Machine learning-based prediction of 30-day mortality in critically ill patients with rheumatoid arthritis.

BACKGROUND: Rheumatoid arthritis patients in the ICU face a high risk of mortality. While traditional ICU scoring systems are not specifically designed for the unique pathophysiological profile of RA. This study aimed to develop a machine learning framework to accurately predict 30-day mortality for these patients. METHODS: Data from 400 RA patients in the MIMIC-IV database were analyzed. LASSO regression identified nine pivotal predictors: age, BUN (urea), PT, respiratory rate, SpO2, glucose, urine output, and coronary artery disease. Six ML models were constructed using SMOTE to handle class imbalance. Performance was evaluated via AUC, sensitivity, and SHAP analysis for interpretability. RESULTS: The LR-SMOTE model demonstrated the best discriminative ability (AUC = 0.69), while the Stacking ensemble achieved the highest sensitivity (0.8). External validation on the eICU dataset yielded an AUC of 0.747 for the LR-SMOTE model. SHAP analysis identified urine output, CAD, and age as the most influential predictors. CONCLUSIONS: The machine learning framework demonstrates superior performance compared to most traditional scoring systems, while offering the distinct advantages of easier data acquisition and lower computational complexity. By leveraging readily accessible clinical parameters, it supports proactive, individualized clinical decision-making in the ICU.