Predicting the risk of venous thromboembolism events in older adults with rheumatoid arthritis after initiating targeted disease-modifying antirheumatic drugs: a comparison of the random survival forest and regularized cox regression models

OBJECTIVES: To develop a random survival forest (RSF) machine learning (ML) model for predicting venous thromboembolism (VTE) risk in rheumatoid arthritis (RA) patients initiating biological (b) or targeted synthetic (ts) disease-modifying antirheumatic drugs (DMARDs) and compare its model performance with a regularized Cox regression (RegCox) model. METHODS: This retrospective cohort study using the 5% Medicare data (2012-2020) identified older RA patients (≥ 65 years) initiating b/tsDMARDs (index date), including tumor necrosis factor inhibitors (TNFi) bDMARDs, non-TNFi bDMARDs, and tsDMARDs between January 1, 2013, through December 31, 2019. Study cohort was followed until an incident composite VTE event or censoring. Data were divided into training (75%) and testing (25%) sets. The RSF model was trained to predict VTE events during the follow-up period in the training set, with the RegCox model as the reference model. The performance of these models was evaluated in the testing data using the C-index. Variable importance of the predictors was assessed. RESULTS: = 0.0021). Variables commonly identified as the top influential variables were varicose veins, inpatient visits, Elixhauser score, emergency room visits, and outpatient visits. CONCLUSIONS: The RSF model performed slightly better in identifying VTE in RA patients after b/tsDMARDs initiation than RegCox. Incorporating additional clinical and contextual information beyond claims data may further enhance predictive accuracy in future studies.