In a recent study published in Frontiers in Medicine, researchers developed and validated machine learning–based models to improve early identification of interstitial lung disease associated with rheumatoid arthritis (RA-ILD) using routinely available clinical and laboratory data.
The researchers conducted a retrospective analysis of 410 patients with rheumatoid arthritis, among whom 100 (24.39%) had confirmed RA-ILD.
These variables were then used to construct five different machine learning models: CatBoost, logistic regression, support vector machine, decision tree, and random forest.
This approach identified CA199, CA125, and age as the most influential predictors in determining RA-ILD risk.
Overall, the study concludes that both CatBoost and decision tree models provide promising approaches for classifying RA-ILD with the use of simple clinical data.
In a recent study published in Frontiers in Medicine, researchers developed and validated machine learning–based models to improve early identification of interstitial lung disease associated with rheumatoid arthritis (RA-ILD) using routinely available clinical and laboratory data. The goal was to create a practical risk stratification tool that could be applied in primary care or general hospital settings, where access to advanced imaging or specialist evaluation may be limited.
The researchers conducted a retrospective analysis of 410 patients with rheumatoid arthritis, among whom 100 (24.39%) had confirmed RA-ILD. After initial data preprocessing, the cohort was split into training and validation sets. A broad set of clinical variables was first screened using univariate analysis, and those showing potential significance were further refined by means of LASSO regression to identify the most informative predictors.
Seven key features were ultimately selected to build the predictive models: age, smoking history, lymphocyte count, lactate dehydrogenase, rheumatoid factor, cancer antigen 125 (CA125), and carbohydrate antigen 199 (CA199). These variables were then used to construct five different machine learning models: CatBoost, logistic regression, support vector machine, decision tree, and random forest.
Model performance was assessed using standard classification metrics, including accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUC). Among all models, CatBoost demonstrated the strongest overall performance, achieving an AUC of 0.784 (95% CI, 0.656-0.885) and the lowest Brier score of 0.158, indicating good calibration and discrimination. The decision tree model performed similarly, with an AUC of 0.783 (95% CI, 0.661-0.818), and showed the highest recall (0.653) and F1 score (0.603), suggesting strong sensitivity in identifying RA-ILD cases.
To improve interpretability, the authors applied SHapley Additive exPlanations (SHAP) analysis to the CatBoost model. This approach identified CA199, CA125, and age as the most influential predictors in determining RA-ILD risk. These findings suggest that tumor-associated markers, alongside demographic and inflammatory indicators, may carry meaningful signals for pulmonary involvement in RA.
Overall, the study concludes that both CatBoost and decision tree models provide promising approaches for classifying RA-ILD with the use of simple clinical data. Although performance was moderate rather than definitive, the results highlight the potential of machine learning tools to assist in early risk stratification. The authors emphasize that external validation in larger, independent cohorts is necessary before clinical application, but the approach may eventually support earlier detection and referral for patients at risk for RA-associated lung disease.
