This cross-sectional study was conducted within the Veterans Affairs Rheumatoid Arthritis (VARA) registry, which is a prospective, multicenter cohort of U.S. Veterans with RA fulfilling the 1987 American College of Rheumatology classification criteria that was initiated in 2003^{16, 17}. This study was approved by the VA Nebraska-Western Iowa Health Care System institutional review board (IRB; 1576193), with each site receiving IRB approval and all participants providing written informed consent. Study participants and the public were not involved in the design, reporting, or dissemination plans of this research.

Clinical and demographic data, including age, sex, race/ethnicity, body mass index (BMI, kg/m²), RA disease duration, and cigarette smoking history (categorized as ever or never smoker), were recorded at the time of registry enrollment. Race/ethnicity was self-reported from a fixed set of categories, and subsequently categorized for this study as non-Hispanic White or other. The disease activity score with 28-joint count (DAS28) and multidimensional Health Assessment Questionnaire (MD-HAQ) were documented by treating rheumatologists at enrollment. Disease-modifying anti-rheumatic drug (DMARD) use at enrollment was determined by linkage with administrative pharmacy dispensing data obtained through the VA Corporate Data Warehouse. Participants with missing demographic or biomarker data were excluded from the analysis (i.e., complete case analysis).

ILD was assessed in the cohort using screening and validation steps, as previously reported^{7, 8, 18}. All participants were screened for possible ILD based on diagnostic codes for ILD or ILD-related terms in chest computed tomography (CT) reports^{19, 20}. ILD diagnoses were subsequently validated by a board-certified rheumatologist based on documentation of a diagnosis of ILD in the medical record by the treating provider with supporting findings from either imaging or lung biopsy per the radiology or pathology report (chest CT in 97% of ILD cases)¹⁹. Diagnostic evaluation (CT, pulmonary function testing, lung biopsy) was performed according to clinical indication at the discretion of treating providers. Given disease latency, prevalent ILD status was defined as the presence of ILD at any time before and up to two years after registry enrollment. Participants who later developed RA-ILD or who had indeterminate ILD were excluded from analyses. When available, ILD pattern was extracted from clinical radiology reports, biopsy reports, and clinical notes. The presence of honeycombing on chest CT reports was considered additional evidence of a UIP pattern¹⁸. Forced vital capacity (FVC) percent predicted was extracted from medical records at the most proximate time to enrollment.

Biomarker assessment was performed using serum and plasma that were collected at enrollment on all participants and stored at −70°C. Biomarkers were selected based on prior associations with RA and/or RA-ILD as part of previous studies within the VARA cohort^{5, 15, 18, 21–23}. Biomarker measurement was performed via either the Mesoscale Discovery platform, enzyme-linked immunosorbent assay (ELISA), or nephelometry (Table 1). Adipokines included adiponectin, fibroblast growth factor (FGF)-21, and leptin. Alarmins included interleukin (IL)-17e-IL-25, IL-33, and thymic stromal lymphopoietin (TSLP). RA-related antibodies included anti-MAA (IgA, IgM, IgG; albumin, collagen, fibrinogen, and vimentin; ELISA), anti-CCP (IgG ELISA), and RF (nephelometry). Chemokines included eotaxin, monocyte chemoattractant proteins (MCPs), macrophage-derived chemokine (MDC), macrophage inflammatory proteins (MIPs), and thymus and activation-regulated chemokine (TARC). Collagenases included matrix metalloproteinases (MMPs). Cytokines included Fms-like tyrosine kinase 3 ligand (FL), fractalkine, granulocyte colony stimulating factor (G-CSF), granulocyte-macrophage colony-stimulating factor (GM-CSF), interferons, interleukins, tumor necrosis factor (TNF)α, and vascular endothelial growth factor (VEGF). All biomarker concentrations were log-transformed and standardized for analyses, with a mean of 0 and a standard deviation of 1.

Genotyping was performed using the Infinium Global Screening Array-24 v2.0 microarray (Illumina, Inc.; San Diego, CA, USA). MUC5B rs35705950 was directly measured and passed quality control measures including Hardy-Weinberg equilibrium. The MUC5B rs35705950 promoter variant was analyzed via an additive model according to the number of thymine (T) copies as the variant allele⁷.

Principal component analysis (PCA) was utilized to identify unique peripheral biomarker signatures for RA-ILD. PCA is a statistical technique that has the powerful ability to compress large amounts of predictor variables into more accessible groups, as well as illustrate patterns in the data based on these variables^{24, 25}. All standardized, log-transformed biomarkers (Table 1) were included in the PCA. As part of PCA, a correlation matrix was constructed for the aforementioned biomarkers, Eigenvalues were calculated from this matrix, and PCs were selected based on Eigenvalue (>1, per Kaiser-Guttman rule), scree plots, and cumulative variance^25–27. An orthogonal varimax rotation was used to improve the interpretability of extracted PCs, forcing variables to align strongly with a single component²⁴. Biomarkers with loadings >0.3 were considered to meaningfully represent the respective PC²⁶. Each PC was descriptively characterized based on the function of the biomarkers with a meaningful loading. The relationship of each PC with the outcome of RA-ILD was initially evaluated as a PC score within a multivariable logistic regression model that adjusted for age, sex, race/ethnicity, and smoking status.

Clinical variables, including age, sex, race/ethnicity, and smoking history, were included in a multivariable logistic regression model, evaluating the outcome of RA-ILD among the entire RA cohort. The 15 PCs were included together in a multivariable model assessing the predictive capacity of biomarkers alone. Subsequently, a combined clinical and biomarker logistic regression model was assessed, including the clinical factors and all 15 PCs extracted from PCA. A final logistic regression model included the combination of clinical factors, PCs, and the MUC5B promoter variant. The area under the receiver operating characteristic curve (AUC) was compared among the clinical, clinical + PCA, and clinical + PCA + MUC5B models, and statistical comparisons were made via the nested likelihood ratio test to evaluate for improvement in model fit. Performance of a clinical + MUC5B model was assessed and compared with the clinical + PCA + MUC5B model. As a sensitivity analysis, this process was then repeated using an expanded clinical model that included BMI, DAS28, prednisone use, and DMARD use. Additional sensitivity analysis was performed with the inclusion of only those PCs which were individually associated with RA-ILD within the models. Internal validation was performed for all models using bootstrapping with 50 different random samples and optimism-adjusted measures were obtained²⁸.

We performed subgroup analyses of individual PC associations with RA-ILD subtypes (UIP, non-UIP) and FVC severity (FVC<80%, FVC≥80% predicted) versus no-ILD via multivariable logistic regression with adjustment for age, sex, race, and smoking history. Logistic regression model performance was also evaluated on the UIP subgroup. All analyses were completed with Stata v18.

Among 2,001 participants with RA, 121 (6.4%) were classified as having prevalent ILD. Participants with RA-ILD were older (average age 67.4 vs. 63.5 years; p<0.001) and more frequently male (94.2% vs. 88.3%; p=0.047) compared to RA patients without ILD (Table 2). A cigarette smoking history (84.3% vs. 77.4%; p=0.07), as well as positivity for RF (86.8% vs. 76.3%; p=0.01) and anti-CCP (83.5% vs. 76.6%; p=0.08) were more frequent in those with ILD. In patients with RA-ILD, there was more frequent use of biologic/targeted synthetic disease-modifying anti-rheumatic drugs (b/tsDMARDs; 33.9% vs. 25.1%; p=0.03) and prednisone (33.9% vs. 17.7%; p<0.001). Among those with RA-ILD, evidence of a UIP pattern was present in 62.8% of cases and the mean FVC% predicted was 77.6. The MUC5B promoter variant occurred more frequently among those with RA-ILD (36.8% vs. 16.8%).

Fifteen PCs were extracted based on Eigenvalue >1 according to the Kaiser-Guttman rule²⁷. These PCs explained 66% of the cumulative variance (Table S1), with the first two PCs accounting for the largest amount of variance (PC1 15.5% and PC2 11.9%, respectively).

The peripheral biomarkers that loaded onto each respective PC, a descriptive characterization of each PC, and the PC score association with RA-ILD are provided in Table 3. The first two PCs predominantly comprised pro-inflammatory cytokines and chemokines and were not associated with RA-ILD. Eight PCs were individually associated with RA-ILD. PC3 (adjusted odds ratio [aOR] 1.16, 95% CI 1.05–1.29) was composed of cytokines involved in the innate immune response (IL-3, IFNα−2a, IL-15, and IL-5). PC4 (aOR 1.28, 95% CI 1.11–1.47), PC7 (aOR 1.18, 95% CI 1.01–1.37), and PC15 (aOR 1.26, 95% CI 1.07–1.49) were composed of distinct anti-MAA antibodies to albumin, fibrinogen, collagen, and/or vimentin. PC6 (aOR 1.17, 95% CI 1.02–1.34) was composed of anti-CCP antibodies and RF. PC8 (aOR 1.43, 95% CI 1.23–1.66) was composed of adipokines and cytokines (FL, leptin, fractalkine, and FGF-21). PC10 (aOR 1.21, 95% CI 1.03–1.41) was composed of alarmins (IL-17e-IL-25 and TSLP). PC12 (aOR 1.20, 95% CI 1.03–1.40) was composed of biomarkers related to tissue remodeling and neutrophil chemotaxis (MMP9 and IL-8). Other PCs were not significantly associated with RA-ILD.

A logistic regression model including only clinical risk factors was able to discriminate RA-ILD with an AUC of 0.630 (95% CI 0.582–0.679). A logistic regression model composed of all 15 PCs was able to discriminate RA-ILD with an AUC of 0.713 (95% CI 0.662–0.764). When all PC scores were included with the clinical risk factors, the predictive ability of the model was significantly increased over the clinical model (AUC 0.739, 95% CI 0.693–0.786; Figure 1a). The addition of the MUC5B promoter variant to the clinical + PC model further improved performance (AUC 0.751, 95% CI 0.707–0.802; Figure 1a). This was also improved over the clinical + MUC5B model (AUC 0.681, 95% CI 0.637–0.726; Table 4). All model comparisons were significant via the nested likelihood ratio test (p<0.001).

Sensitivity analyses with BMI, DAS28, and DMARD use included as additional clinical variables yielded slightly improved performance in clinical (0.706, 95% CI 0.659–0.752), clinical + PC (AUC 0.767, 95% CI 0.723–0.811), and clinical + PC + MUC5B models (AUC 0.776, 95% CI 0.733–0.823; Figure 1b). Again, all model comparisons were significant via the nested likelihood ratio test (p<0.001). Sensitivity analysis with PCs selected based on individual associations (PC3, PC4, PC6, PC7, PC8, PC10, PC12, PC15) did not significantly impact the results (Table 4). Internal validation was performed on all models using bootstrapping. Model performance was attenuated with optimism correction, but the overall trend in performance between the various models was preserved, as shown Table 4.

Subgroup analyses were performed to evaluate the association of each individual PC with RA-ILD based on ILD subtype and FVC impairment. Most PC associations with RA-ILD were consistent across subgroups (Table S2). The exception was PC10 (IL-17e-IL-25 and TSLP), which was positively associated with UIP and negatively associated with non-UIP ILD. Additionally, PC13 (IgA/IgM antibodies to MAA-adducted collagen and adiponectin) was more closely associated with non-UIP ILD and PC7 (IgA/IgG/IgM antibodies to MAA-adducted fibrinogen, collagen, and vimentin) was more closely associated with severe ILD (FVC<80%), though 95% CI for estimates were overlapping.

While a clinical model performed similarly in the UIP subgroup (AUC 0.635 vs. 0.630), there was a substantial improvement in performance of the PC model (AUC 0.768 vs. 0.713) and combined models (clinical+PC+MUC5B model AUC 0.834 vs. 0.755; Table S3).