Combining radiomics based on high-resolution computed tomography with plasma metabolomics for diagnosing subtypes of rheumatoid arthritis-associated interstitial lung disease

Objectives Rheumatoid arthritis-associated interstitial lung disease (RA-ILD) is primarily classified into usual interstitial pneumonia (UIP) and non-specific interstitial pneumonia (NSIP) patterns, which exhibit significant differences in prognosis and treatment response. This study aimed to differentiate UIP and NSIP patterns in RA-ILD by integrating HRCT-based radiomics with plasma metabolomics. Methods We included 350 RA-ILD patients, assigned to training, internal validation, and external validation sets. Radiomics features were extracted from HRCT images, and optimal features were selected using variance threshold, univariate selection, and LASSO regression. Three machine learning algorithms—random forest (RF), logistic regression (LR), and multi-layer perceptron (MLP)—were each used to construct three categories of models: clinical models, radiomics models, and combined clinical-radiomics models. This design resulted in a total of 9 models (3 algorithms × 3 model categories). A nomogram integrating the radiomics score (Rad-score) and clinical variables was developed. Model performance was assessed via AUC, calibration curves, and decision curve analysis (DCA). Plasma metabolomics analysis was performed using liquid chromatography-tandem mass spectrometry, and Spearman correlation was used to explore the relationship between radiomics features and metabolites. Results Twenty-five optimal radiomics features were selected, and 77 metabolites were significantly correlated with radiomics features (R > 0.4, P < 0.05). L-glutamate and alpha-ketoglutarate were co-enriched in key metabolic pathways (e.g., arginine biosynthesis, D-glutamine/D-glutamate metabolism). Radiomics and combined models outperformed clinical models: the MLP combined model showed more favourable discrimination in the training set (AUC = 0.915; 95% CI: 0.884–0.940) and internal validation set (AUC = 0.828; 95% CI: 0.728–0.916), while the RF combined model showed the best discrimination in the external validation set (AUC = 0.841; 95% CI: 0.729–0.916, specificity = 67%). The nomogram exhibited excellent calibration (Hosmer–Lemeshow: P = 0.051, 0.382, 0.073 for training, internal, external sets) and clinical utility (DCA: high net benefits across thresholds). Conclusions The combined clinical-radiomics model provides a promising auxiliary tool for distinguishing subtypes of RA-ILD, which may support early and accurate identification of RA-ILD subtypes to aid clinical decision-making. The Rad-score improves RA-ILD subtype discrimination, and L-glutamate/alpha-ketoglutarate merit further investigation as potential metabolic markers.