identified the most predictive features: DAS28-ESR (mean rank = 2.3), CDAI (3.0), duration of biologic/targeted therapy (3.5), C-reactive protein (3.7), and HAQ disability score (5.3). Elevated inflammation markers (ESR, CRP), high disease activity scores, and longer biologic treatment duration strongly correlated with the D2T phenotype. Overall, the ML approach revealed that worsening disease activity and functional impairment become detectable up to one year before patients meet the D2T RA definition, suggesting a potential window for early therapeutic intervention.