Rheumatoid arthritis-associated interstitial lung disease (RA-ILD) can significantly shorten patients' life expectancy, severely impair their quality of life, and impose an additional physical, psychological, and economic disease burden on patients and their families. This disease is mainly manifested as the usual interstitial pneumonia (UIP) pattern, followed by the nonspecific interstitial pneumonia (NSIP) pattern. Clinically, accurately distinguishing these two patterns is crucial for optimizing individualized clinical decision-making and improving patient management, yet it remains a great challenge due to the complex clinical manifestations and high heterogeneity of RA-ILD. Based on clinical features combined with radiomics, this study aims to establish and validate a CT radiomic integrated model to distinguish the UIP and NSIP patterns of RA-ILD, and further analyze metabolites related to radiomic features, so as to explore the potential biological significance of radiomic features and provide a scientific basis for clinical diagnosis and treatment.

A total of 181 RA-ILD patients (92 with UIP pattern, 89 with NSIP pattern) diagnosed in the First Affiliated Hospital of Army Medical University from January 2011 to June 2024, and 49 RA-ILD patients (27 with UIP pattern, 22 with NSIP pattern) diagnosed in Dazhou Central Hospital from January 2017 to December 2023 were enrolled retrospectively. Patients from the First Affiliated Hospital of Army Medical University were randomly divided into the training set (144 cases) and internal validation set (37 cases), while patients from Dazhou Central Hospital served as the independent external validation set (49 cases). Univariate analysis was used to evaluate clinical variables associated with RA-ILD subtypes. A total of 1688 radiomic features were extracted from the volume of interest (VOI) in CT images, followed by feature screening to eliminate redundant and irrelevant features and calculation of the Radiomics score (Rad-score). Three machine learning algorithms—logistic regression (LR), random forest (RF), and K-nearest neighbor (KNN)—were employed to construct clinical models, radiomic models, and clinical-radiomic integrated models respectively. A nomogram was built by combining the Rad-score and independent clinical variables. The diagnostic efficacy, clinical benefit, and calibration degree of each model were systematically evaluated. Liquid chromatography-tandem mass spectrometry was used for plasma metabolomic analysis. Spearman correlation analysis was performed between radiomic features and metabolomic results to identify significantly correlated metabolites, followed by metabolite enrichment analysis to explore the potential metabolic pathways involved.

Univariate analysis showed that smoking history (OR: 1.988 [95%CI, 1.05–3.835], P=0.037) and erythrocyte sedimentation rate (ESR) (OR: 1.012 [95%CI, 1.002–1.022], P=0.018) were independent predictors for distinguishing the UIP and NSIP patterns of RA-ILD. After a series of feature screening steps, 12 optimal radiomic features with high stability and strong discriminative ability were retained. A total of 292 metabolites were identified to be significantly correlated with radiomic features (R>0.4, P<0.05), among which 4-aminobutyrate and L-glutamate were co-enriched in multiple metabolic pathways, suggesting their potential association with the pathological mechanism of RA-ILD subtypes. In the three datasets (training set, internal validation set, and external validation set), the diagnostic performance of the radiomic model and clinical-radiomic integrated model was significantly better than that of the clinical model. The nomogram showed high consistency between the predicted results and the actual clinical situation, indicating good calibration. Decision curve analysis (DCA) confirmed that the clinical-radiomic integrated model could exhibit higher net benefit compared with the clinical model and radiomic model, verifying its good clinical utility.

The clinical-radiomic integrated model (nomogram) constructed in this study by combining smoking history, ESR, and Rad-score can be used as a reliable tool to distinguish the UIP and NSIP patterns of RA-ILD, which is conducive to the early classification and accurate diagnosis of RA-ILD and provides strong support for clinical decision-making. The Rad-score can effectively improve the ability to distinguish RA-ILD subtypes. In addition, 4-aminobutyrate and L-glutamate, which are significantly correlated with radiomic features and co-enriched in multiple metabolic pathways, have great potential for further research, and may provide new insights into the biological mechanism of RA-ILD subtypes.