A machine learning-based classification model for interstitial lung disease in rheumatoid arthritis.

Objective: This study aimed to develop and validate a preliminary classification and diagnostic model for rheumatoid arthritis-associated interstitial lung disease (RA-ILD) using routine, readily available clinical and laboratory parameters. Multiple machine learning algorithms were employed to construct a practical risk assessment tool suitable for use in primary hospital settings. Methods: Clinical data were retrospectively collected. Patients were divided into RA and RA-ILD groups. After preprocessing, the cohort was randomly divided into training and validation sets. Variables demonstrating a trend toward significance on univariate analysis were subjected to LASSO regression, and feature variables were ultimately identified. Five machine learning models were constructed: CatBoost, logistic regression, support vector machine, decision tree, and random forest. Model performance was assessed on the validation set using accuracy, precision, recall, F1 score, and the area under the receiver operating characteristic (ROC) curve (AUC). The SHapley Additive exPlanations (SHAP) framework was used to identify key features and quantify their contributions to the optimal predictive model. Results: A total of 410 patients with RA were enrolled, among whom 100 (24.39%) were diagnosed with RA-ILD. 23 variables with a trend toward significance on univariate analysis were subjected to LASSO regression. Finally, seven features were selected for model construction: age, smoking history, LYMPH, LDH, RF, CA125, and CA199. Based on these features, five machine learning models were established for RA-ILD classification. In the validation set, the CatBoost model achieved the highest AUC of 0.784 (95% CI: 0.656-0.885) and the lowest Brier score of 0.158, demonstrating robust overall performance. The decision tree (DT) model exhibited comparable discriminatory ability, with an AUC of 0.783 (95% CI: 0.661-0.818), and attained the highest recall (0.653) and F1-score (0.603) across all models, reflecting strong classification efficacy. Conclusion: Among the five evaluated models, CatBoost and DT models showed comparable and favorable overall performance for RA-ILD classification. SHAP analysis based on the CatBoost model identified CA199, CA125, and age as the most important contributors to model prediction. Both models hold promise for RA-ILD risk stratification in clinical practice, although further external validation is warranted.