Factors associated with concurrent cardiovascular disease in patients with rheumatoid arthritis and development of...

To analyze factors associated with concurrent cardiovascular disease (CVD) in patients with rheumatoid arthritis (RA) and to develop machine learning models for identifying current CVD comorbidity status using routinely available clinical data.

This was a cross-sectional observational study. We included 4,767 patients with RA who attended Beijing Jishuitan Hospital from January 2021 to December 2024 and met the diagnostic criteria. According to the current presence of CVD, patients were divided into an RA-CVD group (n = 366) and an RA-non-CVD group (n = 4,401). Demographic characteristics, comorbidities, and laboratory indicators were collected. Univariate analysis was used to compare baseline characteristics between groups. Multivariable logistic regression models were then built after considering data completeness, clinical relevance, and collinearity. A high-completeness primary model was used as the main analysis, and a clinically enhanced model was used as a supplementary analysis. Multiple imputation was further performed as a missing-data sensitivity analysis to assess robustness. A simplified conventional-risk baseline model was additionally constructed using available traditional cardiovascular risk factors as a benchmark. Based on clinical data, we then developed core-variable models, wide-feature models, threshold-adjusted wide-feature models, and medication-information-available subgroup models. The dataset was divided into training and test sets at a ratio of 8:2, and 5-fold stratified cross-validation was performed in the training set. Model performance was evaluated by the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (PR-AUC), accuracy, sensitivity, specificity, and F1 score. The modeling task was defined as identification of concurrent CVD status rather than prediction of incident cardiovascular events.

The current prevalence of CVD in patients with RA was 7.68%. Compared with the RA-non-CVD group, the RA-CVD group had a higher proportion of men, older age, longer disease duration, and higher rates of hypertension and diabetes. Several indicators reflecting inflammatory burden, renal function, metabolic status, and hematologic characteristics also differed significantly between groups. Multivariable logistic regression showed that age, male sex, and disease duration were stable factors associated with CVD in RA in the high-completeness primary model. Multiple imputation sensitivity analysis further supported the robustness of these findings. In the clinically enhanced model, age and male sex remained stably associated with CVD, whereas hypertension, diabetes, erythrocyte sedimentation rate, and estimated glomerular filtration rate did not show stable independent statistical associations. The simplified conventional-risk baseline model showed comparable discrimination to the core-variable logistic regression model, but its sensitivity was very low at the default threshold of 0.50. The non-resampled baseline model achieved an AUC of 0.78 and a PR-AUC of 0.22, with a sensitivity of 0.04 at the default threshold. Core-variable models had moderate ability to distinguish patients with and without concurrent CVD. Adding a broader set of structured variables improved model discrimination, but sensitivity remained very low at the default threshold of 0.50. Training-derived Youden-threshold adjustment improved minority-class identification. For example, the sensitivity of the wide-feature XGBoost model increased from 0.06 at the default threshold to 0.81 after training-derived threshold adjustment.

The current presence of CVD in RA reflects a complex comorbidity pattern characterized by multiple associated clinical factors. Age, male sex, and longer disease duration were relatively stable associated factors. Within the present internally validated setting, machine learning models based on routinely available clinical data may provide preliminary support for identifying concurrent CVD status in patients with RA. Expanding the variable set and applying training-derived threshold adjustment may improve recognition of minority-class patients in this imbalanced dataset. These findings provide a preliminary data-driven framework for comorbidity-oriented cardiovascular assessment in RA, but model stability, threshold transportability, and generalizability require further validation in multicenter prospective cohorts.